betaFeedback

WorksAgricultural and Biological Sciences

Efficient and Accurate Zero Shot Generation of Symmetric Protein Complexes

Abstract

Symmetric organization is a fundamental principle of biomolecular complex assembly. The same principle underlies engineered applications, from vaccine platforms to nanomaterials, making symmetric de novo generation a core capability for computational protein design. We establish a theoretical foundation equating generation of an assembly with underlying symmetry group GG to generation within a quotient space over GG. This directly allows SE(3)\mathrm{SE}(3)-equivariant models to generate symmetric structures with minimal distribution shift. With this insight we introduce ZeUS (Ze\textbf{Ze}ro-shot U\textbf{U}nified S\textbf{S}ymmetrization), a framework that converts any pretrained SE(3)\mathrm{SE}(3)-equivariant backbone generative model into an efficient generator of symmetric assemblies with no retraining. ZeUS operates on a single canonical subunit\textit{canonical subunit} and utilizes phantom subunits\textit{phantom subunits} to preserve internal model operations, reducing memory and runtime cost by O(∣G∣)\mathcal{O}(|G|) for finite groups GG, compared to a full-complex forward pass. We implement ZeUS for IPA-based models (FrameDiff, FoldFlow) and axial-attention-based RFdiffusion, showing high generation efficiency with minimal degradation to generation quality. We then apply ZeUS to generate symmetric complexes of over 10,00010{,}000 residues with exact, untruncated attention, a scale previously intractable with any model.

Introduction

Symmetry is a defining organizational principle of biological macromolecular assemblies. From proteasomes to the nuclear pore complex, the most functionally consequential large-scale protein architectures in nature adopt regular, symmetric quaternary structures, and over ∼40%\sim40\% of proteins in the Protein Data Bank [2] exhibit some form of symmetry [13]. Symmetry is also a naturally attractive design strategy for large protein architectures, as an assembly of kk identical subunits is encoded by a single gene and self-organizes under more favourable free-energy minimization [Zandi et al., 2004]. The same logic motivates engineered symmetric design, which underpins applications from vaccine platforms and targeted drug delivery to enzyme cascades and biomaterials. Therefore, de novo generation of symmetric protein complexes is a core capability for computational protein design. However, training data for symmetric structures is sparse. For example, dimers are well represented in the PDB, but higher order assemblies, such as tetrahedral complexes, number at most a few hundred structures [27], making training and finetuning on symmetric structures difficult. We argue in favour of leveraging the generative capacity of existing models for the task of symmetric protein design, since modern protein backbone generators [17] (Watson et al. 2023) are already highly capable of producing diverse, designable protein structures.

In this work we introduce ZEUS (Zero-shot Unified Symmetrization), a zero shot generation framework that converts any SE(3)\mathrm{SE}(3)-equivariant backbone generative model into an efficient symmetric assembly generator with no retraining and no modification of model weights.

A central observation of ZEUS is that symmetric generation is naturally a generative problem in a quotient space over the symmetry group. A protein assembly with kk-fold symmetry group GG with ∣G∣=k|G| = k is fully determined by its single representative subunit, and the full complex is the orbit of that subunit under the group action. A symmetric assembly within an ambient configuration space X\mathcal{X} is therefore equivalent to its subunit assembly within the quotient space X/G\mathcal{X}/G, and a generative model for symmetric assemblies is precisely a generative model within this quotient.

Our work makes the following contributions:

  1. Quotient space formulation. First, we establish that symmetric protein generation can be interpreted as generation in the quotient space X/G\mathcal{X}/G. This is further realized as generation on a single canonical subunit, made aware of its copies defined by group actions.

  1. ZEUS, an efficient zero shot symmetrization algorithm. We introduce ZEUS, which is a framework that allows any backbone diffusion or flow-matching model that satisfies SE(3)\mathrm{SE}(3)-equivariance to generate protein complexes of arbitrary finite symmetry GG, including cyclic (CnC_n), dihedral (DnD_n), and polyhedral (TT, OO, II), without requiring any additional training or fine-tuning.

  1. Empirical validation of ZEUS. We implement ZEUS for two architecture classes: IPA-based models (FrameDiff, FoldFlow) and the axial-attention-based RFdiffusion. We ensure that our method retains generation quality compared to the baseline full-complex inference by evaluating generation designability (scTM, scRMSD), diversity (pairwiseTM), and novelty (pdbTM) across cyclic groups with separate distributional comparisons for polyhedral groups. ZEUS generates distributionally similar designs well within one standard deviation for all metrics compared to the baseline.

  1. Scaling symmetric complex generation. We show theoretically and demonstrate empirically, a significant time and memory reduction of O(k)O(k) during ZEUS’s forward pass, allowing us to tractably generate assemblies that exceed 10,000 total residues in length, an order of magnitude inaccessible to prior methods on a single GPU. We demonstrate this with generation of C50C_{50}, D50D_{50}, HH, and II groups, all diverse and unique.

Background and Related Work

Bottom-up vs. top-down design. Approaches to symmetric protein design can be broadly characterized as bottom-up or top-down. Bottom-up methods [1, 3, 12, 15] (Shen et al. 2024) design modular, reusable building blocks that are subsequently assembled into a symmetric complex. Top-down methods instead design the entire assembly at once, optionally conditioned on high-level design goals. Modern diffusion- and flow-matching-based generative models fall in this category.

Generative models for protein design. Top-down generative models can be grouped into three categories: Backbone-only models generate residue frames that are then passed to an inverse-folding network such as ProteinMPNN [11] to recover an amino acid sequence; representative examples include RFDiffusion [Watson et al., 2023], FrameDiff [Yim et al., 2023], FoldFlow [4], and FrameFlow [Yim et al., 2024]. Backbone–sequence co-design models jointly generate structure and sequence: FoldFlow2 [16], MultiFlow [7], and ProteinGenerator [24]. Finally, All-atom models additionally model sidechain atoms; this includes APM [9], RFdiffusion3 [6], Protpardelle [10], and Chroma [17].

We restrict our study to backbone generative models, which form the largest and most mature family of SE(3)\mathrm{SE}(3)-equivariant protein generators and are the natural setting for our quotient-space construction. To our knowledge, no backbone or all-atom generative model has been retrained or fine-tuned specifically for symmetric generation.

Existing methods for symmetric backbone generation. Among generative models, only Chroma [17] and RFdiffusion [Watson et al., 2023] have been applied to symmetric complex generation. Both treat symmetric design as full-complex generation with explicit re-symmetrization: at each denoising step, the network performs a forward pass over all residues in the assembly, and the predicted update from a designated canonical subunit is propagated to all other subunits according to the target symmetry group. This approach has two limitations. First, the forward pass scales with the size of the full assembly, making it prohibitively expensive in time and memory for higher order symmetries. Second, intermediate model representations are not constrained to be symmetric. Identical subunits are processed independently within the network, with symmetry imposed only on the coordinate update. In effect, these methods do not leverage the symmetry of the problem; they merely enforce it.

Theoretical Foundations

This section develops the theoretical framework underlying ZEUS. § 3.1 establishes notation for backbone frames, symmetry groups, and the quotient space of symmetric assemblies. § 3.2 contains our core theoretical contribution, that full-complex generation is equivalent to quotient-space generation, which is realized practically through the use of phantom copies of a canonical subunit. Detailed proofs and extended discussions are deferred to Appendix § 8.

Notation and Preliminaries

Backbone models and equivariance. Following Jumper et al. [2021], Yim et al. [2023], an NN-residue protein backbone is parameterized as T=[T1,…,TN]∈SE(3)N\mathbf{T} = [T_1, \ldots, T_N] \in\mathrm{SE}(3)^N, where each frame Tn=(rn,xn)T_n = (r_n, x_n) decomposes into a rotation rn∈SO(3)r_n \in\mathrm{SO}(3) and a translation xn∈R3x_n \in\mathbb{R}^3 at the Cα_\alpha position of residue nn. The group SE(3)\mathrm{SE}(3) acts on frames by left multiplication: T0⋅Tn:=(r0rn,r0xn+x0)T_0 \cdot T_n := (r_0r_n, r_0x_n + x_0). We denote SE(3)0N\mathrm{SE}(3)^N_0 the centered submanifold satisfying ∑nxn=0\sum_n x_n = 0, which supports diffusion processes in practice [Yim et al., 2023].

A generative model fθf_\theta is SE(3)\mathrm{SE}(3)-equivariant if fθ(T0⋅T)=T0⋅fθ(T)f_\theta(T_0 \cdot\mathbf{T}) = T_0 \cdot f_\theta(\mathbf{T}) for any global rigid transform T0∈SE(3)T_0 \in\mathrm{SE}(3), with the action extending element-wise. Correspondingly, all internal scalar representations, such as pairwise distances, attention logits, or node embeddings projected to invariant features, are invariant and satisfy s(T0⋅T)=s(T)s(T_0 \cdot\mathbf{T}) = s(\mathbf{T}). This is realized by the IPA module in Jumper et al. [2021], Yim et al. [2023], Bose et al. [2024], and enforced throughout RoseTTAFold in Watson et al. [2023].

In addition, consider a multi-chain protein as the generative target of fθf_\theta: T=concat⁡[T1,⋯ ,Tm]\mathbf{T} = \operatorname{concat}[T_1, \cdots, T_m], where T1,⋯ ,TmT_1, \cdots, T_m are individual amino acid chains. We say fθf_\theta satisfies chain-permutation equivariance if reordering the input chains leaves the output assembly unchanged up to the same reordering, i.e. fθ(Pπ⋅T)=Pπ⋅fθ(T)f_\theta(P_\pi\cdot\mathbf{T}) = P_\pi\cdot f_\theta(\mathbf{T}) for any π∈Sm\pi\in S_m, where PπP_\pi permutes the mm chain blocks while preserving residue order within each chain, i.e. Pπ⋅concat⁡[T1,…,Tm]=concat⁡[Tπ(1),…,Tπ(m)]P_\pi\cdot\operatorname{concat}[T_1,\ldots,T_m] = \operatorname{concat}[T_{\pi(1)},\ldots,T_{\pi(m)}]. This is achieved by enforcing internal node and pair representations to satisfy h(Pπ⋅T)=Pπ⋅h(T)h(P_\pi\cdot\mathbf{T}) = P_\pi\cdot h(\mathbf{T}) and s(Pπ⋅T)=Pπ⋅s(T)⋅PπTs(P_\pi\cdot\mathbf{T}) = P_\pi\cdot s(\mathbf{T}) \cdot P_\pi^{\mathsf{T}}, respectively. In practice, some models break chain-permutation-equivariance and thus we discuss the effect of the non-strict equivalence in Section B.2.

Symmetry groups. Let GG be a subgroup of SE⁡(3)\operatorname{SE}(3) acting on R3\mathbb{R}^3. We consider two classes:

  • Finite point groups: cyclic CnC_n, dihedral DnD_n, and polyhedral TT, OO, II (with orders 12, 24, 60, respectively).

  • Infinite discrete groups: translational lattices, screw-axis groups, and their semi-direct products with point groups; generating periodic chains, planar lattices, helical filaments, and space-group structures. Experiments in this class are limited since backbone models fall highly out-of-distribution.

We write G={g1=Id⁡,g2,…,gk}G = \{g_1 = \operatorname{Id}, g_2, \ldots, g_k\} with ∣G∣=k|G| = k for finite groups, and in practice approximate infinite group orbits by a finite neighborhood as discussed in Appendix § 9.§ 3.

Canonical subunit, phantom copies, and the quotient space. We refer to a specific single representative of an assembly as the canonical subunit T(1)∈SE⁡(3)0N\mathbf{T}^{(1)} \in\operatorname{SE}(3)^N_0, and to its image gi⋅T(1)g_i \cdot\mathbf{T}^{(1)} under each non-identity group element gig_i as a phantom copy. The corresponding GG-symmetric assembly is the orbit

Tsym=[g1⋅T(1),…,gk⋅T(1)]∈SE⁡(3)kN.(1)\mathbf{T}^{\mathrm{sym}} = [g_1 \cdot\mathbf{T}^{(1)}, \ldots, g_k \cdot\mathbf{T}^{(1)}] \in\operatorname{SE}(3)^{kN}. \tag*{(1)}

For the sake of notational clarity, we mostly consider the case where ∣G∣=k<∞|G| = k < \infty. The key observation is that a GG-symmetric assembly is completely and uniquely determined by the group GG together with a single canonical subunit. This establishes a bijection between GG-symmetric assemblies and a single subunit existing within the quotient space SE⁡(3)0kN/G\operatorname{SE}(3)^{kN}_0/G, where equivalent elements under symmetric group actions are collapsed into a single copy. The next section builds upon this intuition, that generating a GG-symmetric assembly is equivalent to sampling over SE⁡(3)0N/G\operatorname{SE}(3)^N_0/G, a space of dimension kk times smaller than the ambient SE⁡(3)kN\operatorname{SE}(3)^{kN} of full-complex generation.

Equivalence of Generation Strategies

Applying our quotient space view, we propose three strategies for generating symmetric structures:

  1. Full-complex generation: a model operates on the full kNkN-residue assembly Tsym\mathbf{T}^{\mathrm{sym}}, updating all copies simultaneously. With augmented symmetric updates, this is the strategy taken by current methods [17] (Watson et al. 2023).

  2. Quotient-space generation: a model operates directly on SE⁡(3)0N/G\operatorname{SE}(3)^N_0/G, taking equivalence classes as inputs and outputs. This is the theoretical ideal as it removes all symmetry redundancy by construction and assumes parameterized access to the quotient manifold.

  3. Phantom-copy generation: a model operates on a single canonical representative T(1)∈SE⁡(3)0N\mathbf{T}^{(1)} \in\operatorname{SE}(3)^N_0, while consulting its phantom copies {gi⋅T(1)}i=2k\{g_i \cdot\mathbf{T}^{(1)}\}_{i=2}^k as auxiliary geometric context. This is the ZeUS approach.

The two relationships of interest are the descent (I)⇒\Rightarrow(II), where a full-assembly model induces a well-defined operator on the quotient space, and the expansion (II)⇒\Rightarrow(III), where this quotient-space operator is constructed concretely through phantom copies. We now discuss both relationships.

Descent from full-complex to quotient-space. SE⁡(3)\operatorname{SE}(3)-equivariance ensures GG-symmetric inputs yield GG-symmetric outputs, and the action of fθf_\theta on a GG-symmetric input is fully determined by what it does to one canonical subunit. This in turn means that fθf_\theta descends to a well-defined map on equivalence classes:

Theorem 1 (Descent of full-complex models to quotient space). Let fθ:SE⁡(3)kN→SE⁡(3)kNf_\theta: \operatorname{SE}(3)^{kN} \to\operatorname{SE}(3)^{kN} be SE⁡(3)\operatorname{SE}(3)-equivariant and chain-permutation-equivariant. Then there exists a unique map

fθ/G:SE⁡(3)0N/G→SE⁡(3)0N/Gf_\theta^{/G} : \operatorname{SE}(3)^N_0/G \to\operatorname{SE}(3)^N_0/G

such that, for every canonical subunit T(1)\mathbf{T}^{(1)} with assembly Tsym=[g1⋅T(1),…,gk⋅T(1)]\mathbf{T}^{\mathrm{sym}} = [g_1 \cdot\mathbf{T}^{(1)}, \ldots, g_k \cdot\mathbf{T}^{(1)}] and every i∈[k]i \in[k],

fθ(Tsym)∣(i,n)=gi⋅fθ/G([T(1)])∣n.(2)\left.f_\theta(\mathbf{T}^{\mathrm{sym}})\right|_{(i,n)} = \left.g_i \cdot f_\theta^{/G}([\mathbf{T}^{(1)}])\right|_n. \tag*{(2)}

What this quotient space model looks like. Theorem 1 guarantees that an idealized quotient-space model fθGf_\theta^G exists. We now describe these operations directly, working component-by-component in quotient space. First, operations in ambient space can act on residues (node features), on pairs of residues (edge features, attention matrices), and more generally on rr-tuples of residues. In the quotient space, however, each node is an equivalence class with kk ambient representatives, so an operation that consumes a pair of quotient-space nodes, for example, corresponds to kk distinct types of pairs in ambient space (one per relative group element g∈Gg \in G; see Figure 1) and must produce kk outputs to capture all of them. Every primitive of fθf_\theta admits such a quotient-space counterpart, and fθGf_\theta^G is the composition of these counterparts. The precise statement of the operation typology and its completeness across SE(3)\mathrm{SE}(3)-equivariant architectures is given in Appendix § 8.§ 1.

Overview of ZEUS

Figure 1. Overview of ZEUS. Left: Pairwise matrices under symmetric input are degenerate. ZEUS realizes O(k)O(k) memory savings by storing the reduced block row representation during model forward. Middle: O(k)O(k) compute savings are realized by computing symmetry-aware attentional updates to the main/canonical subunit only. Right: ZEUS has minimal output degradation compared to full-complex diffusion, with a mean difference in scRMSD of 0.131 over all evaluated models and symmetries.

Expansion from quotient-space to ambient phantom-copy generation. The quotient-space description above operates on equivalence classes, which are hard to parameterize for existing architectures. Expansion recovers a concrete computation by translating each quotient-space operation back into the ambient space. For example, an edge in quotient space can be represented by kk pair of residues, between a canonical representative in T(1)\mathrm{T}^{(1)} and a phantom copy in g⋅T(1)g \cdot\mathrm{T}^{(1)}, where gg cycles through GG. This expansion preserves the architecture of fθf_\theta while focusing memory and compute on the canonical representative. We formalize the construction as an operator fθphf_\theta^{\mathrm{ph}} and state the resulting equivalence as the following theorem.

Theorem 2 (Phantom-copy realization). The phantom-copy operator fθph:SE(3)0N→SE(3)0Nf_\theta^{\mathrm{ph}} : \mathrm{SE}(3)_0^N \to\mathrm{SE}(3)_0^N, constructed by per-operation ambient-space expansion of fθGf_\theta^G as described above, satisfies

fθG([T(1)])=[fθph(T(1))](3)f_\theta^G\left(\left[\mathrm{T}^{(1)}\right]\right) = \left[f_\theta^{\mathrm{ph}}\left(\mathrm{T}^{(1)}\right)\right] \tag*{(3)}

for every representative T(1)∈SE(3)0N\mathrm{T}^{(1)} \in\mathrm{SE}(3)_0^N.

Remark 1. Beyond edge features, each primitive component of fθGf_\theta^G has an ambient-space expansion centered around the canonical chain: a triplet operation becomes k2k^2 ambient triplets, and aggregations sum these contributions back via group-indexed einsum’s. The composition of these expansions is exactly fθphf_\theta^{\mathrm{ph}}. A precise statement of the operation typology and an inductive proof for tuples are given in Appendix § 8.§ 1, while empirical deployments are presented in § 4.

Together, Theorem 1 and Theorem 2 give the practical equivalence underlying ZEUS, reducing full-complex generation fθ(Tsym)f_\theta(\mathrm{T}^{\mathrm{sym}}) to phantom-copy generation fθph(T(1))∣nf_\theta^{\mathrm{ph}}(\mathrm{T}^{(1)})|_n.

Local geometry preservation under quotient. A central implication of Theorem 1 is that a model trained in SE(3)0N\mathrm{SE}(3)_0^N can be applied, zero shot, to generate GG-symmetric assemblies. The quotient space SE(3)0N/G\mathrm{SE}(3)_0^N/G, where each residue of the projection of a GG-symmetric assembly lives, is a manifold whose local geometry is (a.e.) identical to that of the ambient SE(3)\mathrm{SE}(3). The distribution shift therefore lies only in global geometry, and any short-range structural priors the model has internalized transfer without loss.

Methods

We translate Theorems 1 and 2 into concrete computational savings, guided by the central observation that every internal representation of an SE(3)\mathrm{SE}(3)-equivariant model, evaluated on a GG-symmetric assembly, is equivariant under the GG-action, and is therefore duplicated kk times in the full-complex forward pass. Recovering this redundancy yields a uniform O(k)O(k) reduction in both memory and compute, applied across every layer and every operation type as they now act on the canonical subunit alone.

Memory Savings: Reduction of Phantom Copies

  • Node features reduce from kNkN to NN: scalar (invariant) components of the phantom copies are identical to those of the canonical subunit, and equivariant (frame-valued) components are recovered as the group-transformed canonical embeddings under gig_i.

  • Edge features reduce from k2N2k^2N^2 to kN2kN^2: by the same principle, only the kk distinct relative-group blocks incoming to the canonical subunit need be maintained. Phantom-to-phantom edges are never required, since each admits a counterpart involving the canonical chain via gig_i.

Compute Savings: Symmetry-Guided Expansion of Canonical Chain

The runtime counterpart to § 4.1 is to compute updates only for the canonical subunit, which by Theorems 1 and 2 is exact provided all attention-like operations remain phantom-aware. We describe the most salient modifications required for each of the two architecture classes targeted by ZEUS. Full details of additional modifications to architectures can be found in Appendix § 11.

IPA architectures (FrameDiff, FoldFlow). We modify the IPA layer to efficiently pass messages from phantom residues into the canonical subunit. To compute the attention softmax over all kk group transforms of the kNkN residues, we only need queries from the canonical subunit, while keys need to come from all copies. Invariant scalar keys are tied across copies, but point-valued keys are equivariant and must be transformed by the group action. Dropping constants for brevity, the attention logit becomes

ai,j←Softmax⁡j∈{1,…,kN}, g∈G(qiTkj+bi,j−∑p∥Ti∘q⃗ip−g∘Tj∘k⃗jp∥2)(4)a_{i,j} \leftarrow\operatorname{Softmax}_{j\in\{1,\ldots,kN\},\,g\in G}\left(\mathbf{q}_i^{\mathsf{T}}\mathbf{k}_j+b_{i,j}-\sum_p\left\lVert T_i\circ\vec{\mathbf{q}}_i^{p}-g\circ T_j\circ\vec{\mathbf{k}}_j^{p}\right\rVert^2\right) \tag*{(4)}

where ii ranges over the residues in the canonical subunit and jj ranges over all residues in the complex.

Axial-attention architectures (RFdiffusion). Pair-track axial attention [Watson et al., 2023] is the most memory-intensive block in the RoseTTAFold / RFdiffusion architecture. Standard Axial Attention initiates its sequences on the full (kN)×(kN)(kN)\times(kN) pair-wise tensor and thus requires all k2N2k^2N^2 elements. As discussed in § 4.3, however, the canonical block-row contains all distinct pair interactions, as every phantom-centered block is a block-column permutation. We replace axial attention with an implementation that consumes only the canonical block-row and computes the full pair update via group-indexed shuffles, never materializing the (kN)×(kN)(kN)\times(kN) tensor:

ai,j=Softmax⁡j(∑g∈G[(πg(Q))T(πg(K))1+((j−1)(modN))]+bi,j)(5)a_{i,j}=\operatorname{Softmax}_{j}\left(\sum_{g\in G}\left[(\pi_g(\mathbf{Q}))^{\mathsf{T}}(\pi_g(\mathbf{K}))_{1+((j-1)\pmod N)}\right]+b_{i,j}\right) \tag*{(5)}

where ii ranges over the residues in the canonical subunit and jj ranges over all residues in the complex. Q\mathbf{Q} and K\mathbf{K} are the block-row pair query and key matrices. πg(X)\pi_g(X) denotes the column permutation operation of a block-row XX by the group permutation matrix determined through gg. The output is exactly equivalent to the full axial update, and computation never exceeds O(kN2)O(kN^2) pair memory.

Block Structure of Pair-wise Interaction Matrices

Here we expand upon the discussion on axial-attention by studying pair-wise feature matrices. Quotient to phantom equivalence implies a specific exploitable structure in any pair-wise interaction matrix. Indexing residues as (i,n)(i,n) for chain-i∈[k]i\in[k] and position n∈[N]n\in[N], we arrange a kN×kNkN\times kN interaction matrix AA in N×NN\times N blocks A[i,j]A^{[i,j]} collecting entries from chain-ii’s residue to chain-jj’s.

Theorem 3 (Block Structure of Interaction Matrices). Let AA be a pair-wise interaction matrix, featured within an SE(3)-equivariant model evaluated on a GG-symmetric assembly, whose entries are all SE(3)-invariant pairwise features. Then for any i,j∈[k]i,j\in[k], A[i,j]=A[1,idx⁡(gi−1gj)]A^{[i,j]}=A^{[1,\operatorname{idx}(g^{-1}_{i}g_j)]}, where idx⁡:G→[k]\operatorname{idx}:G\to[k] is the inverse of group element indexing.

By the block-permutation theorem, the block-row ii associated with g∈Gg\in G satisfies $A^{[i,:]}=\pi_g(A^{[1,:]}).

A proof is given in Appendix § 8.§ 2. Theorem 3 implies a block-wise permutation pattern for pair-wise interaction matrices specific to the symmetry group GG, where each row of blocks is a permutation of the first row. For abelian groups, consecutive rows are cyclic shifts of the first row; for non-abelian groups, the permutation is determined by the group multiplication table. A visualization is in Figure 1, and more extensive visualizations are in Appendix § 10.

Experiments

Symmetric Generation Quality

By Theorem 1, ZEUS produces samples equivalent to full-complex symmetric diffusion under pure SE(3)-invariant attention and chain-permutation equivariance. We now empirically characterize the equivalence of generation quality across architectures, symmetry groups, and design metrics and the efficiency of the observed runtime and memory savings.

Setup. We benchmark three backbone models across two architectural classes: IPA-based FrameDiff and FoldFlow, and axial attention based RFdiffusion. For each model we unconditionally generate 200 designs at five cyclic symmetries (C2C_2, C3C_3, C4C_4, C6C_6, C8C_8), and 100 designs at C12C_{12} with subunit length N=50N = 50, yielding total complex sizes from 100 to 600 residues. In practice, model design choices can violate chain-permutation equivariance. FrameDiff and FoldFlow were trained on single chain input. We implement symmetric multi chain inference that respects this invariance. RFdiffusion is not chain-permutation equivariant and thus we implement a full-complex diffusion ablation that enforces chain-permutation equivariance during inference, but without any ZEUS reductions. We compare ZEUS with both the default RFdiffusion full-complex inference and this ablated inference to distinguish the effect of symmetry reduction from that of modifying the pretrained model. In addition, we discuss the theoretical implications in Appendix § 9.§ 2

Each configuration is run with full-complex diffusion and with ZeUS, and an additional chain-permutation ablation for RFdiffusion. All experiments were run on a single NVIDIA A100-40GB GPU.

Metrics. We report three classes of generation-quality metrics, following prior work.

Designability. For each backbone design we predict 8 sequences with ProteinMPNN [11], refold each with ESMFold [23], and measure the self-consistency RMSD (scRMSD) and TM-score (scTM) against the original backbone. A backbone is designable if its best scRMSD falls below 2 Å or its best scTM exceeds 0.5. We report the fraction of designable structures together with the mean best scTM and scRMSD over all refolded designs. FrameDiff and FoldFlow are not trained on multi-chain structures, so their baseline designability rates on symmetric assemblies are modest. The mean scTM and scRMSD complement the fraction designable by enabling a more sensitive comparison.

Novelty. We measure novelty as the maximum TM-score against the PDB for designable complexes only, retrieved via Foldseek [Van Kempen et al., 2024]. A lower TM-score means lower structural similarity to the PDB and is thus desirable. We report it at both the subunit and full-complex levels, since these capture independent aspects of complex novelty.

Diversity. Following FoldFlow [4], we report the mean pairwise TM-score among designable complexes only, computed at the full-complex level.

Efficiency. We report peak GPU memory during a denoising step (torch.cuda.max_memory_allocated) and average wall-clock time per denoising step.

Equivalence across architectures and symmetries. Table 1 shows that ZeUS achieves equivalent design quality to the chain-permutation equivariant RFdiffusion ablation, with a mean signed scRMSD difference of only 0.04 Å. While there is some degradation in both designability and scRMSD when moving from full-complex diffusion to the ablation, this degradation is attributable to the ablation itself rather than ZeUS. Our results show that ZeUS preserves the design quality of the corresponding ablated model.

SymmetryDesignability (%) ↑ fullDesignability (%) ↑ ablationDesignability (%) ↑ ZeUSDesignability (%) ↑ Δ (pp)mean best scRMSD (Å) ↓ fullmean best scRMSD (Å) ↓ ablationmean best scRMSD (Å) ↓ ZeUSmean best scRMSD (Å) ↓ Δ (Å)
C2C_297.096.595.5−1.01.040.900.89−0.01
C3C_382.076.076.00.01.691.941.76−0.18
C4C_467.061.554.5−7.02.432.392.72+0.33
C6C_657.051.047.0−4.03.533.833.98+0.16
C8C_848.541.046.5+5.54.705.045.10+0.06
C12C_{12}51.036.040.0+4.05.686.045.91−0.12

Table 1. RFdiffusion Designability (frac. designable, scRMSD) quality across cyclic symmetry groups. Full, Ablation, and ZeUS denote full-complex inference, the chain-invariance ablation, and ZeUS inference, respectively. The Δ columns report the difference Δ = ZeUS − Ablation.

Similarly, Table 2 shows that diversity and novelty remain largely unchanged between ZeUS and full-complex diffusion across all evaluated models, with mean signed differences of −0.0053 in diversityTM and −0.0076 in pdbTM, respectively. Together, these results are consistent with the equivalence established in Theorem 1, across different architectures and symmetry groups.

ModelSymmetrydiversityTM ↓\downarrow fulldiversityTM ↓\downarrow ZeUSdiversityTM ↓\downarrow Δ\DeltapdbTM (novelty) ↓\downarrow fullpdbTM (novelty) ↓\downarrow ZeUSpdbTM (novelty) ↓\downarrow Δ\Delta
FrameDiffC2C_20.2230.231+0.0080.1570.132−0.025
FrameDiffC3C_30.2260.202−0.0240.0970.078−0.019
FrameDiffC4C_40.2370.232−0.0050.1030.068−0.035
FrameDiffC6C_60.2600.286+0.0260.0590.044−0.015
FrameDiffC8C_80.2990.292−0.0070.0470.026−0.022
FrameDiffC12C_{12}0.374––0.0330.020−0.013
FoldFlowC2C_20.2490.247−0.0010.1260.130+0.003
FoldFlowC3C_30.2140.215+0.0010.0690.075+0.005
FoldFlowC4C_40.2190.220+0.0010.0580.046−0.012
FoldFlowC6C_60.2630.263−0.0010.0540.040−0.015
FoldFlowC8C_80.3200.283−0.0370.0210.035+0.014
FoldFlowC12C_{12}0.3950.368−0.0270.0170.018+0.001
RFdiffusionC2C_20.2520.260+0.0080.0370.038+0.001
RFdiffusionC3C_30.2290.225−0.0040.0340.031−0.002
RFdiffusionC4C_40.2550.243−0.0110.0240.026+0.002
RFdiffusionC6C_60.2590.244−0.0150.0170.0170.000
RFdiffusionC8C_80.2710.268−0.0030.0130.012−0.001
RFdiffusionC12C_{12}0.3040.306+0.0010.0100.006−0.004

Table 2. Complex-level diversity and novelty of designable structures. Lower values indicate greater diversity and novelty, respectively. The Δ\Delta columns report the difference Δ=ZeUS−Full\Delta= \mathrm{ZeUS} - \mathrm{Full}.

Table 3 in Appendix F provides comprehensive results across all models and symmetries. Designability, diversity, and novelty agree within 1σ1\sigma between both methods across all six cyclic symmetries.

modelsymmDesignability (↑\uparrow) fullDesignability (↑\uparrow) ZEUSmean best scRMSD (↓\downarrow) fullmean best scRMSD (↓\downarrow) ZEUSdiversityTMcomplex_{\text{complex}} (↓\downarrow) fulldiversityTMcomplex_{\text{complex}} (↓\downarrow) ZEUSpdbTMcomplex_{\text{complex}} (↓\downarrow) fullpdbTMcomplex_{\text{complex}} (↓\downarrow) ZEUS
FrameDiffC2C_20.6800.7602.026±1.0372.026 \pm1.0372.060±0.8912.060 \pm0.8910.223±0.0480.223 \pm0.0480.231±0.0540.231 \pm0.0540.157±0.0600.157 \pm0.0600.132±0.0540.132 \pm0.054
FrameDiffC3C_30.2800.2003.471±0.8943.471 \pm0.8943.356±1.0503.356 \pm1.0500.226±0.0520.226 \pm0.0520.202±0.0420.202 \pm0.0420.097±0.0280.097 \pm0.0280.078±0.0270.078 \pm0.027
FrameDiffC4C_40.1600.2203.348±0.8813.348 \pm0.8813.427±0.9353.427 \pm0.9350.237±0.0610.237 \pm0.0610.232±0.0620.232 \pm0.0620.103±0.0580.103 \pm0.0580.068±0.0480.068 \pm0.048
FrameDiffC6C_60.0800.1403.310±0.8923.310 \pm0.8923.413±0.9993.413 \pm0.9990.260±0.0670.260 \pm0.0670.286±0.0690.286 \pm0.0690.059±0.0440.059 \pm0.0440.044±0.0210.044 \pm0.021
FrameDiffC8C_80.2000.1403.028±0.9043.028 \pm0.9043.119±0.9383.119 \pm0.9380.299±0.0750.299 \pm0.0750.292±0.0910.292 \pm0.0910.047±0.0290.047 \pm0.0290.026±0.0080.026 \pm0.008
FrameDiffC12C_{12}0.2200.0203.013±0.9543.013 \pm0.9544.154±1.0714.154 \pm1.0710.374±0.0910.374 \pm0.091–0.033±0.0150.033 \pm0.0150.020
FoldFlowC2C_20.8400.8201.838±0.7531.838 \pm0.7531.686±0.7391.686 \pm0.7390.249±0.0510.249 \pm0.0510.247±0.0410.247 \pm0.0410.126±0.0600.126 \pm0.0600.130±0.0640.130 \pm0.064
FoldFlowC3C_30.2200.2803.336±0.9243.336 \pm0.9243.029±0.9513.029 \pm0.9510.214±0.0530.214 \pm0.0530.215±0.0470.215 \pm0.0470.069±0.0250.069 \pm0.0250.075±0.0190.075 \pm0.019
FoldFlowC4C_40.1800.1202.983±0.8992.983 \pm0.8993.285±0.7123.285 \pm0.7120.219±0.0420.219 \pm0.0420.220±0.0490.220 \pm0.0490.058±0.0250.058 \pm0.0250.046±0.0220.046 \pm0.022
FoldFlowC6C_60.1400.1603.202±0.7773.202 \pm0.7772.964±0.8902.964 \pm0.8900.263±0.0640.263 \pm0.0640.263±0.0770.263 \pm0.0770.054±0.0380.054 \pm0.0380.040±0.0180.040 \pm0.018
FoldFlowC8C_80.1400.1202.957±0.7142.957 \pm0.7143.062±0.7343.062 \pm0.7340.320±0.0840.320 \pm0.0840.283±0.0680.283 \pm0.0680.021±0.0070.021 \pm0.0070.035±0.0190.035 \pm0.019
FoldFlowC12C_{12}0.1400.1602.922±0.8632.922 \pm0.8632.947±0.6542.947 \pm0.6540.395±0.0790.395 \pm0.0790.368±0.1080.368 \pm0.1080.017±0.0120.017 \pm0.0120.018±0.0060.018 \pm0.006
RFdiffusionC2C_20.9700.9551.036±0.7761.036 \pm0.7760.889±0.6790.889 \pm0.6790.252±0.0770.252 \pm0.0770.260±0.0800.260 \pm0.0800.037±0.0190.037 \pm0.0190.038±0.0180.038 \pm0.018
RFdiffusionC3C_30.8200.7601.694±1.3671.694 \pm1.3671.758±1.4451.758 \pm1.4450.229±0.0810.229 \pm0.0810.225±0.0730.225 \pm0.0730.034±0.0190.034 \pm0.0190.031±0.0150.031 \pm0.015
RFdiffusionC4C_40.6700.5452.429±1.7122.429 \pm1.7122.722±1.5912.722 \pm1.5910.255±0.0960.255 \pm0.0960.243±0.0880.243 \pm0.0880.024±0.0160.024 \pm0.0160.026±0.0190.026 \pm0.019
RFdiffusionC6C_60.5700.4723.534±2.0463.534 \pm2.0463.984±1.8283.984 \pm1.8280.259±0.1060.259 \pm0.1060.244±0.0900.244 \pm0.0900.017±0.0110.017 \pm0.0110.017±0.0120.017 \pm0.012
RFdiffusionC8C_80.4850.4654.703±1.9754.703 \pm1.9755.105±1.7475.105 \pm1.7470.271±0.0980.271 \pm0.0980.268±0.1030.268 \pm0.1030.013±0.0070.013 \pm0.0070.012±0.0070.012 \pm0.007
RFdiffusionC12C_{12}0.5100.4005.681±1.8655.681 \pm1.8655.914±1.9215.914 \pm1.9210.304±0.1140.304 \pm0.1140.306±0.1230.306 \pm0.1230.010±0.0080.010 \pm0.0080.006±0.0030.006 \pm0.003

Table 3. Designability (frac. designable, scRMSD), Diversity (TMcomplexTM_{\mathrm{complex}}), and Novelty (pdbTMcomplexpdbTM_{\mathrm{complex}}) results for cyclic generation for FrameDiff, FoldFlow, and RFdiffusion. Row labels CkC_k denote symmetry group with subunit length n=50n = 50. For each metric, paired columns report the full-complex baseline and ZEUS. All reported uncertainties are 1-sigma standard deviations.

Memory and runtime scale linearly in kk. Figure 2 shows that for a fixed memory budget, the total number of residues ZEUS can fit in memory grows as k\sqrt{k}, relative to the baseline. At C50C_{50}, this enables generation of complexes up to ∼16,000\sim16{,}000 residues, compared to ∼2,400\sim2{,}400 under the full-complex baseline. The runtime curve follows the same pattern, with the per-step advantage of ZEUS growing with kk. Conversely this shows the O(k)O(k) linear reduction predicted in § 4.1 when NN is kept fixed.

GPU VRAM usage and forward pass wall-clock time

Figure 2. Left: GPU VRAM usage. Right: forward pass wall-clock time for the full-complex baseline compared to ZEUS, at varying cyclic group sizes. Models are run until out-of-memory, with the largest NresN_{\mathrm{res}} that fits in memory marked.

Polyhedral symmetries. We do not report designability metrics for polyhedral generations as self-consistency-based designability is currently unreliable for polyhedral assemblies. Current established folding models struggle to correctly predict the structure of large symmetric complexes [Zhu et al., 2023]. This implies that we cannot determine whether a given design under TT, OO, or II symmetry is genuinely undesignable, or whether the folding model itself is the bottleneck. Nevertheless we report distributional comparisons of polyhedral generations with the baseline full-complex and ZEUS, available in Appendix § 15. Finally, ZEUS results remain distributionally similar for additional metrics such as steric clashes, subunit-level novelty, and secondary structure composition, available in Appendix § 14.

Generated I, D50, H, C50 designs with their interchain contacts

Figure 4. Generated II, D50D_{50}, HH, C50C_{50} designs with Nres=6000N_{\mathrm{res}} = 6000, 10000, 4500, 10000 respectively, displayed along with their interchain contacts. Contacts respect the imposed symmetry in each case.

Large Scale Design

Generation of large-scale biomolecular complexes (>∼1000>\sim1000 residues) is an active frontier of de novo design [14, 5]. At this size, existing generative models exhibit substantially higher failure rates and are unable to calculate full attention on all residues on a single GPU [21]. ZEUS’s quotient space reformulation makes this regime tractable to compute exact, untruncated attention over the full symmetric complex. To showcase this, we use ZEUS on FrameDiff to generate large C50C_{50}, D50D_{50}, II (Icosahedral), and HH (Helical) complexes, at sizes of 10,000, 10,000, 6000, and 4500 total residues, respectively.

The generated designs are structurally coherent, with interchain interfaces that respect the imposed symmetry and characteristic motifs emerging per symmetry class. Helical assemblies are dominated by α\alpha-helices aligned along the tilt axis, mirroring both natural helical proteins and prior de novo helical designs [Shen et al., 2018], while dihedral interfaces respect both the cyclic and stacking components of D50D_{50}.

Normalized buried surface area of ZEUS generations compared to PDB proteins matched by symmetry

Figure 3. Normalized buried surface area of ZEUS generations compared to PDB proteins matched by symmetry.

To probe the global plausibility of our generated complexes, we calculate the buried surface area of each subunit protein normalized by the unbound solvent-accessible surface area as a proxy for interchain contact. We find that the property distributions of our large-scale designs closely overlap with natural PDB proteins in each symmetry class. In addition to being globally plausible, the designs themselves remain geometrically distinct from any PDB structure as measured by radius and present reasonable alternative large-scale geometries (closest PDB matches reported in Appendix § 17). We note that these are backbone-only structures, and assessing designability is intractable for such proposed structure at this scale [5]. Our contribution here is demonstrating that existing backbone generators can produce coherent symmetric complexes at a previously unattainable scale; improving validity in this regime is left to future work.

Conclusions, Limitations, and Future Work

We introduced ZEUS, a method that converts any SE(3)\mathrm{SE}(3)-equivariant and chain-permutation-equivariant backbone generative model into a symmetric assembly generator with lossless O(k)O(k) reductions in memory and runtime. The central insight is to formulate symmetric backbone design as a generative problem within a quotient space. We implement this insight for the two dominant backbone architecture classes, IPA-based models (FrameDiff, FoldFlow) and the axial-attention-based RFdiffusion. Empirically, ZEUS matches the designability, diversity, and novelty of full-complex symmetric inference across all settings tested while achieving the predicted O(k)O(k) savings. The same reduction unlocks generation of symmetric complexes an order of magnitude larger than feasible under any explicit-symmetrization method, without truncation or approximation. We acknowledge that the equivalence of transfer is not as strong when the base model is not fully compliant, nevertheless we anticipate that fine-tuning on symmetric designs and tightening the SE(3)\mathrm{SE}(3)-equivariance and chain-permutation equivariance of existing architectures will be natural next steps toward even more accurate and scalable symmetric complex design.

Acknowledgments and Disclosure of Funding

This work was partially supported by the National Institutes of Health (OT2OD038003, U54OD036472, U54DK097771), the National Science Foundation (2312501, 2106859), and Amazon, NEC, and the Easton Center.

Theoretical Details

Proof of Theorems 1 and 2: Equivalence of Generation Strategies

This appendix provides full proofs of Theorems 1 and 2. Appendix A.1.2 establishes the descent (I)⇒\Rightarrow(II), proving that an SE(3)-equivariant and chain-permutation-equivariant full-complex model induces a unique well-defined operator on the quotient space. Appendix A.1.3 introduces the operation typology used by both proofs, characterizing the primitives that compose any SE(3)-equivariant graph network in terms of rr-tuple operations on residue frames. Appendix A.1.4 establishes the expansion (II)⇒\Rightarrow(III) by constructing the phantom-chain operator primitive-by-primitive and proving its equivalence with fθ/Gf_\theta^{/G} on canonical representatives.

Setup and Notation

Let fθ:SE(3)kN→SE(3)kNf_\theta:\mathrm{SE}(3)^{kN}\to\mathrm{SE}(3)^{kN} be a backbone model, where the input T=[T1,…,Tk]\mathbf{T}=[\mathbf{T}_1,\ldots,\mathbf{T}_k] can be separated into kk distinct components of size NN (i.e., every chain belongs to a single component in our context). We say fθf_\theta is SE(3)\mathrm{SE}(3)-equivariant if fθ(g⋅T)=g⋅fθ(T)f_\theta(g\cdot\mathbf{T})=g\cdot f_\theta(\mathbf{T}) for every g∈SE(3)g\in\mathrm{SE}(3), where the action is element-wise on the kNkN residues. We say fθf_\theta is chain-permutation-equivariant if fθ(σ⋅T)=σ⋅fθ(T)f_\theta(\sigma\cdot\mathbf{T})=\sigma\cdot f_\theta(\mathbf{T}) for every permutation σ∈Sk\sigma\in\mathcal{S}_k of the kk component blocks of T∈SE(3)kN\mathbf{T}\in\mathrm{SE}(3)^{kN}. Let G={g1=Id,g2,…,gk}G=\{g_1=\mathrm{Id},g_2,\ldots,g_k\} be a finite symmetry group, and write Tsym=(g1⋅T(1),…,gk⋅T(1))\mathbf{T}^{\mathrm{sym}}=(g_1\cdot\mathbf{T}^{(1)},\ldots,g_k\cdot\mathbf{T}^{(1)}) for the GG-symmetric assembly generated by a canonical subunit T(1)∈SE(3)N\mathbf{T}^{(1)}\in\mathrm{SE}(3)^N. Equivalence classes [T(1)][\mathbf{T}^{(1)}] in SE(3)0N/G\mathrm{SE}(3)^N_0/G are taken under the relation T∼T′\mathbf{T}\sim\mathbf{T}' if and only if T′=g⋅T\mathbf{T}'=g\cdot\mathbf{T} for some g∈Gg\in G.

Proof of Theorem 1: Descent

The proof proceeds in three steps: first, we show that fθ(Tsym)f_\theta(\mathbf{T}^{\mathrm{sym}}) is itself GG-symmetric, allowing the equivalence class of any single chain to represent the entire output; second, we use this to construct fθ/Gf_\theta^{/G} on equivalence classes; third, we verify that the construction is independent of the choice of representative and is the unique map satisfying (2).

Step 1: GG-symmetry of the output. The construction of fθ/Gf_\theta^{/G} requires that fθ(Tsym)f_\theta(\mathbf{T}^{\mathrm{sym}}) be GG-symmetric, since taking the equivalence class of a single copy is well-defined only when all copies are GG-related. For any g∈Gg\in G, the diagonal action of gg on Tsym\mathbf{T}^{\mathrm{sym}} produces g⋅Tsym=(gg1⋅T(1),…,ggk⋅T(1))g\cdot\mathbf{T}^{\mathrm{sym}}=(gg_1\cdot\mathbf{T}^{(1)},\ldots,gg_k\cdot\mathbf{T}^{(1)}), which by closure of GG under multiplication is a permutation of the original orbit; concretely, g⋅Tsym=πg⋅Tsymg\cdot\mathbf{T}^{\mathrm{sym}}=\pi_g\cdot\mathbf{T}^{\mathrm{sym}} where πg∈Sk\pi_g\in\mathcal{S}_k is the permutation defined by ggi=gπg(i)gg_i=g_{\pi_g(i)}. Applying fθf_\theta to both sides and using SE(3)-equivariance on the left and chain-permutation-equivariance on the right yields

g⋅fθ(Tsym)=πg⋅fθ(Tsym).(6)g\cdot f_\theta(\mathbf{T}^{\mathrm{sym}})=\pi_g\cdot f_\theta(\mathbf{T}^{\mathrm{sym}}). \tag*{(6)}

Reading off the ii-th block of each side, the ii-th copy of the output equals gg applied to the πg−1(i)\pi_g^{-1}(i)-th copy, so the output orbit is closed under GG and fθ(Tsym)f_\theta(\mathbf{T}^{\mathrm{sym}}) is itself GG-symmetric.

Step 2: Construction of fθ/Gf_\theta^{/G}. Define fθ/G:SE(3)0N/G→SE(3)0N/Gf_\theta^{/G}:\mathrm{SE}(3)^N_0/G\to\mathrm{SE}(3)^N_0/G as follows. Given an equivalence class [T(1)][\mathbf{T}^{(1)}], choose any representative T(1)\mathbf{T}^{(1)}, form the orbit Tsym\mathbf{T}^{\mathrm{sym}}, apply fθf_\theta to obtain the updated assembly fθ(Tsym)f_\theta(\mathbf{T}^{\mathrm{sym}}), and set fθ/G([T(1)])=[fθ(Tsym)∣(1,⋅)]f_\theta^{/G}([\mathbf{T}^{(1)}])=[f_\theta(\mathbf{T}^{\mathrm{sym}})|_{(1,\cdot)}], the equivalence class of the first canonical chain. By Step 1, the output is GG-symmetric, so the choice of which copy to read off is immaterial: any other copy yields the same equivalence class.

Step 3: Well-definedness and uniqueness. We verify that fθ/Gf_\theta^{/G} is independent of the choice of representative T(1)\mathbf{T}^{(1)} within [T(1)][\mathbf{T}^{(1)}]. Suppose T(1)′=h⋅T(1)\mathbf{T}^{(1)\prime}=h\cdot\mathbf{T}^{(1)} for some h∈Gh\in G. The orbit from T(1)′\mathbf{T}^{(1)\prime} is Tsym′=(g1h⋅T(1),…,gkh⋅T(1))\mathbf{T}^{\mathrm{sym}\prime}=(g_1h\cdot\mathbf{T}^{(1)},\ldots,g_kh\cdot\mathbf{T}^{(1)}), which by closure of GG is a permutation σh\sigma_h of the original orbit Tsym\mathbf{T}^{\mathrm{sym}}. By chain-permutation-equivariance, fθ(Tsym′)=σh⋅fθ(Tsym)f_\theta(\mathbf{T}^{\mathrm{sym}\prime})=\sigma_h\cdot f_\theta(\mathbf{T}^{\mathrm{sym}}), and the first copy of fθ(Tsym′)f_\theta(\mathbf{T}^{\mathrm{sym}\prime}) is therefore some copy of fθ(Tsym)f_\theta(\mathbf{T}^{\mathrm{sym}}), which by Step 1 belongs to the same equivalence class as the first. The output equivalence class is therefore the same regardless of representative, establishing well-definedness. Finally, equation (2) follows directly from (6) by reading off the ii-th block: fθ(Tsym)∣(i,n)=gi⋅fθ(Tsym)∣(1,n)=gi⋅fθ/G([T(1)])∣nf_\theta(\mathbf{T}^{\mathrm{sym}})|_{(i,n)}=g_i\cdot f_\theta(\mathbf{T}^{\mathrm{sym}})|_{(1,n)}=g_i\cdot f_\theta^{/G}([\mathbf{T}^{(1)}])|_n. Uniqueness of fθGf_{\theta}^{G} satisfying (2) follows because the right-hand side determines the equivalence class of the canonical-copy output for every input class.

Operation Typology for Equivariant Graph Networks

The proof of Theorem 2 requires a structural proposition on fθf_{\theta}: that it decomposes into a sequence of SE(3)\mathrm{SE}(3)-equivariant primitive operations on rr-tuples of residue frames. We make this proposition precise here.

Primitive operations. A primitive operation consumes a finite collection of input features, each indexed by an ordered tuple of residue indices, and produces a single output feature indexed by an ordered tuple of residue indices. For rigidity, we generally write the input arities as a multiset {rin(1),…,rin(p)}\{r_{\mathrm{in}}^{(1)},\ldots,r_{\mathrm{in}}^{(p)}\} and the output arity as routr_{\mathrm{out}}, so that the operation maps an input collection

(xn1(1)⋯nrin(1)(1),…,xn1(p)⋯nrin(p)(p))(7)\left(x_{n_{1}^{(1)}\cdots n_{r_{\mathrm{in}}^{(1)}}}^{(1)},\ldots,x_{n_{1}^{(p)}\cdots n_{r_{\mathrm{in}}^{(p)}}}^{(p)}\right) \tag*{(7)}

to an output ym1⋯mrouty_{m_{1}\cdots m_{r_{\mathrm{out}}}}. Each output index ranges independently over residues (a “for each” over the corresponding axis); indices that appear in some input tuple but not in the output tuple are contracted by a pooling operation (sum, mean, max, etc.) over the residues in the missing axes, altogether definable as an einsum-like operation.

This typology, while abstruse, covers all standard primitives used in SE(3)\mathrm{SE}(3)-equivariant backbone models, with mostly low tuple dimensionalities:

  • A pointwise node map is the case p=1p=1, rin(1)=rout=1r_{\mathrm{in}}^{(1)}=r_{\mathrm{out}}=1 with no contraction.

  • An edge construction from a pair of node features is p=2p=2, rin(1)=rin(2)=1r_{\mathrm{in}}^{(1)}=r_{\mathrm{in}}^{(2)}=1, rout=2r_{\mathrm{out}}=2, also without contraction: the output edge ynmy_{nm} is built from the node features xn(1)x_{n}^{(1)} and xm(2)x_{m}^{(2)} via an SE(3)\mathrm{SE}(3)-equivariant function of frames TnT_{n} and TmT_{m}.

  • An edge-to-edge transformation is p=1p=1, rin(1)=rout=2r_{\mathrm{in}}^{(1)}=r_{\mathrm{out}}=2.

  • Standard attention is p=2p=2 with one input feature of arity 2 (the attention logits or kernel αnm\alpha_{nm}) and one of arity 1 (the values vmv_m), and output arity rout=1r_{\mathrm{out}}=1 (the aggregated node update); the index nn is preserved as a “for each” axis, and the index mm is contracted by attention-weighted summation yn=∑mαnmvmy_n=\sum_m\alpha_{nm}v_m.

  • Triplet operations such as those appearing in axial attention have rin=3r_{\mathrm{in}}=3 for some input feature, and may either preserve the triplet structure (rout=3r_{\mathrm{out}}=3) or collapse it to a pair or node by contraction.

Proposition. It is plain to see that all graph-based models operating on tensors can be decomposed into a string of these primitive operations. We therefore assume throughout that fθf_{\theta} decomposes as a finite composition of such primitive operations.

This also proposes an important new perspective for Theorem 1: the descent of the compound model fθf_{\theta} onto quotient space fθGf_{\theta}^{G} is the same as the descent and re-composition of each individual component in fθf_{\theta}. Proof immediately follows from the same line of reasoning on each SE(3)\mathrm{SE}(3)-equivariant primitive operation.

Proof of Theorem 2: Expansion

We construct the phantom-copy operator fθphf_{\theta}^{\mathrm{ph}} by giving the ambient-space expansion of each primitive operation in fθGf_{\theta}^{G}, then prove the equivalence (3) by induction on the composition.

Step 1: Group-indexed features in the quotient space. Before expanding individual primitives, we describe the data layout of features in the quotient-space picture and its ambient-space realization. A feature of arity rr in the ambient space is indexed by an ordered tuple of residues (n1,…,nr)(n_1,\ldots,n_r). On the quotient space, each residue index becomes an equivalence class with kk ambient representatives, so an rr-tuple of equivalence classes corresponds to krk^r distinct ambient tuples; gauge-fixing the first index to a canonical representative leaves kr−1k^{r-1} distinct ambient tuples, indexed by relative group elements in Gr−1G^{r-1}. Concretely, a quotient-space feature of arity rr is realized in the ambient space as a feature indexed by (n1,…,nr;g2,…,gr)(n_1,\ldots,n_r;g_2,\ldots,g_r), where the first rr indices range over canonical-copy residues and the remaining r−1r-1 group indices specify which phantom-copy representatives provide the second through rr-th frames in the implicit ambient tuple (T(1,n1),g2⋅T(1,n2),…,gr⋅T(1,nr))(\mathbf{T}^{(1,n_1)},g_2\cdot\mathbf{T}^{(1,n_2)},\ldots,g_r\cdot\mathbf{T}^{(1,n_r)}). We refer to this as the group-indexed representation of the quotient-space feature, and write x~n1⋯nrg2,…,gr\widetilde{x}^{g_2,\ldots,g_r}_{n_1\cdots n_r} for its components.

Step 2: Ambient-space expansion of a single primitive. Consider a primitive operation PP with input arities {rin(1),…,rin(p)}\{r_{\mathrm{in}}^{(1)},\ldots,r_{\mathrm{in}}^{(p)}\} and output arity routr_{\mathrm{out}} in fθf_\theta, taking inputs x(1),…,x(p)x^{(1)},\ldots,x^{(p)} and producing output yy. The corresponding quotient-space primitive P/GP/G takes group-indexed inputs x~(i)\widetilde{x}^{(i)} with rin(i)−1r_{\mathrm{in}}^{(i)}-1 group-element axes each, and produces a group-indexed output y~\widetilde{y} with rout−1r_{\mathrm{out}}-1 group-element axes. Its ambient-space expansion PphP^{\mathrm{ph}} proceeds as follows:

  1. For each output residue tuple (m1,…,mrout)(m_1,\ldots,m_{r_{\mathrm{out}}}) and output group-index tuple (h2,…,hrout)(h_2,\ldots,h_{r_{\mathrm{out}}}), evaluate PP on the ambient representatives implied by these indices: specifically, instantiate each input x(i)x^{(i)} at the ambient tuple of frames specified by the appropriate canonical-copy residues and group elements, gauge-fixed so that the first input index aligns with the canonical-copy output index m1m_1.

  1. Indices that appear in some input but not in the output are contracted by the pooling operation specified by PP, summed (or otherwise reduced) over the residues in those axes. Crucially, indices that appear as group-element axes in some input but not in the output must also be contracted; this corresponds to summing over the corresponding group elements in GG, weighted by the original pooling rule.

  1. The resulting value is stored at y~m1⋯mrouth2,…,hrout\widetilde{y}^{h_2,\ldots,h_{r_{\mathrm{out}}}}_{m_1\cdots m_{r_{\mathrm{out}}}}.

We claim that for any GG-symmetric input Tsym\mathbf{T}^{\mathrm{sym}}, the output of PP applied to the full assembly, restricted to output tuples whose first index lies in the canonical copy, equals the output of PphP^{\mathrm{ph}} applied to the canonical representative. The argument has two parts.

(a) Per-tuple equivariance. For any ambient input tuple of the form (gi1⋅T(1,n1),…,gir⋅T(1,nr))(g_{i_1}\cdot\mathbf{T}^{(1,n_1)},\ldots,g_{i_r}\cdot\mathbf{T}^{(1,n_r)}) arising from the orbit of a canonical residue tuple under group elements (gi1,…,gir)(g_{i_1},\ldots,g_{i_r}), SE(3)\mathrm{SE}(3)-equivariance of PP gives

P(gi1⋅T(1,n1),…,gir⋅T(1,nr))=gi1⋅P(T(1,n1),gi1−1gi2⋅T(1,n2),…,gi1−1gir⋅T(1,nr)).(8)P(g_{i_1}\cdot\mathbf{T}^{(1,n_1)},\ldots,g_{i_r}\cdot\mathbf{T}^{(1,n_r)})=g_{i_1}\cdot P(\mathbf{T}^{(1,n_1)},g_{i_1}^{-1}g_{i_2}\cdot\mathbf{T}^{(1,n_2)},\ldots,g_{i_1}^{-1}g_{i_r}\cdot\mathbf{T}^{(1,n_r)}). \tag*{(8)}

When the output first index is in the canonical copy (i1=1i_1=1, so gi1=Idg_{i_1}=\mathrm{Id}), this is exactly the value PphP^{\mathrm{ph}} stores at the corresponding output index, with relative group elements gi1−1gij=gijg_{i_1}^{-1}g_{i_j}=g_{i_j} providing the input-side phantom-copy assignments.

(b) Pooling consistency. For pooled axes, the full-assembly computation of PP on Tsym\mathbf{T}^{\mathrm{sym}} contracts a pooled input index over all kNkN residues in the assembly. By the orbit structure of the input, this kNkN-residue sum decomposes into kk subsums, each over the NN residues of one copy. The subsum corresponding to copy ii is, by part (a), exactly the value computed by PphP^{\mathrm{ph}} at the corresponding canonical-residue tuple with that input axis assigned phantom-copy group element gig_i. The total full-assembly contraction is therefore the sum of PphP^{\mathrm{ph}} values over all kk phantom-copy assignments of the contracted axis, which is exactly the group-element contraction performed by PphP^{\mathrm{ph}} in step 2 of its expansion. The same argument applies to all pooled input axes simultaneously, since the orbit structure factors over independent residue indices.

Together, parts (a) and (b) establish that the full-assembly output of PP, restricted to canonical-copy output tuples, agrees with PphP^{\mathrm{ph}} applied to the canonical representative.

Step 3: Composition. The phantom-copy operator fθphf_\theta^{\mathrm{ph}} is the composition of the per-primitive expansions PℓphP_\ell^{\mathrm{ph}} in the same order as the primitives in fθ=PL∘⋯∘P1f_\theta=P_L\circ\cdots\circ P_1. We prove fθph([T(1)])=[fθ(T(1))]f_\theta^{\mathrm{ph}}([\mathbf{T}^{(1)}])=[f_\theta(\mathbf{T}^{(1)})] by induction on LL.

For L=1L=1, the equality follows directly from Step 2. For the inductive step, suppose the equality holds for compositions of fewer than LL primitives. Let Tsym,ℓ\mathbf{T}^{\mathrm{sym},\ell} denote the intermediate assembly produced after the first ℓ\ell primitives are applied to Tsym\mathbf{T}^{\mathrm{sym}}. Each prefix composition Pℓ∘⋯∘P1P_\ell\circ\cdots\circ P_1 is itself SE(3)\mathrm{SE}(3)-equivariant and chain-permutation-equivariant (these properties are preserved under composition of equivariant primitives), so by Theorem 1 applied to each prefix, every Tsym,ℓ\mathbf{T}^{\mathrm{sym},\ell} is GG-symmetric. Step 2 applied to PLP_L at the input Tsym,L−1\mathbf{T}^{\mathrm{sym},L-1} gives that the canonical-copy output of PL(Tsym,L−1)P_L(\mathbf{T}^{\mathrm{sym},L-1}) equals PLphP_L^{\mathrm{ph}} applied to the canonical representative of Tsym,L−1\mathbf{T}^{\mathrm{sym},L-1}. By the inductive hypothesis, this canonical representative equals (PL−1ph∘⋯∘P1ph)(T(1)).(P_{L-1}^{\mathrm{ph}} \circ\cdots\circ P_1^{\mathrm{ph}})(\mathbf{T}^{(1)}). Composing, the canonical-copy output of fθ(Tsym)f_\theta(\mathbf{T}^{\mathrm{sym}}) equals (PLph∘⋯∘P1ph)(T(1))=fθph(T(1)),(P_L^{\mathrm{ph}} \circ\cdots\circ P_1^{\mathrm{ph}})(\mathbf{T}^{(1)}) = f_\theta^{\mathrm{ph}}(\mathbf{T}^{(1)}), completing the induction.

Remarks on the construction. The phantom-copy realization preserves the architecture of fθf_\theta in the sense that each primitive PphP^{\mathrm{ph}} uses the same parameters and computational structure as the original PP, applied to a different input set. The architectural changes are limited to bookkeeping: intermediate features carry kr−1k^{r-1} group-indexed copies for arity-rr features, and pooling operations now span both residue and group-element axes. Memory and compute reductions follow because the canonical-copy outputs are computed only once rather than kk times as in the full-complex forward pass.

Proof of Theorem 3: Block Structure of Pair-wise Interaction Attention Matrices

We prove the block-structure equality A[ij]=A[1,idx⁡(gi−1gj)]A^{[ij]} = A^{[1,\operatorname{idx}(g_i^{-1}g_j)]} stated in Theorem 3, then expand on the resulting permutation patterns for abelian and non-abelian groups.

Setup. Let A∈RkN×kNA \in\mathbb{R}^{kN \times kN} denote a pair-wise interaction matrix featured in an SE(3)-equivariant model, being evaluated on a GG-symmetric assembly Tsym=(g1⋅T(1),…,gk⋅T(1))\mathbf{T}^{\mathrm{sym}} = (g_1 \cdot\mathbf{T}^{(1)}, \ldots, g_k \cdot\mathbf{T}^{(1)}). We index residues as (i,n)(i,n) for copy i∈[k]i \in[k] and position n∈[N]n \in[N], with A(i,n),(j,m)A_{(i,n),(j,m)} denoting the logit from key residue (j,m)(j,m) to query residue (i,n)(i,n). Arranging AA into k×kk \times k blocks of size N×NN \times N, the (i,j)(i,j)-th block A[ij]A^{[ij]} has entries Anm[ij]=A(i,n),(j,m)A^{[ij]}_{nm} = A_{(i,n),(j,m)}. Entries are SE(3)-invariant scalars.

Proof of Theorem 3. By chain-permutation-equivariance, the permutation πg:i↦idx⁡(g⋅gi)\pi_g : i \mapsto\operatorname{idx}(g \cdot g_i) acts equivariantly on the matrix AA, so for a single entry:

A(i,n),(j,m)(Tsym)=A(i,n),(j,m)(πg∘Tsym)=πg∘A(i,n),(j,m)(Tsym)=A(πg(i),n),(πg(j),m)(Tsym).A_{(i,n),(j,m)}(\mathbf{T}^{\mathrm{sym}}) = A_{(i,n),(j,m)}(\pi_g \circ\mathbf{T}^{\mathrm{sym}}) = \pi_g \circ A_{(i,n),(j,m)}(\mathbf{T}^{\mathrm{sym}}) = A_{(\pi_g(i),n),(\pi_g(j),m)}(\mathbf{T}^{\mathrm{sym}}).

where the first equation follows since the permutation πg\pi_g acting on the GG-symmetric assembly is equivalent to the geometric action of gg, which keeps it intact. Taking g=gi−1g = g_i^{-1}, we have immediately

A(i,n),(j,m)=A(1,n),(idx⁡(gi−1gj),m),A_{(i,n),(j,m)} = A_{(1,n),(\operatorname{idx}(g_i^{-1}g_j),m)},

which is to say,

Anm[ij]=Anm[1,idx⁡(gi−1gj)].(9)A^{[ij]}_{nm} = A^{[1,\operatorname{idx}(g_i^{-1}g_j)]}_{nm}. \tag*{(9)}

Proof of Theorem 3.

Permutation patterns: abelian and non-abelian groups. The block equality says that the ii-th row of blocks of AA is a permutation of the first row, with the jj-th block of the ii-th row equal to the idx⁡(gi−1gj)\operatorname{idx}(g_i^{-1}g_j)-th block of the first row. This permutation pattern is fully determined by the group multiplication table of GG together with the chosen indexing idx.

For abelian GG (cyclic groups CnC_n and direct products thereof), the block at position (i,j)(i,j) depends only on gi−1gjg_i^{-1}g_j, which itself depends only on the index difference j−ij-i modulo nn (under the natural indexing gi=gi−1g_i = g^{i-1} for a generator gg). The block matrix is therefore a circulant: each row is a cyclic shift of the first row, with shift amount equal to the row index. This is the pattern exploited by efficient circulant-matrix multiplications for cyclic-symmetric assemblies.

For non-abelian GG (dihedral DnD_n, polyhedral TT, OO, II), the permutation indexed by gi−1gjg_i^{-1}g_j is the action of gi−1g_i^{-1} on the indexing of GG via left multiplication, which is the regular representation of GG. The block matrix is therefore a GG-equivariant matrix in the sense of group representation theory: it commutes with the action of GG on the row and column indices. Concretely, the row indexed by gig_i is obtained from the first row by relabeling the column indices according to j↦idx⁡(gi⋅gj)j \mapsto\operatorname{idx}(g_i \cdot g_j) (the action of gig_i by left multiplication on the indexing of GG). For CnC_n this reduces to the cyclic shift above; for DnD_n the shifts are interleaved with reflections; for the polyhedral groups the permutations correspond to the rotational subgroup action on the orbit of the canonical copy. A visualization of the permutation patterns for each group used in our experiments is provided in Appendix C.

Detailed Feature Reductions

This appendix expands the per-component reductions sketched in the main text into a precise account of how each feature type stored by an SE(3)-equivariant backbone model is realized in the phantom-copy operator fθphf_{\theta}^{\mathrm{ph}}. The general principle, formalized in Appendix § 8.1.3, is that an arity-rr feature on a GG-symmetric assembly is determined by the canonical-residue tuple together with r−1r-1 relative group elements; the phantom-copy realization stores only this gauge-fixed representation. We instantiate the principle for each feature type appearing in FrameDiff/FoldFlow and RFdiffusion.

Node features (arity $r = 1). Node features have no group-element axis (kr−1=1k^{r-1} = 1). Scalar (invariant) node features are identical across all kk copies of a GG-symmetric assembly by SE(3)-invariance, so storing them only for the canonical subunit reduces the memory footprint from kN⋅DhkN \cdot D_h to N⋅DhN \cdot D_h, where DhD_h is the node embedding dimension. Frame-valued (equivariant) node features—residue frames T(i,n)T^{(i,n)} themselves, predicted frame updates ΔT(i,n)\Delta T^{(i,n)}—are tied across copies by T(i,n)=gi⋅T(1,n)T^{(i,n)} = g_i \cdot T^{(1,n)}, so only the canonical-copy frames T(1,n)T^{(1,n)} are stored explicitly; phantom-copy frames are reconstructed on demand by applying gig_i. This reduction is exact and applies uniformly to every node-level tensor in the model, including IPA local-frame coordinates.

Edge features (arity $r = 2). Edge features carry one group-element axis (kr−1=kk^{r-1} = k), corresponding to the relative group element between query and key residues. The phantom-copy realization stores edge features as a tensor z~nmg\tilde{z}_{nm}^{g} for n,m∈[N]n,m \in[N] and g∈Gg \in G, with memory footprint kN2DzkN^2D_z where DzD_z is the edge embedding dimension. Reading off the relevant blocks: z~nmid\tilde{z}_{nm}^{\mathrm{id}} contains the intra-canonical-subunit edge from residue mm to residue nn; z~nmgj\tilde{z}_{nm}^{g_j} for j≥2j \ge2 contains the inter-copy edge from phantom residue (j,m)(j,m) to canonical residue (1,n)(1,n). Phantom-to-phantom edges (e.g., from (j1,m)(j_1,m) to (j2,n)(j_2,n) with j1,j2≥2j_1,j_2 \ge2) are not stored explicitly; by the block-structure result of Theorem 3, every such edge equals an edge in the canonical block-row indexed by group element gj1−1gj2g_{j_1}^{-1}g_{j_2}. The compression from k2N2Dzk^2N^2D_z to kN2DzkN^2D_z corresponds to retaining only the canonical block-row of the full kN×kNkN \times kN edge tensor.

Edge-feature initialization with self-conditioning. FrameDiff and FoldFlow initialize edge features at each diffusion step using a binned Cα_\alpha pairwise distance matrix from the previous self-conditioned prediction x^sc\hat{x}^{\mathrm{sc}} [Yim et al., 2023]. On a GG-symmetric assembly, this distance matrix has the same block structure as the attention matrix: d(i,n),(j,m)=d(1,n),(gi−1gj,m)d^{(i,n),(j,m)} = d^{(1,n),(g_i^{-1}g_j,m)}, since pairwise distances are SE(3)-invariant. The phantom-copy realization computes binned distances only for the canonical block-row, indexed by g∈Gg \in G, and uses these as the initialization of z~nmg\tilde{z}_{nm}^{g}. The self-conditioned prediction x^sc\hat{x}^{\mathrm{sc}} is itself GG-symmetric in ZEUS by induction (the canonical-copy prediction at each diffusion step is propagated to phantom copies by explicit group action), so the resulting z~nmg\tilde{z}_{nm}^{g} matches the canonical block-row of the full-complex computation exactly. At the first diffusion step, x^sc\hat{x}^{\mathrm{sc}} is initialized to zero, consistent with the training protocol.

Attention logits and weights (arity $r = 2). Attention logits inherit the same arity and group-axis structure as edge features. The phantom-copy realization stores a~nmg\tilde{a}_{nm}^{g} for n,m∈[N]n,m \in[N] and g∈Gg \in G, with memory footprint kN2kN^2 per attention head. After softmax normalization, the attention weights αnmg\alpha_{nm}^{g} inherit the same shape; aggregation of values to canonical-residue updates contracts the key residue index mm and the group-element axis gg:

vn=∑g∈G∑m=1Nα~nmg⋅v~mg(10)v_n = \sum_{g \in G} \sum_{m=1}^{N} \tilde{\alpha}_{nm}^{g} \cdot\tilde{v}_{m}^{g} \tag*{(10)}

which is precisely the einsum-style group-axis contraction described in the proof of Theorem 2.

Frame updates. Frame updates are pointwise frame-valued node maps and follow the node-feature reduction. Predicted updates ΔT(1,n)\Delta T^{(1,n)} are computed for the canonical subunit only; phantom-copy updates are obtained as ΔT(i,n)=gi⋅ΔT(1,n)\Delta T^{(i,n)} = g_i \cdot\Delta T^{(1,n)} during the assembly reconstruction at the end of each diffusion step. This step is the practical mechanism that propagates the canonical-copy update to the full assembly, and it incurs O(kN)O(kN) rather than O(kN⋅model-cost)O(kN \cdot\text{model-cost}) work.

Further Discussions

Extended Related Work

Representing symmetric structures through symmetry-inequivalent components and their group orbits is a well-established construction used across crystallography and protein design. In crystal generation, DiffCSP++ [18] and SGEquiDiff [8] model symmetry-inequivalent atoms under Wyckoff constraints and recover symmetry-related copies through space-group operations. Related approaches exploit the same symmetry structure in different ways: WyFormer [20] generates discrete descriptions of occupied Wyckoff sites that are subsequently converted into crystal structures, while Space Group Conditional Flow Matching [26] initializes symmetry-compatible atomic configurations and preserves their Wyckoff constraints through an equivariant flow over the full unit cell. In amyloid-binder design, Wallace et al. [Wallace et al., 2024] use the target fibril’s screw symmetry to replicate helical binder motifs along its surface and jointly optimize binder–fibril and binder–binder interactions. Related symmetry enforcement also appears in continuous pattern design: Lin et al. [22] construct continuous representations invariant under any of the 17 wallpaper groups, with applications to pattern, paper-cutting, topology, and material design.

Violation of Permutation Invariance in Chain Embeddings

The chain-permutation-equivariance assumption of Theorems 1 and 2 is violated in practice by the relative-residue-index encoding ϕ(m−n)\phi(m-n) used to initialize edge features and chain offsets in RFdiffusion [Watson et al., 2023]. RFdiffusion’s default inference splits the chains with LL, before initializing edge features with relative sequential distances. In general, any such method that instantiates signed relative embeddings is not chain-permutation equivariant, as the swapping of chains ii and jj changes the sign of the relative-index encoding. This has the effect of breaking the equivalence claim of ZEUS with respect to pairwise interaction blocks. Recall the block structure of the kN×kNkN \times kN pair-wise interaction matrix from Theorem 3: under exact chain-permutation equivariance, A[ij]=A[1,gi−1gj]A^{[ij]} = A^{[1,g_i^{-1}g_j]} for all i,j∈[k]i,j \in[k]. The intra-copy diagonal blocks A[ii]A^{[ii]} are computed exclusively from intra-copy edge features, and are therefore unaffected by the inter-copy offset Δ\Delta. Hence they match exactly across all copies even in the presence of this violation. The inter-copy off-diagonal blocks A[ij]A^{[ij]}’s, with i≠ji \ne j, are affected by this discrepancy. The equality A[ij]=A[1,gi−1gj]A^{[ij]} = A^{[1,g_i^{-1}g_j]} then becomes approximate rather than exact.

Proposed fixes. Several possible fixes can restore exact chain-permutation equivariance, though they certainly require model retraining and lie outside the zero shot scope of this work. The simplest fix is an out-of-range token, which replaces inter-copy relative indices with a fixed special token, which is assigned a learned embedding distinct from any in-range relative index. This is the approach adopted by some structure-prediction models for chain breaks, including AF-Multimer’s chain-relative encoding [19]. A more advanced method may involve a symmetry-aware encoding that represents the relative group relation between the pair, designed to be invariant across chain permutations. Both preserve the within-copy semantics of relative position while restoring exact chain-permutation equivariance, and we view this development as a natural direction for future work.

Phantom-Copy Truncation for Infinite Symmetry Groups

For finite point groups, all ∣G∣=k|G| = k phantom copies are retained and Theorems 1 and 2 hold exactly outside the chain-embedding caveat of Appendix B.2. For infinite groups such as lattice and helical symmetries, the orbit is unbounded and exact phantom-copy generation is impossible. We propose to implement distance-based truncation as a natural way to deal with this, motivated by the fact that many architectures already imbue distance priors. For example, IPA has a distance term that penalizes residue pairs quadratically.

Distance Truncation. Given a canonical subunit T(1)T^{(1)} and a possibly infinite symmetry group GG, we truncate the phantom-copy set to those chains whose minimum residue–residue distance to the canonical chain falls below a hyperparameter dmax⁡d_{\max}, that is, {g⋅T(1):g∈G, dist⁡(g⋅T(1),T(1))<\{g \cdot T^{(1)} : g \in G,\ \operatorname{dist}(g \cdot T^{(1)}, T^{(1)}) < dmax⁡d_{\max}} where dist denotes the minimum pairwise residue distance between the candidate phantom copy and the canonical chain. Concretely, we enumerate group elements in order of increasing minimum distance and include them until the cutoff is reached, yielding a finite phantom-copy set of size κ\kappa even when GG is infinite. For lattice and helical groups acting with bounded density on R3\mathbb{R}^3, κ=O(1)\kappa= O(1) independent of the (formal) group order, with the constant determined by dmax⁡d_{\max} and the geometric density of the orbit.

Visualization of Block Symmetric Attention

We empirically validate Theorem 3 by visualizing the attention matrices that arise during full-complex symmetric inference. We run full-complex (non-ZEUS) symmetric diffusion on FrameDiff, FoldFlow, and RFdiffusion under C8C_8 and tetrahedral symmetry. For FrameDiff and FoldFlow we use an inter-chain offset of 1000 residue positions.

For each model-symmetry pair, Figure 5–Figure 10 display attention logits at head 0 across layers and denoising timesteps. Also displayed is the analytical block structure predicted by the symmetry group GG. The predicted block structure emerges robustly across all three architectures, all sampled layers, and all sampled timesteps.

RFdiffusion attention logits under C8 symmetry across network layers and denoising timesteps

Figure 5. RFdiffusion attention logits under C8C_8 symmetry, at attention head 0, across a grid of network layers and denoising timesteps. The analytical block structure predicted by Theorem 3 for G=C8G = C_8 is shown for reference.

FoldFlow attention logits under tetrahedral symmetry across network layers and denoising timesteps

Figure 10. FoldFlow attention logits under tetrahedral (TT) symmetry, at attention head 0, across a grid of network layers and denoising timesteps. The analytical block structure predicted by Theorem 3 for G=TG = T is shown for reference.

RFdiffusion attention logits under tetrahedral symmetry

Figure 6. RFdiffusion attention logits under tetrahedral (TT) symmetry, at attention head 0, across a grid of network layers and denoising timesteps. The analytical block structure predicted by Theorem 3 for G=TG = T is shown for reference.

FrameDiff attention logits under C8 symmetry across network layers and denoising timesteps

Figure 7. FrameDiff attention logits under C8C_8 symmetry, at attention head 0, across a grid of network layers and denoising timesteps. The analytical block structure predicted by Theorem 3 for G=C8G = C_8 is shown for reference.

FrameDiff attention logits under tetrahedral symmetry across network layers and denoising timesteps

Figure 8. FrameDiff attention logits under tetrahedral (TT) symmetry, at attention head 0, across a grid of network layers and denoising timesteps. The analytical block structure predicted by Theorem 3 for G=TG = T is shown for reference.

FoldFlow attention logits under $C_8$ symmetry, at attention head 0, across a grid of network layers and denoising timesteps

Figure 9. FoldFlow attention logits under C8C_8 symmetry, at attention head 0, across a grid of network layers and denoising timesteps. The analytical block structure predicted by Theorem 3 for G=C8G = C_8 is shown for reference.

Algorithm Implementation

Symmetric Invariant Point Attention for Framediff and FoldFlow

Algorithm D.1 describes our symmetric aware IPA implementation, used in the Framediff and FoldFlow Architecture.

Algorithm 1 Framediff SE(3) Symmetric Diffusion Block (single block ℓ→ℓ+1\ell\to\ell+ 1)

Input:

Node features: hℓ∈RLcan×dh\mathbf{h}_{\ell} \in\mathbb{R}^{L_{can} \times d_h}

Edge features: zℓ∈RLcan×L×dz\mathbf{z}_{\ell} \in\mathbb{R}^{L_{can} \times L \times d_z},

Backbone frames Tℓ∈SE(3)Lcan\mathbf{T}_{\ell} \in\mathrm{SE}(3)^{L_{can}}

Initial node features h0\mathbf{h}_0

Symmetry Group G∈SE(3)G \in\mathrm{SE}(3)

Index helpers (j∈{1,…,L}j \in\{1, \ldots, L\}):

subunit: σ(j)=⌈j/Lcan⌉\sigma(j) = \lceil j/L_{can} \rceil // map residue idx to group idx

canonical residue: ρ(j)=((j−1) mod Lcan)+1\rho(j) = ((j - 1) \bmod L_{can}) + 1 // map residue idx to canonical idx

Output:

Updated Node features: hℓ+1\mathbf{h}_{\ell+1},

Updated Edge features: zℓ+1\mathbf{z}_{\ell+1},

Updated Frames: Tℓ+1\mathbf{T}_{\ell+1}

NodeUpdate:

  1. h′←hℓ+SymmetricInvariantPointAttention⁡(hℓ,zℓ,Tℓ,G)\mathbf{h}^{\prime} \leftarrow\mathbf{h}_{\ell} + \operatorname{SymmetricInvariantPointAttention}(\mathbf{h}_{\ell}, \mathbf{z}_{\ell}, \mathbf{T}_{\ell}, G)

  2. h′′←Transformer⁡(Concat⁡(h′,Linear⁡(h0)))\mathbf{h}^{\prime\prime} \leftarrow\operatorname{Transformer}(\operatorname{Concat}(\mathbf{h}^{\prime}, \operatorname{Linear}(\mathbf{h}_0)))

  3. hℓ+1←MLP⁡(h′′+h′)\mathbf{h}_{\ell+1} \leftarrow\operatorname{MLP}(\mathbf{h}^{\prime\prime} + \mathbf{h}^{\prime})

EdgeUpdate:

  1. for each edge (n∈Lcan,m∈L)(n \in L_{can}, m \in L) do

  2. e←Concat⁡(Linear⁡(hℓ+1n),Linear⁡(hℓ+1ρ(m)),zℓnm)\mathbf{e} \leftarrow\operatorname{Concat}(\operatorname{Linear}(\mathbf{h}_{\ell+1}^{n}), \operatorname{Linear}(\mathbf{h}_{\ell+1}^{\rho(m)}), \mathbf{z}_{\ell}^{nm})

  3. e′←MLP⁡(e)+e\mathbf{e}^{\prime} \leftarrow\operatorname{MLP}(\mathbf{e}) + \mathbf{e}

  4. zℓ+1nm←Linear⁡(e′)\mathbf{z}_{\ell+1}^{nm} \leftarrow\operatorname{Linear}(\mathbf{e}^{\prime})

  5. end for

BackboneUpdate:

  1. Tℓ+1←BackboneUpdate⁡(Tℓ,hℓ+1)\mathbf{T}_{\ell+1} \leftarrow\operatorname{BackboneUpdate}(\mathbf{T}_{\ell}, \mathbf{h}_{\ell+1})

  2. return hℓ+1,zℓ+1,Tℓ+1\mathbf{h}_{\ell+1}, \mathbf{z}_{\ell+1}, \mathbf{T}_{\ell+1}

Algorithm 2 Efficient Symmetry-Aware Invariant Point Attention (IPA)

Input:

Canonical node features: s∈RLcan×dss \in\mathbb{R}^{L_{\mathrm{can}}\times d_s}

Block-row Edge features: z∈RLcan×L×dzz \in\mathbb{R}^{L_{\mathrm{can}}\times L\times d_z}, with L=k⋅LcanL = k \cdot L_{\mathrm{can}}

Canonical backbone frames: {Ti}i=1Lcan∈SE(3)\{T_i\}_{i=1}^{L_{\mathrm{can}}} \in\mathrm{SE}(3)

Symmetry group: {g1,g2,…,gk}⊂G\{g_1, g_2, \ldots, g_k\} \subset G

Index helpers (j∈{1,…,L}j \in\{1, \ldots, L\}):

subunit: σ(j)=⌈j/Lcan⌉\sigma(j) = \lceil j/L_{\mathrm{can}}\rceil // map residue idx to group idx

canonical residue: ρ(j)=((j−1) mod Lcan)+1\rho(j) = ((j - 1) \bmod L_{\mathrm{can}}) + 1 // map residue idx to canonical idx

Output: Updated Canonical single representations s~∈RLcan×ds\tilde{s} \in\mathbb{R}^{L_{\mathrm{can}}\times d_s}

// Convert invariant/equivariant repr. under the symmetry group.

  1. sj′←sρ(j)s'_j \leftarrow s_{\rho(j)} j∈{1,…,L}j \in\{1, \ldots, L\}

  1. Tj′←gσ(j)∘Tρ(j)T'_j \leftarrow g_{\sigma(j)} \circ T_{\rho(j)} j∈{1,…,L}j \in\{1, \ldots, L\}

// Queries from canonical residues, keys/values from the phantoms.

  1. qih←LinearNoBias⁡(si)\mathbf{q}_i^h \leftarrow\operatorname{LinearNoBias}(s_i) i∈{1,…,Lcan}i \in\{1, \ldots, L_{\mathrm{can}}\}

  1. kjh,vjh←LinearNoBias⁡(sj′)\mathbf{k}_j^h,\mathbf{v}_j^h \leftarrow\operatorname{LinearNoBias}(s'_j) j∈{1,…,L}j \in\{1, \ldots, L\}

  1. q⃗ihp←Ti∘LinearNoBias⁡(si)∈R3, p∈{1,…,Nqp}\vec{\mathbf{q}}_i^{hp} \leftarrow T_i \circ\operatorname{LinearNoBias}(s_i) \in\mathbb{R}^3,\ p \in\{1, \ldots, N_{qp}\}

  1. k⃗jhp←Tj′∘LinearNoBias⁡(sj′)∈R3, p∈{1,…,Nqp}\vec{\mathbf{k}}_j^{hp} \leftarrow T'_j \circ\operatorname{LinearNoBias}(s'_j) \in\mathbb{R}^3,\ p \in\{1, \ldots, N_{qp}\}

  1. v⃗jhp←Tj′∘LinearNoBias⁡(sj′)∈R3, p∈{1,…,Nvp}\vec{\mathbf{v}}_j^{hp} \leftarrow T'_j \circ\operatorname{LinearNoBias}(s'_j) \in\mathbb{R}^3,\ p \in\{1, \ldots, N_{vp}\}

  1. bijh←LinearNoBias⁡(zij)b_{ij}^h \leftarrow\operatorname{LinearNoBias}(z_{ij}) i∈{1,…,Lcan}, j∈{1,…,L}i \in\{1, \ldots, L_{\mathrm{can}}\},\ j \in\{1, \ldots, L\}

  1. wC←29Nqp;wL←13w_C \leftarrow\sqrt{\frac{2}{9N_{qp}}};\quad w_L \leftarrow\sqrt{\frac{1}{3}}

// Compute attention weights for phantom →\to canonical only.

  1. aijh←Softmax⁡j∈{1,…,L}(wL(1cqihTkjh+bijh−γhwC2∑p=1Nqp∥Ti∘q⃗ihp−Tj′∘k⃗jhp∥2))a_{ij}^h \leftarrow\operatorname{Softmax}_{j\in\{1,\ldots,L\}}\left(w_L\left(\frac{1}{\sqrt{c}}\mathbf{q}_i^{h\mathsf{T}}\mathbf{k}_j^h+b_{ij}^h-\frac{\gamma^h w_C}{2}\sum_{p=1}^{N_{qp}}\left\lVert T_i\circ\vec{\mathbf{q}}_i^{hp}-T'_j\circ\vec{\mathbf{k}}_j^{hp}\right\rVert^2\right)\right)

  1. o~ih←∑j=1Laijhzij\tilde{o}_i^h \leftarrow\sum_{j=1}^{L}a_{ij}^h z_{ij}

  1. oih←∑j=1Laijhvjho_i^h \leftarrow\sum_{j=1}^{L}a_{ij}^h\mathbf{v}_j^h

  1. o⃗ihp←Ti−1∘∑j=1Laijh(Tj′∘v⃗jhp)\vec{o}_i^{hp} \leftarrow T_i^{-1}\circ\sum_{j=1}^{L}a_{ij}^h\left(T'_j\circ\vec{\mathbf{v}}_j^{hp}\right)

  1. s~i←Linear⁡(Concat⁡h,p(o~ih,oih,o⃗ihp,∥o⃗ihp∥))\tilde{s}_i \leftarrow\operatorname{Linear}\left(\operatorname{Concat}_{h,p}\left(\tilde{o}_i^h,o_i^h,\vec{o}_i^{hp},\left\lVert\vec{o}_i^{hp}\right\rVert\right)\right)

  1. return s~∈RLcan×ds\tilde{s} \in\mathbb{R}^{L_{\mathrm{can}}\times d_s}

Symmetric BiasedAxialAttention for RFdiffusion

Algorithm D.2 describes the native Axial Attention used in RFdiffusion and RoseTTAFold. We include it to help the reader visualize ZEUS’ implementation differences.

Algorithm 3 BiasedAxialAttention (RFdiffusion/RoseTTAFold-native axial attention)

Input: z∈RL×L×dzz \in\mathbb{R}^{L \times L \times d_z}, bias b∈RL×Lb \in\mathbb{R}^{L \times L}

Output: Updated pair z′z^{\prime} of shape (L,L,dz)(L, L, d_z)

  1. if IsRow then z←zT,b←bTz \leftarrow z^{\mathsf{T}}, \quad b \leftarrow b^{\mathsf{T}}

  1. end if

  1. z←LayerNorm⁡(z);b←LayerNorm⁡(b)z \leftarrow\operatorname{LayerNorm}(z); \quad b \leftarrow\operatorname{LayerNorm}(b)

  1. Q←Wqz,K←Wkz,V←Wvz,B←WbbQ \leftarrow W_q z, \quad K \leftarrow W_k z, \quad V \leftarrow W_v z, \quad B \leftarrow W_b b \hfill ▹\triangleright shapes (L,L,h,d)(L, L, h, d)

  1. g←sigm⁡(Wgz)g \leftarrow\operatorname{sigm}(W_g z)

  1. Q←Q/d;K←K/LQ \leftarrow Q/\sqrt{d}; \quad K \leftarrow K/\sqrt{L}

  1. A[:,:,h]=∑k=1dQ[:,:,h,k]TK[:,:,h,k],Q[:,:,h,k],K[:,:,h,k]∈RL×LA[:,:,h] = \sum_{k=1}^{d} Q[:,:,h,k]^{\mathsf{T}} K[:,:,h,k], \quad Q[:,:,h,k], K[:,:,h,k] \in\mathbb{R}^{L \times L}

  1. A←Softmax⁡(A+B)A \leftarrow\operatorname{Softmax}(A+B)

  1. O[:,:,h,k]=A[:,:,h]V[:,:,h,k]T,A[:,:,h],V[:,:,h,k]∈RL×L,O[:,:,h,k]∈RL×LO[:,:,h,k] = A[:,:,h]V[:,:,h,k]^{\mathsf{T}}, \quad A[:,:,h], V[:,:,h,k] \in\mathbb{R}^{L \times L}, \quad O[:,:,h,k] \in\mathbb{R}^{L \times L}

  1. O←g⊙OO \leftarrow g \odot O

  1. z′←WoOz^{\prime} \leftarrow W_o O

  1. if IsRow then z′←z′Tz^{\prime} \leftarrow z^{\prime\mathsf{T}}

  1. end if

return z′z^{\prime}

Algorithm 4 SymmetricBiasedAxialAttention (symmetry-aware variant of BiasedAxialAttention)

Input:

Block-row pair z∈RLcan×L×dzz \in\mathbb{R}^{L_{\mathrm{can}} \times L \times d_z}, bias b∈RLcan×Lb \in\mathbb{R}^{L_{\mathrm{can}} \times L}

Group order SS with L=SLcanL = S L_{\mathrm{can}}

Permutation table P∈ZS×SP \in\mathbb{Z}^{S \times S}, where P[g,h]=idx⁡(g−1h)P[g,h] = \operatorname{idx}(g^{-1}h)

Output: Updated restricted pair z′∈RLcan×L×dzz' \in\mathbb{R}^{L_{\mathrm{can}} \times L \times d_z}

// Per-block transpose for row attention.

  1. if ISROW then

  2. for β=0,…,S−1\beta= 0,\ldots,S-1 do

  3. z[:,βLcan:(β+1)Lcan,:]←z[:,P[β,0]Lcan:(P[β,0]+1)Lcan,:]Tz[:,\beta L_{\mathrm{can}}:(\beta+1)L_{\mathrm{can}},:] \leftarrow z[:,P[\beta,0]L_{\mathrm{can}}:(P[\beta,0]+1)L_{\mathrm{can}},:]^{\mathsf{T}}

  4. b[:,βLcan:(β+1)Lcan]←b[:,P[β,0]Lcan:(P[β,0]+1)Lcan]b[:,\beta L_{\mathrm{can}}:(\beta+1)L_{\mathrm{can}}] \leftarrow b[:,P[\beta,0]L_{\mathrm{can}}:(P[\beta,0]+1)L_{\mathrm{can}}]

  5. end for

  6. end if

  7. z←LayerNorm⁡(z); b←LayerNorm⁡(b)z \leftarrow\operatorname{LayerNorm}(z);\ b \leftarrow\operatorname{LayerNorm}(b)

  8. Q←Wqz, K←Wkz, V←Wvz, B←WbbQ \leftarrow W_qz,\ K \leftarrow W_kz,\ V \leftarrow W_vz,\ B \leftarrow W_bb ▷ shapes (Lcan,L,h,d)(L_{\mathrm{can}},L,h,d)

  9. g←sigm⁡(Wgz)g \leftarrow\operatorname{sigm}(W_gz)

  10. Q←Q/d; K←K/LQ \leftarrow Q/\sqrt{d};\ K \leftarrow K/\sqrt{L}

// Split along the column axis into SS group-blocks.

  1. {Q(s)}s=0S−1←SplitAlongColumns⁡(Q,nsplits=S)\{Q^{(s)}\}_{s=0}^{S-1} \leftarrow\operatorname{SplitAlongColumns}(Q,n_{\mathrm{splits}}=S);

  2. {K(s)}s=0S−1←SplitAlongColumns⁡(K,nsplits=S)\{K^{(s)}\}_{s=0}^{S-1} \leftarrow\operatorname{SplitAlongColumns}(K,n_{\mathrm{splits}}=S);

  3. {V(s)}s=0S−1←SplitAlongColumns⁡(V,nsplits=S)\{V^{(s)}\}_{s=0}^{S-1} \leftarrow\operatorname{SplitAlongColumns}(V,n_{\mathrm{splits}}=S); ▷ each block ∈RLcan×Lcan×h×d\in\mathbb{R}^{L_{\mathrm{can}}\times L_{\mathrm{can}}\times h\times d}

// Compute attention blockwise with permutation table for phantoms →\to canonical only.

  1. A(s)[:,:,h]←∑k=1d∑g=0S−1Q(P[g,0])[:,:,h,k]TK(P[g,s])[:,:,h,k], s=0,…,S−1A^{(s)}[:,:,h] \leftarrow\sum_{k=1}^{d}\sum_{g=0}^{S-1}Q^{(P[g,0])}[:,:,h,k]^{\mathsf{T}}K^{(P[g,s])}[:,:,h,k],\ s=0,\ldots,S-1

  2. A[:,:,h]←[A(0)[:,:,h]∥⋯∥A(S−1)[:,:,h)]∈RLcan×LA[:,:,h] \leftarrow[A^{(0)}[:,:,h]\mathbin\Vert\cdots\mathbin\Vert A^{(S-1)}[:,:,h)] \in\mathbb{R}^{L_{\mathrm{can}}\times L}

  3. A←Softmax⁡(A+B)A \leftarrow\operatorname{Softmax}(A+B)

  4. {A(s)}s=0S−1←SplitAlongColumns⁡(A,nsplits=S)\{A^{(s)}\}_{s=0}^{S-1} \leftarrow\operatorname{SplitAlongColumns}(A,n_{\mathrm{splits}}=S) ▷ each A(s)[:,:,h]∈RLcan×LcanA^{(s)}[:,:,h] \in\mathbb{R}^{L_{\mathrm{can}}\times L_{\mathrm{can}}}

// Assemble Output by gathering VV blocks.

  1. if ISROW then

  2. O(bi)[:,:,h,k]←∑kj=0S−1A(kj)[:,:,h]V(P[P[bi,0],kj])[:,:,h,k]T, bi=0,…,S−1O^{(b_i)}[:,:,h,k] \leftarrow\sum_{k_j=0}^{S-1}A^{(k_j)}[:,:,h]V^{(P[P[b_i,0],k_j])}[:,:,h,k]^{\mathsf{T}},\ b_i=0,\ldots,S-1

  3. else

  4. O(cb)[:,:,h,k]←∑kj=0S−1A(kj)[:,:,h]V(P[cb,kj])[:,:,h,k]T, cb=0,…,S−1O^{(c_b)}[:,:,h,k] \leftarrow\sum_{k_j=0}^{S-1}A^{(k_j)}[:,:,h]V^{(P[c_b,k_j])}[:,:,h,k]^{\mathsf{T}},\ c_b=0,\ldots,S-1

  5. end if

  6. O[:,:,h,k]←[O(0)[:,:,h,k]∥⋯∥O(S−1)[:,:,h,k)]∈RLcan×LO[:,:,h,k] \leftarrow[O^{(0)}[:,:,h,k]\mathbin\Vert\cdots\mathbin\Vert O^{(S-1)}[:,:,h,k)] \in\mathbb{R}^{L_{\mathrm{can}}\times L}

  7. O←g⊙OO \leftarrow g \odot O

  8. z′←WoOz' \leftarrow W_oO

  9. return z′z'

Implementations of Symmetry Transformations

We apply the group symmetry operator during generation as a left matrix multiplication with the group elements’ rotation matrices:

G⋅Tn=[(Rg1⋅rn,Rg1⋅xn),…,(Rgk⋅rn,Rgk⋅xn)](11)G \cdot T_n = [(R_{g_1}\cdot r_n,R_{g_1}\cdot x_n),\ldots,(R_{g_k}\cdot r_n,R_{g_k}\cdot x_n)] \tag*{(11)}

As denoted before, let ∣G∣=k|G|=k. For cyclic symmetries (CnC_n, k=nk=n) the rotation matrices RgiR_{g_i} are rotations of 2πik\frac{2\pi i}{k} about zz.

For dihedral (DnD_n, k=2nk=2n) they are the cyclic rotations above, composed with a π\pi rotation about xx.

For tetrahedral (T,k=12)(T, k = 12) the rotation matrices RgiR_{g_i} comprise the identity; eight rotations of ±2π/3\pm2\pi/3 about the four body-diagonal axes (±1,±1,±1)/3(\pm1, \pm1, \pm1)/\sqrt{3} (with an even number of minus signs); and three rotations of π\pi about the coordinate axes.

For octahedral (O,k=24)(O, k = 24) the rotation matrices RgiR_{g_i} comprise the identity; nine rotations of ±π/2\pm\pi/2 and π\pi about the three coordinate axes; eight rotations of ±2π/3\pm2\pi/3 about the four body-diagonal axes (±1,±1,±1)/3(\pm1, \pm1, \pm1)/\sqrt{3}; and six rotations of π\pi about the face-diagonal axes (±1,±1,0)/2(\pm1, \pm1, 0)/\sqrt{2} and cyclic permutations.

For icosahedral (I,k=60)(I, k = 60) the rotation matrices RgiR_{g_i} comprise the identity, twenty-four rotations of ±2π/5\pm2\pi/5 and ±4π/5\pm4\pi/5 about the six axes through opposite vertices, located at cyclic permutations of (0,±1,±φ)(0, \pm1, \pm\varphi) where φ=(1+5)/2\varphi= (1 + \sqrt{5})/2, twenty rotations of ±2π/3\pm2\pi/3 about the ten axes through opposite face centers; and fifteen rotations of π\pi about the fifteen axes through opposite edge midpoints.

In practice, we find an alternative symmetric operation that parameterizes the symmetry operation as a rotation about an offset center to yield more valid results for Framediff and Foldflow.

G⋅Tn=[(Rg1⋅rn,Rg1⋅(xn−t)+t),…,(Rgk⋅rn,Rgk⋅(xn−t)+t)](12)G \cdot T_n = [(R_{g_1} \cdot r_n, R_{g_1} \cdot(x_n - t) + t), \ldots, (R_{g_k} \cdot r_n, R_{g_k} \cdot(x_n - t) + t)] \tag*{(12)}

where t∈R3t \in\mathbb{R}^3 is a translation vector placing the canonical subunit at distance ∥t∥\lVert t\rVert from the symmetry axis. For the CkC_k case this causes the kk subunits to initialize on a ring of radius ∥t∥\lVert t\rVert rather than all at the origin.

We choose TT so that adjacent subunits at t=0t = 0 are separated by the characteristic spread of the initialization distribution. Two subunits related by RkR_k and placed at distance rr from the origin are separated by chord length 2rsin⁡(π/k)2r\sin(\pi/k); setting this equal to 4σ4\sigma and applying a cube-root correction for subunit volume gives

∥t∥=2σsin⁡(π/k)⋅N503(13)\lVert t\rVert= \frac{2\sigma}{\sin(\pi/k)} \cdot\sqrt[3]{\frac{N}{50}} \tag*{(13)}

where σ\sigma is the standard deviation of the translational frame initialization and NN is the subunit length.

Justification. For our SE(3) diffusion models, at t=0t = 0, frame translations are drawn from an isotropic Gaussian centered at the origin with scale σ\sigma. Applying T0T_0 results in all kk symmetric copies of the canonical subunit initializing on top of one another. The denoising process must push these copies apart over the course of generation. Empirically, we find this manifests as a sharp drop in designability for k≥6k \ge6: the model exhausts its denoising budget on spatial separation, and in some cases, cannot separate chains at all, leading to highly clashing designs. This is especially pronounced for models not pretrained on multi-chain generation such as Framediff and FoldFlow. The application of this offset symmetrization constrains the trajectory’s center of mass to lie at radius ∥t∥\lVert t\rVert from the axis, freeing the model to spend its capacity on local geometry.

Experimental details

We benchmark on cyclic symmetries C2,C3,C4,C6,C8,C12C_2, C_3, C_4, C_6, C_8, C_{12} with subunit length n=50n = 50, yielding total complex sizes from 100 to 600 residues. The choice of n=50n = 50 balances designability against known out of distribution behavior of FrameDiff and FoldFlow above ∼300\sim300 residues at the full-complex level. All experiments are run on a single NVIDIA A100 (40 GB VRAM).

We use the FrameDiff paper checkpoint at https://github.com/Immortals-33/Framediff/tree/master, the FoldFlow-Base https://github.com/DreamFold/FoldFlow/releases/download/0.1.0/foldflow-base.pth, and the RFdiffusion base checkpoint http://files.ipd.uw.edu/pub/RFdiffusion/6f5902ac237024bdd0c176cb93063dc4/Base_ckpt.pt.

Each is run with the published default hyperparameters. FrameDiff with 500 diffusion steps, translation schedule 0.1→200.1 \rightarrow20, and IGSO(3) schedule 0.1→1.50.1 \rightarrow1.5; FoldFlow with translation schedule 0.01→200.01 \rightarrow20 and IGSO(3) schedule 0.01→1.50.01 \rightarrow1.5; and RFdiffusion with 50 diffusion steps and β\beta schedule 0.01→0.070.01 \rightarrow0.07.

For FrameDiff and FoldFlow full-complex inference, we preserve within-chain residue indexes and assign an inter-chain separation offset of 1000. For RFdiffusion full-complex inference we maintain the native inter-chain separation offsets of ±L\pm L where LL is the total number of residues in the complex.

By definition, any sign on the chain offset disrupts chain-permutation equivariance. For RFdiffusion ablation and ZEUS inference, all chains instead share identical sequential indices. Interchain offsets are set to LL for RFdiffusion.

For RFdiffusion we centre the diffusion process about the whole-complex. At each step the model receives the entire kNkN-residue complex; we extract the predicted p(x0)p(x_0) for the canonical subunit, apply the reverse step, and propagate the update to all phantom subunits via the standard rotational symmetry operation 10, with the reverse trajectory centered on the full complex.

For Framediff and FoldFlow we use the offset symmetry parameterization 11 with prior radius determined using equation (13).

Each generated design’s sequence is predicted with ProteinMPNN at sampling temperature 0.1, using tied positions across symmetric subunits, followed by refolding prediction with ESMFold v1 in multimer mode. For FrameDiff and FoldFlow, ESMFold prediction is restricted to the canonical subunit plus its two nearest phantom neighbors.

Novelty against the PDB is computed with Foldseek [Van Kempen et al., 2024] with foldseek easy-search at the subunit level and foldseek easy-multimersearch -e inf at the complex level.

Additional Experimental Metrics

Boxplot visualization of distributional metrics

Figure 11. Boxplot visualization of Table 3. Distributional metrics for Cyclic generation with subunit length=50 with ZEUS, for all models. Grey box is the baseline full-complex inference, while coloured is ZEUS. ZEUS achieves equivalent performance across all metrics.

settinghelixsheetcoil
C2_1000.863±0.036/0.859±0.0410.863 \pm0.036 / 0.859 \pm0.0410±0/0±00 \pm0 / 0 \pm00.137±0.036/0.141±0.0410.137 \pm0.036 / 0.141 \pm0.041
C3_1500.870±0.064/0.852±0.0490.870 \pm0.064 / 0.852 \pm0.0490.006±0.042/0±00.006 \pm0.042 / 0 \pm00.124±0.033/0.148±0.0490.124 \pm0.033 / 0.148 \pm0.049
C4_2000.871±0.031/0.864±0.0390.871 \pm0.031 / 0.864 \pm0.0390±0/0±00 \pm0 / 0 \pm00.129±0.031/0.136±0.0390.129 \pm0.031 / 0.136 \pm0.039
C6_3000.870±0.034/0.851±0.0420.870 \pm0.034 / 0.851 \pm0.0420±0/0±00 \pm0 / 0 \pm00.130±0.034/0.149±0.0420.130 \pm0.034 / 0.149 \pm0.042
C8_4000.858±0.088/0.864±0.0400.858 \pm0.088 / 0.864 \pm0.0400.007±0.048/0±00.007 \pm0.048 / 0 \pm00.135±0.053/0.136±0.0400.135 \pm0.053 / 0.136 \pm0.040
C12_6000.872±0.046/0.860±0.0460.872 \pm0.046 / 0.860 \pm0.0460±0/0.002±0.0110 \pm0 / 0.002 \pm0.0110.128±0.046/0.138±0.0450.128 \pm0.046 / 0.138 \pm0.045
T_6000.883±0.030/0.866±0.0410.883 \pm0.030 / 0.866 \pm0.0410±0/0±00 \pm0 / 0 \pm00.117±0.030/0.134±0.0410.117 \pm0.030 / 0.134 \pm0.041
O_12000.879±0.042/0.870±0.0400.879 \pm0.042 / 0.870 \pm0.0400±0/0±00 \pm0 / 0 \pm00.121±0.042/0.130±0.0400.121 \pm0.042 / 0.130 \pm0.040

Table 4. FoldFlow – DSSP secondary-structure fractions. Each cell shows full-complex / ZEUS as mean ± std.

settinghelixsheetcoil
C2_1000.679±0.2510.679 \pm0.251 / 0.736±0.1740.736 \pm0.1740.070±0.1550.070 \pm0.155 / 0.039±0.1210.039 \pm0.1210.251±0.1420.251 \pm0.142 / 0.225±0.0860.225 \pm0.086
C3_1500.735±0.2060.735 \pm0.206 / 0.723±0.2040.723 \pm0.2040.059±0.1310.059 \pm0.131 / 0.057±0.1230.057 \pm0.1230.206±0.1060.206 \pm0.106 / 0.220±0.1200.220 \pm0.120
C4_2000.785±0.1720.785 \pm0.172 / 0.714±0.2190.714 \pm0.2190.030±0.0980.030 \pm0.098 / 0.021±0.0670.021 \pm0.0670.184±0.0910.184 \pm0.091 / 0.265±0.1960.265 \pm0.196
C6_3000.700±0.2540.700 \pm0.254 / 0.719±0.1780.719 \pm0.1780.059±0.1360.059 \pm0.136 / 0.036±0.0860.036 \pm0.0860.242±0.1830.242 \pm0.183 / 0.245±0.1410.245 \pm0.141
C8_4000.770±0.2010.770 \pm0.201 / 0.712±0.2370.712 \pm0.2370.049±0.1230.049 \pm0.123 / 0.067±0.1590.067 \pm0.1590.180±0.0960.180 \pm0.096 / 0.221±0.1080.221 \pm0.108
C12_6000.747±0.2170.747 \pm0.217 / 0.705±0.1990.705 \pm0.1990.049±0.1230.049 \pm0.123 / 0.044±0.1240.044 \pm0.1240.204±0.1110.204 \pm0.111 / 0.251±0.1250.251 \pm0.125
T_6000.836±0.1400.836 \pm0.140 / 0.712±0.2130.712 \pm0.2130.019±0.0990.019 \pm0.099 / 0.049±0.1350.049 \pm0.1350.145±0.0650.145 \pm0.065 / 0.239±0.1230.239 \pm0.123
O_12000.846±0.0810.846 \pm0.081 / 0.713±0.1760.713 \pm0.1760.004±0.0310.004 \pm0.031 / 0.042±0.1140.042 \pm0.1140.150±0.0690.150 \pm0.069 / 0.244±0.1010.244 \pm0.101

Table 5. Framediff – DSSP secondary-structure fractions. Each cell shows full-complex / ZEUS as mean ± std.

settinghelixsheetcoil
C2_1000.773±0.2140.773 \pm0.214 / 0.770±0.2080.770 \pm0.2080.082±0.1520.082 \pm0.152 / 0.088±0.1540.088 \pm0.1540.145±0.0790.145 \pm0.079 / 0.142±0.0650.142 \pm0.065
C3_1500.665±0.3030.665 \pm0.303 / 0.656±0.2700.656 \pm0.2700.160±0.2190.160 \pm0.219 / 0.163±0.2040.163 \pm0.2040.175±0.1000.175 \pm0.100 / 0.180±0.1160.180 \pm0.116
C4_2000.642±0.2520.642 \pm0.252 / 0.619±0.3150.619 \pm0.3150.163±0.1890.163 \pm0.189 / 0.178±0.2150.178 \pm0.2150.196±0.1160.196 \pm0.116 / 0.204±0.1340.204 \pm0.134
C6_3000.432±0.2480.432 \pm0.248 / 0.550±0.2730.550 \pm0.2730.294±0.1910.294 \pm0.191 / 0.216±0.1650.216 \pm0.1650.274±0.1110.274 \pm0.111 / 0.234±0.1530.234 \pm0.153
C8_4000.459±0.2830.459 \pm0.283 / 0.441±0.2730.441 \pm0.2730.310±0.2120.310 \pm0.212 / 0.315±0.1890.315 \pm0.1890.231±0.1180.231 \pm0.118 / 0.244±0.1360.244 \pm0.136
C12_6000.346±0.2730.346 \pm0.273 / 0.297±0.2730.297 \pm0.2730.382±0.2100.382 \pm0.210 / 0.444±0.1930.444 \pm0.1930.273±0.1230.273 \pm0.123 / 0.259±0.1240.259 \pm0.124
T_6000.723±0.2610.723 \pm0.261 / 0.670±0.2990.670 \pm0.2990.119±0.1780.119 \pm0.178 / 0.131±0.2310.131 \pm0.2310.158±0.1100.158 \pm0.110 / 0.199±0.1170.199 \pm0.117
O_12000.810±0.1070.810 \pm0.107 / 0.750±0.1390.750 \pm0.1390.020±0.0430.020 \pm0.043 / 0.041±0.0750.041 \pm0.0750.170±0.0880.170 \pm0.088 / 0.209±0.1170.209 \pm0.117

Table 6. RFdiffusion – DSSP secondary-structure fractions. Each cell shows full-complex / ZEUS as mean ± std.

Polyhedral Generation Experiments

The self-consistency evaluation is uninformative for polyhedral symmetries (we restrict ourselves to tetrahedral, octahedral, and icosahedral when we mention polyhedral as they are the representative set.) Folding models routinely fail on these complexes even at small sizes [5]. A failure to refold cannot distinguish between poor design from folding-model failure. Therefore we restrict the analysis of polyhedral generations to distributional comparisons only. We seek to determine equivalence of validity for ZEUS using proxy metrics for designability.

Setup. For tetrahedral (TT) and octahedral (OO) symmetries, we generate 50 designs at subunit length n=50n = 50 for each model (FrameDiff, FoldFlow, RFdiffusion), using full-complex baseline and ZEUS inference. Icosahedral (II) is omitted from this comparison because full-complex inference at n=50n = 50 does not fit on a single A100. Reducing the icosahedral subunit lengths even further is uninformative and biologically implausible [25] and we do not consider them. All reported uncertainties are 1-sigma standard deviations.

Metrics. Since designability cannot be directly evaluated, all metrics below are computed on the full set of 50 generated designs rather than restricted to a designable subset as in the cyclic evaluation.

  • Diversity and novelty. Defined as in the main experiments (§ 5), but computed across all 50 generated designs.

  • Radius of gyration (subunit and complex). RgR_g is an RMS distance of a protein’s atoms from its centre of mass. It is a measure of spatial extent and compactness:

Rgsubunit=∑i∈C∥ri−rˉ∥2R_g^{\mathrm{subunit}} = \sqrt{\sum_{i \in\mathcal{C}} \lVert\mathbf{r}_i - \bar{\mathbf{r}} \rVert^2}

where C\mathcal{C} is the set of backbone atoms in the canonical chain and rˉ\bar{\mathbf{r}} the chain centre of mass. The complex-level RgcomplexR_g^{\mathrm{complex}} is computed analogously over the union of all backbone atoms with respect to the assembly center of mass.

  • Clash score. Number of atom pairs separated by less than 1.5 Å and at least three residues apart in sequence.

  • Shell thickness. We measure the radial extent of the cage wall between the polyhedron’s inner cavity and outer surface. For backbone atoms at coordinates {ri}\{r_i\} in the assembly, with assembly centre of mass rˉ\bar{r}, let ri=∥ri−rˉ∥r_i = \lVert r_i - \bar{r} \rVert. Then

shell thickness=P99({ri})−P1({ri}),\text{shell thickness} = P_{99}(\{r_i\}) - P_{1}(\{r_i\}),

where PqP_q denotes the qq-th percentile.

  • Buried surface area per chain. The solvent-accessible surface area (SASA) of a structure is the total area accessible to a 1.4 Å solvent probe rolled over its van der Waals surface [Shrake and Rupley, 1973]. Per-chain BSA is the area of a chain’s surface occluded upon assembly: SASAisolated−SASAcomplex\mathrm{SASA}_{\mathrm{isolated}} - \mathrm{SASA}_{\mathrm{complex}}, with both terms restricted to backbone atoms and computed with biotite.structure.sasa. Since all subunits are identical by symmetry, we compute on chain A only.

  • Normalized BSA. The per-chain BSA divided by SASAisolated\mathrm{SASA}_{\mathrm{isolated}}, giving the fraction of a subunit’s surface buried at the assembly interface.

  • pdbTM (subunit and complex). Maximum TM-score against the PDB, returned by Foldseek at both the subunit and full-complex level.

Figure 12 compares the full-complex baseline against ZEUS across TT and OO symmetries for all three models. On global shape metrics such as shell thickness, per-chain BSA, BSA/SASA ratio, and radius of gyration, the two methods agree closely across all three models. Novelty shows a visible distributional shift: ZEUS-generated assemblies are slightly more PDB-distinct than the baseline. This direction is favorable for de novo design. Diversity distributions are nearly identical, with a maximum between-pathway difference of 0.0124 in TM-score (FoldFlow, octahedral). All remaining metrics are distributionally similar. We thus claim that the distributional equivalence established for cyclic symmetries in § 5 extends to polyhedral symmetries.

Distributional metrics for tetrahedral and octahedral generations across three backbones, showing boxplots for Rg assembly, Rg subunit mean, clash score, shell thickness, BSA per chain, BSA / SASA, pdbTM complex, pdbTM subunit, and pairwise TM diversity

Figure 12. Distributional metrics for tetrahedral and octahedral generations under full-complex baseline and ZEUS, across all three backbones. Grey boxplot is full-complex inference, coloured is ZEUS. Symmetries are labelled with total NresN_{\mathrm{res}}. ZEUS is distributionally equivalent.

Symmetric Generation Samples

Cyclic samples using FrameDiff aligned by symmetry type and method, comparing full-complex and ZEUS inference

Figure 13. Cyclic samples using FrameDiff with Lsubunit=50L_{\mathrm{subunit}} = 50 aligned by symmetry type (row) and method (2 samples per col) comparing full-complex and ZEUS inference.

Cyclic samples using FoldFlow with full-complex and ZEUS inference

Figure 14. Cyclic samples using FoldFlow with Lsubunit=50L_{\mathrm{subunit}} = 50 aligned by symmetry type (row) and method (2 samples per col) comparing full-complex and ZEUS inference.

Cyclic samples arranged by symmetry type C2, C3, C4, C6, C8, and C12, comparing RFdiffusion full-complex and ZEUS inference

Figure 15. Cyclic samples using RFdiffusion with Lsubunit=50L_{\mathrm{subunit}} = 50 aligned by symmetry type (row) and method (2 samples per col) comparing full-complex and ZEUS inference.

Polyhedral samples using FrameDiff with $L_{\mathrm{subunit}}=50$ aligned by symmetry type and method

Figure 16. Polyhedral samples using FrameDiff with Lsubunit=50L_{\mathrm{subunit}} = 50 aligned by symmetry type (row) and method (2 samples per col) comparing full-complex and ZeUS inference.

Polyhedral samples using FoldFlow, aligned by symmetry type and method

Figure 17. Polyhedral samples using FoldFlow with Lsubunit=50L_{\mathrm{subunit}} = 50 aligned by symmetry type (row) and method (2 samples per col) comparing full-complex and ZeUS inference.

Polyhedral samples using RFdiffusion aligned by symmetry type and inference method

Figure 18. Polyhedral samples using RFdiffusion with Lsubunit=50L_{\mathrm{subunit}} = 50 aligned by symmetry type (3 samples per row) and method (2 samples per col) comparing full-complex and ZEUS inference.

Large Scale Generations

We run ZEUS on FrameDiff with the following configurations: C50C_{50} at subunit length 200 and radius 225.3 Å; D50D_{50} at subunit length 100, radius 225.3 Å, and ring offset 28 Å; helical (HH) at subunit length 100, radius 79.875 Å, twist 0.31, and rise 1.30 Å; and icosahedral (II) at subunit length 100, outer radius 57 Å with C5C_{5}-axis radius 20 Å. Radius is defined as the median Cα_\alpha distance to the symmetry axis.

For each generated complex, we retrieve the closest matching natural assemblies via the RCSB Search v2 API, restricted to homomers of the matching symmetry class and filtered by minimum chain count: ≥20\geq20 chains for cyclic (n=268n = 268), ≥40\geq40 for dihedral (n=40n = 40), ≥30\geq30 for helical (n=267n = 267), and ≥60\geq60 for icosahedral (n=797n = 797).

Backbone-only buried surface area is computed as SASAisolated_{\mathrm{isolated}} − SASAcomplex_{\mathrm{complex}}, with both terms restricted to backbone atoms. We compute distributions with n=10n = 10 generations for each symmetry.

Figure 19 shows the top-5 retrieved assemblies for each generated complex. ZEUS generated structures are radially distinct from existing proteins and present plausible alternative large-scale geometries.

Top-5 most similar symmetric complexes retrieved from the PDB compared against ZEUS generations

Figure 19. Top k=5k = 5 most similar symmetric complexes retrieved from the PDB compared against ZEUS generations (left column).

Ethics Statement

We acknowledge that because our work enables large-scale symmetric protein design, if modified, it could potentially be used for the design of harmful viral capsid proteins by malicious actors. In order to mitigate this, in this existing work, we do not train or fine-tune on any symmetric complexes, and we expect that this will result in model outputs being unlikely to be biologically expressible as of now. In our future work, we will mitigate risks that could arise from misuse of our method for viral capsid design by carefully selecting our fine-tuning corpus to avoid particularly harmful viral strains. On the other hand, despite the potential risks that could arise from viral capsid generation, we expect that our method will also enable a number of potential benefits. For example, the ability to generate viral capsid proteins could prove useful in the study of viral mechanisms, in vaccine development, and in developing new methods for targeted drug delivery.

References

  1. [1]Jacob B. Bale, Shane Gonen, Yuxi Liu, William Sheffler, Daniel Ellis, Chantz Thomas, Duilio Cascio, Todd O. Yeates, Tamir Gonen, Neil P. King, and David Baker. Accurate design of megadalton-scale two-component icosahedral protein complexes. Science, 353(6297):389–394, 2016. doi: 10.1126/science.aaf8818. URL https://www.science.org/doi/abs/10.1126/science.aaf8818.DOI
  2. [2]Helen M Berman, John Westbrook, Zukang Feng, Gary Gilliland, Talapady N Bhat, Helge Weissig, Ilya N Shindyalov, and Philip E Bourne. The protein data bank. Nucleic acids research, 28(1):235–242, 2000.
  3. [3]Neville P. Bethel, Andrew J. Borst, Fabio Parmeggiani, Matthew J. Bick, T. J. Brunette, Hannah Nguyen, Alex Kang, Asim K. Bera, Lauren Carter, Marcos C. Miranda, Ryan D. Kibler, Mila Lamb, Xinting Li, Banumathi Sankaran, and David Baker. Precisely patterned nanofibres made from extendable protein multiplexes. Nature Chemistry, 15(12):1664–1671, Dec 2023. ISSN 1755-4349. doi: 10.1038/s41557-023-01314-x. URL https://doi.org/10.1038/s41557-023-01314-x.
  4. [4]Avishek Joey Bose, Tara Akhound-Sadegh, Guillaume Huguet, Kilian FATRAS, Jarrid Rector-Brooks, Cheng-Hao Liu, Andrei Cristian Nica, Maksym Korablyov, Michael M. Bronstein, and Alexander Tong. SE(3)-stochastic flow matching for protein backbone generation. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=kJF1H23hKb.arxiv.org/abs/2310.02391
  5. [5]Patrick Bryant, Gabriele Pozzati, Wensi Zhu, Aditi Shenoy, Petras Kundrotas, and Arne Elofsson. Predicting the structure of large protein complexes using alphafold and monte carlo tree search. Nature communications, 13(1):6028, 2022.DOI
  6. [6]Jasper Butcher, Rohith Krishna, Raktim Mitra, Rafael I. Brent, Yanjing Li, Nathaniel Corley, Paul T. Kim, Jonathan Funk, Simon Mathis, Saman Salike, Aiko Muraishi, Helen Eisenach, Tuscan Rock Thompson, Jie Chen, Yuliya Politanska, Enisha Sehgal, Brian Coventry, Odin Zhang, Bo Qiang, Kieran Didi, Max Kazman, Frank DiMaio, and David Baker. De novo design of all-atom biomolecular interactions with rfdiffusion3. bioRxiv, 2025.DOI
  7. [7]Andrew Campbell, Jason Yim, Regina Barzilay, Tom Rainforth, and Tommi Jaakkola. Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design. arXiv, 2024.arxiv.org/abs/2402.04997
  8. [8]Rees Chang, Angela Pak, Alex Guerra, Ni Zhan, Nick Richardson, Elif Ertekin, and Ryan Adams. Space group equivariant crystal diffusion. Advances in Neural Information Processing Systems, 38:72772–72805, 2026.arxiv.org/abs/2505.10994
  9. [9]Ruizhe Chen, Dongyu Xue, Xiangxin Zhou, Zaixiang Zheng, Xiangxiang Zeng, and Quanquan Gu. An all-atom generative model for designing protein complexes. arXiv preprint arXiv:2504.13075, 2025.arxiv.org/abs/2504.13075
  10. [10]Alexander E Chu, Jinho Kim, Lucy Cheng, Gina El Nesr, Minkai Xu, Richard W Shuai, and Po-Ssu Huang. An all-atom protein generative model. Proceedings of the National Academy of Sciences, 121(27):e2311500121, 2024.DOI
  11. [11]J. Dauparas, I. Anishchenko, N. Bennett, H. Bai, R. J. Ragotte, L. F. Milles, B. I. M. Wicky, A. Courbet, R. J. de Haas, N. Bethel, P. J. Y. Leung, T. F. Huddy, S. Pellock, D. Tischer, F. Chan, B. Koepnick, H. Nguyen, A. Kang, B. Sankaran, A. K. Bera, N. P. King, and D. Baker. Robust deep learning–based protein sequence design using proteinmpnn. Science, 378(6615):49–56, 2022. doi: 10.1126/science.add2187. URL https://www.science.org/doi/abs/10.1126/science.add2187.DOI
  12. [12]Robby Divine, Ha V. Dang, George Ueda, Jorge A. Fallas, Ivan Vulovic, William Sheffler, Shally Saini, Yan Ting Zhao, Infencia Xavier Raj, Peter A. Morawski, Madeleine F. Jennewein, Leah J. Homady, Yu-Hsin Wan, Marti R. Tooley, Franziska Seeger, Ali Etemadi, Mitchell L. Fahning, James Lazarovits, Alex Roederer, Alexandra C. Walls, Lance Stewart, Mohammadali Mazloomi, Neil P. King, Daniel J. Campbell, Andrew T. McGuire, Leonidas Stamatatos, Hannele Ruohola-Baker, Julie Mathieu, David Weesler, and David Baker. Designed proteins assemble antibodies into modular nanocages. Science, 372(6537):eabd9994, 2021. doi: 10.1126/science.abd9994. URL https://www.science.org/doi/abs/10.1126/science.abd9994.DOI
  13. [13]Jose M Duarte, Shuchismita Dutta, David S Goodsell, and Stephen K Burley. Exploring protein symmetry at the rcsb protein data bank. Emerging topics in life sciences, 6(3):231–243, 2022.DOI
  14. [14]Mark A. Herzik, Mengyu Wu, and Gabriel C. Lander. High-resolution structure determination of sub-100 kda complexes using conventional cryo-em. Nature communications, 10(1):1032, 2019.DOI
  15. [15]Timothy F. Huddy, Yang Hsia, Ryan D. Kibler, Jinwei Xu, Neville Bethel, Deepesh Nagarajan, Rachel Redler, Philip J. Y. Leung, Connor Weidle, Alexis Courbet, Erin C. Yang, Asim K. Bera, Nicolas Coudray, S. John Calise, Fatima A. Davila-Hernandez, Hannah L. Han, Kenneth D. Carr, Zhe Li, Ryan McHugh, Gabriella Reggiano, Alex Kang, Banumathi Sankaran, Miles S. Dickinson, Brian Coventry, T. J. Brunette, Yulai Liu, Justas Dauparas, Andrew J. Borst, Damian Ekiert, Justin M. Kollman, Gira Bhabha, and David Baker. Blueprinting extendable nanomaterials with standardized protein blocks. Nature, 627(8005):898–904, Mar 2024. ISSN 1476-4687. doi: 10.1038/s41586-024-07188-4. URL https://doi.org/10.1038/s41586-024-07188-4.
  16. [16]Guillaume Huguet, James Vuckovic, Kilian Fatras, Eric Thibodeau-Laufer, Pablo Lemos, Riashat Islam, Cheng-Hao Liu, Jarrid Rector-Brooks, Tara Akhound-Sadegh, Michael Bronstein, Alexander Tong, and Avishke Joey Bose. Sequence-augmented se(3)-flow matching for conditional protein backbone generation. Advances in neural information processing systems, 37:33007–33036, 2024.arxiv.org/abs/2405.20313
  17. [17]John B. Ingraham, Max Baranov, Zak Costello, Karl W. Barber, Wujie Wang, Ahmed Ismail, Vincent Frappier, Dana M. Lord, Christopher Ng-Thow-Hing, Erik R. Van Vlack, Shan Tie, Vincent Xue, Sarah C. Cowles, Alan Leung, João V. Rodrigues, Claudio L. Morales-Perez, Alex M. Ayoub, Robin Green, Katherine Puentes, Frank Oplinger, Nishant V. Panwar, Fritz Obermeyer, Adam R. Root, Andrew L. Beam, Frank J. Poelwijk, and Gevorg Grigoryan. Illuminating protein space with a programmable generative model. Nature, 623(7989):1070–1078, Nov 2023. ISSN 1476-4687. doi: 10.1038/s41586-023-06728-8. URL https://doi.org/10.1038/s41586-023-06728-8.
  18. [18]Rui Jiao, Wenbing Huang, Yu Liu, Deli Zhao, and Yang Liu. Space group constrained crystal generation. In International Conference on Learning Representations, volume 2024, pages 6836–6853, 2024.arxiv.org/abs/2402.03992
  19. [19]John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Židek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardo Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W. Senior, Koray Kavukcuoglu, Pushmeet Kohli, and Demis Hassabis. Highly accurate protein structure prediction with alphafold. Nature, 596(7873):583–589, Aug 2021. ISSN 1476-4687. doi: 10.1038/s41586-021-03819-2. URL https://doi.org/10.1038/s41586-021-03819-2.
  20. [20]Nikita Kazeev, Wei Nong, Ignat Romanov, Ruiming Zhu, Andrey Ustyuzhanin, Shuya Yamazaki, and Kedar Hippalgaonkar. Wyckoff transformer: Generation of symmetric crystals. arXiv preprint arXiv:2503.02407, 2025.arxiv.org/abs/2503.02407
  21. [21]Dejun Lin, Simon Chu, Vishanth Iyer, Youhan Lee, John St John, Kevin Boyd, Brian Roland, Xiaowei Ren, Guoqing Zhou, Zhonglin Cao, Polina Binder, Yuliya Zhautouskaya, Jakub Zakrzewski, Maximilian Stadler, Kyle Gion, Yuxing Peng, Xi Chen, Tianjing Zhang, Philipp Junk, Michelle Dimon, Pawel Gniewek, Fabian Ortega, McKinley Polen, Ivan Grubisic, Ali Bashir, Graham Holt, Danny Kovtun, Matthias Grass, Luca Naef, Rui Wang, Jian Peng, Anthony Costa, Saee Paliwal, Eddie Calleja, Timur Rvachov, Neha Tadimeti, Roy Tal, and Emine Kucukbenli. Fold-cp: A context parallelism framework for biomolecular modeling. arXiv preprint arXiv:2603.14806, 2026a.arxiv.org/abs/2603.14806
  22. [22]Ning Lin, Luxi Chen, Huaquan Chen, Jiacheng Cen, Chongxuan Li, Wenbing Huang, and Hao Sun. Planar symmetric pattern generation. arXiv preprint arXiv:2606.02073, 2026b.arxiv.org/abs/2606.02073
  23. [23]Zeming Lin, Halil Akin, Roshan Rao, Brian Hie, Zhongkai Zhu, Wenting Lu, Nikita Smetanin, Robert Verkuil, Ori Kabeli, Yaniv Shmueli, Allan dos Santos Costa, Maryam Fazel-Zarandi, Tom Sercu, Salvatore Candido, and Alexander Rives. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science, 379(6637):1123–1130, 2023.DOI
  24. [24]Sidney Lyayuga Lisanza, Jacob Merle Gershon, Samuel W. K. Tipps, Jeremiah Nelson Sims, Lucas Arnoldt, Samuel J. Hendel, Miriam K. Simma, Ge Liu, Muna Yase, Hongwei Wu, Claire D. Tharp, Xinting Li, Alex Kang, Evans Brackenbrough, Asim K. Bera, Stacey Gerben, Bruce J. Wittmann, Andrew C. McShan, and David Baker. Multifaceted and functional protein design using rosettafold sequence diffusion. Nature Biotechnology, 43(8):1288–1298, Aug 2025. ISSN 1546-1696. doi: 10.1038/s41587-024-02395-w. URL https://doi.org/10.1038/s41587-024-02395-w.
  25. [25]Anže Lošdorfer Božič, Antonio Šiber, and Rudolf Podgornik. Statistical analysis of sizes and shapes of virus capsids and their resulting elastic properties. Journal of biological physics, 39(2):215–228, 2013.DOI
  26. [26]Omri Puny, Yaron Lipman, and Benjamin Kurt Miller. Space group conditional flow matching. arXiv preprint arXiv:2509.23822, 2025.arxiv.org/abs/2509.23822
  27. [27]RCSB Protein Data Bank. PDB statistics: Homo-oligomers global assembly symmetries per year. https://www.rcsb.org/stats/symmetry/growth, 2025. Accessed: 2026-05-07.

Paper details

vlnc:W_k5cwhm5k
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 11

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Agricultural and Biological Sciences

Not recorded for this paper

View statement
Looking for related papers…
Contents