WorksComputer Science
Blitz: Accelerating the Diffusion Modules of AlphaFold3 and Boltz-2
Abstract
Diffusion-based protein structure prediction models produce all-atom structures in about 200 denoising steps, and that cost has driven a number of few-step samplers. We present Blitz, a few-step distillation framework that uses a geometric noise schedule to distribute steps more evenly across noise scales, then learns those updates through multistep consistency training. Physical steering and target-free candidate selection help reduce the remaining geometric defects at inference. We apply Blitz to AlphaFold 3 and Boltz-2, reducing denoising from 200 steps to eight. Because published few-step results use different benchmarks and scoring procedures, we compare both distilled models with their teachers and other few-step methods under a common evaluation protocol covering structural accuracy, interface quality, ligand placement, and stereochemical validity. Blitz retains teacher-level structural accuracy, with AF3-Blitz outperforming all evaluated few-step baselines and exceeding its teacher on several accuracy metrics.
Introduction
Diffusion-based atom prediction has been a major breakthrough in the biological sciences, with models such as AlphaFold3 {(citep}(Abramson et al., 2024)}, Boltz-1 and Boltz-2 {(citep}(Wohlwend et al., 2024; Passaro et al., 2025)}, Chai-1 {(citep}(Chai Discovery et al., 2024)}, and Protenix {(citep}(ByteDance AML AI4Science Team et al., 2025)}. It builds on progress in protein folding that led CASP organizers to describe AlphaFold2 as a solution to the decades-old folding problem for single protein chains {(citep}(Abbass, 2025)}. However, these models are expensive in pipelines that require tens of thousands of predictions, leading to a growing literature on reducing their cost. These efforts primarily tackle three parts of the pipeline:
Reducing the cost of constructing multiple sequence alignments (MSAs). MMseqs2 and ColabFold accelerate sequence search {(citep}
(Steinegger & Söding, 2017; Mirdita et al., 2022)}, while Kallenborn et al. [12] move the search onto GPUs.
Reducing the number of recycle updates and the computational cost of the Pairformer trunk {(citep}
(Gong et al., 2025; Zhang et al., 2026)}.
Reducing the cost of the diffusion model itself, which typically uses 200 denoising steps in AlphaFold3 and Boltz-2 {(citep}
(Abramson et al., 2024; Passaro et al., 2025)}.
On the model side, these costs have been approached in several ways. Gong et al. [10] introduced Protenix-Mini, reducing both the cost of the trunk and the number of diffusion sampling steps. Zhang et al. [24] introduced DCFold, whose Temporal Geodesic Matching schedule pairs diffusion times during consistency training to improve stability. They also apply consistency training to the Pairformer to reduce recycling. Scarpellini et al. [18] introduced DeCAF, which distills all-atom flow maps in the noise coordinate used by EDM and applies reward-guided search at inference. They report strong results in both protein–ligand pose accuracy and physical validity.
However, we show here that differences in scoring conventions and benchmark construction make these algorithms difficult to compare. Claims of preserved quality do not fully hold under our unified evaluation. This is especially important for few-step samplers, where steric clashes, distorted bonds, and invalid bond angles can coexist with high structural similarity scores. Our matched evaluations show that commonly reported accuracy metrics can obscure these flaws when stereochemical checks are omitted (Table 1 and Figure 1).
Figure 1. Selected 8V9Q example (Collette et al. [9]): experimental reference, (a) released DeCAF, and (b) its Boltz-1x teacher. Scores show lDDT without → with stereochemical checks.
| Sampler | NFE | PXM ↑ lDDT | OST lDDT ↑ Unchecked | OST lDDT ↑ Checked | Events ↓ /target | lDDT-PLI ↑ | PB-valid ↑ valid |
| AF3 teacher | 200 | 0.8554 | 0.8626 | 0.8615 | 0.7 | 0.6446 | 44.8% |
| AF3-Blitz [8] | 8 | 0.8579 | 0.8653 | 0.8635 | 4.8 | 0.6430 | 42.7% |
| Protenix | 200 | 0.8512 | 0.8593 | 0.8557 | 4.3 | 0.6259 | 45.8% |
| Protenix one-step ODE | 1 | 0.8503 | 0.8591 | 0.5426 | 1563.6 | 0.5774 | 26.0% |
| Boltz-2x teacher | 600 | 0.8426 | 0.8514 | 0.8499 | 4.4 | 0.5596 | 60.9% |
| Boltz-2-Blitz [8] | 8 | 0.8451 | 0.8542 | 0.8492 | 14.2 | 0.5545 | 52.0% |
| Boltz-1x teacher | 600 | 0.8272 | 0.8357 | 0.8074 | 35.8 | 0.5932 | 58.0% |
| DeCAF default | 10 | 0.8254 | 0.8340 | 0.6756 | 681.3 | 0.5872 | 25.0% |
| DeCAF SDE () | 10 | 0.8166 | 0.8250 | 0.6635 | 680.6 | 0.5810 | 26.6% |
Table 1. Accuracy with and without stereochemical checks. Bold highlights geometry failures. NFE counts steering particles per prediction. Evaluation details are in Appendix B.
To address these gaps in stereochemical quality and structural accuracy, we present Blitz, a few-step distillation framework. We first show that the default polynomial noise schedule of Karras et al. [13] is ill-suited to the few-step molecular sampling regime studied here, and replace it with a geometric schedule. We then address the remaining accuracy gap with multistep consistency training on this grid, adapting the method of Heek et al. [11]. The eight-step trajectory also leaves room for inference-time steering, as used in Boltz-2 (Passaro et al. [17]). We tune the steering strengths for Boltz-2-Blitz and bring its reference-free validity closer to that of its teacher. Finally, we benchmark structural accuracy and stereochemical quality on three popular datasets used in this literature: the Boltz-2 benchmark ( targets), Runs N’ Poses, and RecentPDB384 ([17, 19, 7]). Among the compared configurations, AF3-Blitz achieves the highest stereochecked complex lDDT on the primary benchmark. Both students retain teacher-level structural accuracy, exceeding their teachers on several reported metrics. We apply the same framework to AlphaFold3 and Boltz-2 to show generality across two publicly available diffusion-based structure prediction models.
In summary, our core contributions are:
A diagnosis. We show that prevailing scoring and benchmark conventions systematically mask the geometric failures of few-step structure predictors, and we expose these failures under matched evaluation.
A method. We introduce Blitz, which pairs a geometric noise schedule and multistep consistency training with inference-time physical controls to reach teacher-level structural accuracy in eight sampling steps, a reduction in sampling steps for both AlphaFold3 and Boltz-2.
A protocol. We establish a standardized, multi-metric evaluation across three benchmarks spanning structural accuracy, interface quality, ligand placement, and stereochemistry, designed to surface the failure modes that single-metric reporting hides.
Background: diffusion and consistency distillation
Diffusion in AlphaFold3. AF3 adopts the EDM formulation and polynomial noise schedule of Karras et al. [13] for all-atom structure prediction (Abramson et al. [2], Supplementary Information, Sections 3.7–3.7.1). Its diffusion module is a non-equivariant transformer conditioned on the single and pair representations from the trunk. Let contain the resolved atomic coordinates and the molecular features and trunk conditioning. In these variance-exploding coordinates, a training input is
The denoiser predicts clean coordinates. AF3 uses EDM input/output preconditioning and samples training noise as , with and Å. The trunk is evaluated once and reused across multiple noised copies; the coordinate loss uses weighted rigid alignment.
The EDM noise schedule. Sampling begins at high noise and follows a decreasing sequence of noise levels. Karras et al. space these levels polynomially. For positive nodes, we write
AF3 uses , with absolute noise endpoints Å and Å ([2 Supplementary Eq. 7]). EDM selected through experiments on image generation.
Consistency models. Consistency models learn to map points on the same probability-flow trajectory to its low-noise endpoint ([21]). For two states on that trajectory, the desired property is , with the boundary condition at a small positive endpoint . In consistency distillation, a frozen diffusion teacher takes a numerical step to form a pair of nearby states. Training brings their student predictions together, stopping gradients through the target branch. Consistency training instead forms pairs directly from noised data, without a pretrained teacher ([21, 20]).
Multistep consistency models. Heek et al. [11] unify TRACT [4] and consistency models [21] by dividing the trajectory into fixed segments. A shared model maps each point to the lower boundary of its segment. Generation traverses the segments in order, using one evaluation per segment. Their formulation uses variance-preserving coordinates,
and constructs the segment maps with DDIM updates.
What does it mean to preserve quality in structure prediction?
The prediction in Figure 1a has severe overlap between its protein chains. Despite having recognizable individual folds, this arrangement is sterically implausible; however, it still scores well on unchecked lDDT. Its Boltz-1x teacher, shown in Figure 1b, retains recognizable folds without this severe overlap; this raises an important question: what must accelerated structure prediction preserve when we remove these denoising steps?
This is particularly important when there is a claim that a faster sampler preserves accuracy. Unchecked lDDT measures how well local interatomic distances are preserved from the reference. However, this local fold preservation can coexist with severe clashes and violations of bond lengths and angles. The original lDDT authors anticipated this distinction and provided optional stereochemical checks to account for these violations ([14]). PoseBusters showed that agreement with an experimental ligand pose does not necessarily establish physical plausibility ([6]). However, evaluations like those done by DCFold report only RMSD thresholds on the PoseBusters benchmark without reporting PB-validity, leaving the failure mode that motivated those checks unexamined ([24]). This concern also applies to cofolding. AlphaFold3 documents overlapping chains ([2]), and DeCAF reports ligand validity alongside pose accuracy ([18]). Figure 1 helps illustrate the discrepancy at the protein-assembly level.
Table 1 helps make this discrepancy in the evaluated configurations more explicit. On matched targets from the Boltz-2 benchmark (2,300 targets), the default Protenix sampler and its one-step ODE counterpart are almost identical on unchecked lDDT. DCFold uses a corresponding Protenix-based one-step ODE configuration as its untrained AF3 ODE baseline ([24]). However, with stereochemical checks enabled, our one-step configuration’s score falls approximately 32 points on a 100-point scale, while the default sampler’s score changes only slightly. Furthermore, the one-step configuration has, on average, approximately 1,500 counted bond-length, bond-angle, and clash violations per target.
PXM: Protenix’s PXMETER lDDT scorer. OST: OpenStructure’s compare-structures scorer, which enables stereochemical checks by default; “Unchecked” disables them.
This gap is one part of the picture. DockQ, PoseBusters validity, ICS, IPS, and lDDT-PLI provide complementary information alongside stereochemically checked lDDT. In essence, we argue that these measures should be reported where applicable, with their definitions and evaluation settings made explicit, particularly for fast samplers to provide an accurate picture. Appendix B gives our scoring rules and check coverage.
Blitz
Noise allocation before distillation
Before detailing our distillation procedure, we identify a problem with the noise schedule in the low-step regime. Following the deterministic analysis of Karras et al. [13], we hold the pretrained AF3 denoiser and solver fixed and replace the default Karras schedule with a geometric schedule. Stereochemically checked lDDT improves while bond violations and clashes decrease, as shown in Figure 2. Critically, the improvement is largest at the smallest step budgets. The same pattern holds across every publicly available structure generation model we evaluated.
Figure 2. Allocation matters most at low budgets. The same untrained AF3 denoiser is evaluated with the default and geometric schedules at matched step budgets and positive endpoints.
The geometric schedule is the limit of Equation (2):
It spaces the positive nodes uniformly in log noise, giving the same relative noise reduction between adjacent nodes. In § 5, we compare students trained identically except for the noise schedule: polynomial with or geometric.
Local approximation error and endpoint damage
We next ask: Why does changing the schedule improve molecular geometry? With fewer steps, each update must cover a larger noise interval. We first measure what is lost by crossing that interval in one step. On selected diagnostic targets, we start at the same state and compare one Euler step with a Heun calculation that divides the interval into smaller steps:
This estimates the coordinate error introduced by the coarse step, measured in the same frame over valid atoms.
The next question is: Does this error damage the final molecule? We continue both endpoints with the same refined sampler , including the same terminal step, and measure the difference in final quality:
Here is stereochemically checked lDDT or a bond, angle, or clash violation count. Changing only one interval isolates its effect on the final structure.
Indeed, the final positive interval of the default schedule causes the largest average loss of stereochemically checked lDDT, even though an earlier interval causes a larger final coordinate displacement (Figure 3). Coordinate error alone therefore does not explain the damage to molecular geometry. This gives us a reason to redistribute the noise intervals before distillation, as the geometric schedule does. Appendix § 18 gives the full analysis and checks on the numerical reference.
Figure 3. Local approximation error and final molecular quality. A: Euler error relative to a refined Heun calculation. B–C: final effects of replacing one refined interval with an Euler step. Details are in Appendix I.
Multistep consistency distillation
We adapt multistep consistency distillation [11] to the variance-exploding EDM coordinates used by AF3 and Boltz-2. The geometric grid divides the trajectory into fixed segments, with boundaries . We refine a separate teacher grid during training, keeping these segment boundaries fixed. We sample an adjacent edge using EDM log-normal importance weights, then identify its student segment.
DCFold reports instability with conventional consistency training ([24]), while DeCAF finds poor physical validity with its consistency-distillation baseline ([18]). Relaxing the one-step requirement shortens the mapping the student must learn. In our setting, multistep distillation works with an aligned coordinate-space loss and ordinary EDM edge sampling, without DCFold’s Temporal Geodesic Matching.
A short Heun step of the frozen teacher is followed by a stopped-gradient student update to the segment boundary. Inverting Equation (8) with gives the denoiser target in Algorithm 1. Here denotes molecular input features, the reference coordinates, and the atom weights, zero for unresolved or padded atoms. The conditioning includes and the frozen trunk embeddings. We apply weighted Kabsch alignment and a pseudo-Huber loss; Appendix A defines the loss and boundary updates.
Algorithm 1 Blitz multistep consistency distillation
Input: training set , frozen trunk , frozen teacher , student initialized from , segment boundaries
for each optimizer update do
Sample , augment , and set
Sample an edge from the current teacher grid
Set for the segment containing , and
Sample and set
if ; otherwise
Average over noised copies and training examples; update
end for
The student is initialized from its teacher; only the diffusion module is trained. Each trunk evaluation conditions multiple noised copies. At inference, Equation (8) composes the learned maps. Appendix A gives training parameters, boundary handling, and deployment scales.
Physical steering and target-free selection
The Boltz-2 students use physical steering during sampling, with strengths approximated from the teacher’s coordinate corrections. A correction to the denoised estimate changes the next state by , where and
For each potential, we sum correction vectors before taking their RMS and compare transferred corrections in matched log-noise bins. This guides the weights, which we then validate using accuracy and geometry. This steering did not work well for AF3 in our tests.
Blitz reranks five candidates without reference structures or further diffusion calls. We prefer its lower geometry-violation rate; Appendix G.3 compares native confidence.
Data and evaluation
We evaluate Blitz using AlphaFold3 [2] and Boltz-2 [17] to assess generality across two all-atom structure prediction systems. The schedule improvements in § 4.1–§ 4.2 also hold for Chai-1 and Protenix [8, 7].
Benchmarks and reporting. We evaluate the Boltz-2 benchmark (2,300 targets), Runs N’ Poses, and Protenix’s 384-target recent-PDB sample [17, 19, 7]. Main configurations use three recycles and five candidates. For brevity, the main text focuses on the Boltz-2 benchmark (2,300 targets). Results for Runs N’ Poses and RecentPDB384, including ligand-evaluation coverage limitations, are reported in Appendices D and E.
Accuracy and physical quality. We use OpenStructure [5] to score complex and backbone lDDT, interfaces, ligand placement, and bond, angle, and clash violations. Complex lDDT includes stereochemical checks [14]. Ligand validity is evaluated separately with DeCAF-compatible PoseBusters [6]. Success requires the same primary ligand to pass PB-valid and RMSD Å in the same candidate. We also report the fraction of targets with geometry defects. Failed and inapplicable evaluations remain separate from invalid predictions. Appendix H reports confidence intervals for mean scores and paired teacher–student differences.
Accuracy and geometry with fewer steps
NFE is per prediction; assigned and primary denote ligand RMSD. All Boltz-2 configurations use three recycle updates and steering; step scales are given in Appendix § 10.§ 2. Evaluation details are in Appendix § 11.
Table 2 shows that both eight-step Blitz students retain much of their teachers’ structural accuracy with a fraction of the sampling steps. The four- and sixteen-step AF3 students show how this trade-off changes with the sampling budget. The composition breakdown in Appendix § 12.§ 1 shows higher checked lDDT for both AF3 students on protein-only and protein–ligand complexes; gains are not uniform across nucleic-acid-containing complexes.
| Metric | AlphaFold3 Teacher | AlphaFold3 | AlphaFold3 Blitz [4] | AlphaFold3 Blitz [8] | AlphaFold3 Blitz [16] | Boltz-2 Teacher | Boltz-2 Blitz [8] | Boltz-2 Blitz [16] |
| NFE | 200 | 8 | 4 | 8 | 16 | 600 | 8 | 16 |
| Complex lDDT | 0.8615 | 0.8603 | 0.8559 | 0.8635 | 0.8634 | 0.8499 | 0.8492 | 0.8533 |
| Backbone lDDT | 0.9136 | 0.9151 | 0.9128 | 0.9150 | 0.9151 | 0.9043 | 0.9055 | 0.9069 |
| TM-score | 0.8898 | 0.8884 | 0.8889 | 0.8888 | 0.8894 | 0.8753 | 0.8768 | 0.8777 |
| DockQ | 0.4198 | 0.4133 | 0.4185 | 0.4207 | 0.4167 | 0.3929 | 0.3897 | 0.3932 |
| QS-global | 0.7784 | 0.7756 | 0.7805 | 0.7768 | 0.7784 | 0.7374 | 0.7347 | 0.7355 |
| ICS | 0.4263 | 0.4211 | 0.4288 | 0.4263 | 0.4247 | 0.3884 | 0.3857 | 0.3864 |
| IPS | 0.4560 | 0.4538 | 0.4573 | 0.4574 | 0.4570 | 0.4225 | 0.4224 | 0.4230 |
| lDDT-PLI | 0.6437 | 0.6434 | 0.6372 | 0.6425 | 0.6437 | 0.5596 | 0.5545 | 0.5573 |
| Ligand RMSD () | 6.63 | 6.51 | 6.44 | 6.44 | 6.51 | 9.00 | 8.75 | 8.78 |
| Assigned (%) | 61.3 | 61.4 | 61.0 | 61.3 | 61.4 | 52.2 | 51.6 | 52.2 |
| PB-valid (%) | 45.0 | 35.2 | 34.7 | 43.0 | 45.7 | 60.9 | 52.0 | 55.9 |
| Primary (%) | 68.0 | 67.3 | 67.5 | 68.0 | 68.3 | 60.9 | 60.3 | 60.7 |
| Success (%) | 34.7 | 29.6 | 30.4 | 35.8 | 35.9 | 36.5 | 30.6 | 33.9 |
| RF-valid (%) | 41.5 | 36.4 | 36.9 | 40.5 | 42.3 | 98.5 | 93.7 | 94.1 |
| Bonds / target | 0.06 | 2.84 | 12.84 | 0.79 | 0.10 | 0.90 | 4.83 | 1.79 |
| Angles / target | 0.09 | 1.02 | 4.00 | 0.23 | 0.08 | 0.71 | 3.05 | 1.23 |
| Clashes / target | 0.54 | 4.03 | 24.06 | 3.80 | 1.60 | 2.82 | 6.30 | 3.63 |
| Any violation (%) | 9.8 | 50.6 | 57.5 | 11.3 | 6.8 | 30.3 | 41.9 | 26.7 |
Table 2. Accuracy and geometry across sampling budgets.
We return to the example that opened Section 3. Figure 4 shows that AF3-Blitz retains the individual folds without the severe chain overlap seen in Figure 1. Its stereochemically checked score also remains close to its unchecked score. This is the improvement we want from a fast sampler: preserving the fold without the same collapse in molecular geometry.
Figure 4. The same 8V9Q example as Figure 1: reference, (a) AF3-Blitz, and (b) its AF3 teacher. Scores show unchecked stereochemically checked lDDT.
This distinction also appears in the comparison with other few-step samplers (Table 3). The evaluated DeCAF configurations and Protenix one-step ODE baseline have backbone scores much closer to Blitz than their stereochemically checked all-atom scores. The fold alone does not capture the difference in molecular geometry.
| Metric | Boltz-1x | DeCAF default | DeCAF SDE | DeCAF FK | Protenix | Protenix ODE | Protenix Mini |
| NFE | 600 | 10 | 10 | 40 | 200 | 1 | 1 |
| Complex lDDT | 0.8051 | 0.6741 | 0.6607 | 0.6654 | 0.8533 | 0.5462 | 0.6571 |
| Backbone lDDT | 0.8928 | 0.8869 | 0.8784 | 0.8791 | 0.9071 | 0.9039 | 0.8938 |
| TM-score | 0.8708 | 0.8622 | 0.8581 | 0.8577 | 0.8807 | 0.8780 | 0.8680 |
| DockQ | 0.3868 | 0.3702 | 0.3626 | 0.3636 | 0.4077 | 0.4026 | 0.3684 |
| QS-global | 0.7332 | 0.7247 | 0.7196 | 0.7257 | 0.7561 | 0.7483 | 0.7183 |
| ICS | 0.3842 | 0.3740 | 0.3698 | 0.3757 | 0.4071 | 0.3996 | 0.3686 |
| IPS | 0.4228 | 0.4171 | 0.4141 | 0.4177 | 0.4393 | 0.4399 | 0.4198 |
| lDDT-PLI | 0.5934 | 0.5874 | 0.5810 | 0.5876 | 0.6244 | 0.5780 | 0.5781 |
| Ligand RMSD (Å) | 7.01 | 6.97 | 7.11 | 7.03 | 7.00 | 6.68 | 7.46 |
| Assigned < 2Å (%) | 57.3 | 56.0 | 56.1 | 56.5 | 59.5 | 54.9 | 54.3 |
| PB-valid (%) | 58.2 | 24.7 | 26.7 | 53.9 | 45.8 | 26.7 | 28.2 |
| Primary < 2Å (%) | 66.3 | 65.7 | 65.8 | 65.6 | 69.4 | 65.7 | 63.8 |
| Success (%) | 39.4 | 21.1 | 21.8 | 38.4 | 34.0 | 25.0 | 24.0 |
| RF-valid (%) | 99.0 | 26.0 | 27.5 | 66.8 | 32.7 | 30.9 | 31.4 |
| Bonds / target | 11.59 | 268.56 | 271.91 | 280.02 | 0.19 | 666.44 | 373.96 |
| Angles / target | 5.45 | 123.85 | 131.80 | 138.37 | 0.58 | 283.84 | 129.20 |
| Clashes / target | 18.79 | 288.56 | 276.60 | 228.94 | 3.71 | 636.58 | 180.51 |
| Any violation (%) | 95.0 | 97.0 | 97.5 | 97.7 | 49.1 | 100.0 | 99.95 |
Table 3. Other sampler configurations. Metrics match Table 2.
DeCAF default uses the released ten-step sampler (); its SDE variant uses without guidance. DeCAF FK uses four particles and ten steps, requiring 40 denoiser evaluations per retained pose. FK steering improves ligand validity, while substantial protein-geometry violations remain.
DeCAF FK uses four steering particles; Boltz-1x uses three. Evaluation details follow Table 2.
What does the average hide?
The polynomial () and geometric students use the same training setup, differing only in the noise schedule. They have close average lDDT, but their errors are distributed very differently. Figure 5 compares them on the same targets using the same ranking rule. Most predictions from the geometric student avoid counted geometry violations altogether. The student’s defects are spread across many more targets, even though the mean clash counts are close.
Figure 5. Similar averages, different failure patterns. A: paired lDDT changes (geometric minus ); zero denotes equal scores. B–C: defect frequency and violation-count tails on matched Boltz-2 benchmark targets. A uses a signed-log change axis; C uses log-scaled counts and percentages, with a linear region near zero count. Zero counted violations does not imply full chemical validity.
Indeed, an average alone misses both parts of this result: how often a structure is damaged and how badly it is damaged when it fails. The paired accuracy changes and the violation tail show the remaining trade-off. Appendices F and G examine how target size and individual outliers affect these results.
Conclusion and future work
This work addresses a central problem in evaluating fast structure prediction: high structural-similarity scores can obscure failures in molecular geometry. We combine a geometric noise schedule with multistep consistency training, physical steering, and reference-free candidate selection. Across AlphaFold 3 and Boltz-2, this approach retains much of the teachers’ structural accuracy under stereochemically checked evaluation. With the full inference procedure, most predictions from the sixteen-step AF3 student have no counted geometry defects, and fewer predictions are affected than for the corresponding teacher. Its remaining clash burden is concentrated in a relatively small subset of targets. Several accuracy measures also improve over the teacher; our paired comparisons and uncertainty estimates distinguish these gains from differences that remain uncertain.
Blitz provides a practical alternative to the expensive diffusion stage in workflows that repeatedly call a structure predictor, including protein design, virtual screening, and antibody candidate evaluation. Our speedup applies to this stage rather than the full prediction pipeline; the trunk remains unchanged and still represents a substantial computational cost. Future work could combine few-step diffusion with methods that reduce trunk computation or strengthen physical constraints during sampling.
Our results show that substantial acceleration need not come at the expense of structural accuracy or molecular geometry, provided both are considered in training, inference, and evaluation. Average similarity scores alone cannot establish that quality has been preserved. Evaluation should make clear what is preserved, how often predictions fail, and how severe those failures are. Claims of successful acceleration should account for all three.
Acknowledgments
We thank the members of the Kumara and Lai labs for their comments and feedback. We also thank Xiowen You, Ankur Verma, and Daniel Cullina for helpful discussions and suggestions, and Charles Pavloski (NCSA NSF-OAC2320345 & NSF-OAC2005572) for feedback and troubleshooting. We gratefully acknowledge funding support from the Cornell University BRC Epigenomics Core Facility (RRID:SCR_021287), the Penn State Institute for Computational and Data Sciences (RRID:SCR_025154), and Penn State AIMI (RRID:SCR_022867). Soundar R. Kumara would also like to thank the Allen E. Pearce and Allen M. Pearce Professorship. This project was supported by NIH grant R35GM155380 to W.K.M.L. Computational resources were provided through an NSF ACCESS Maximize allocation (CIS260116).
Use of Generative AI
We used Claude and ChatGPT as coding assistants for supporting data-processing, evaluation, scoring, and plotting scripts, and for parts of Blitz, including the training loop in Algorithm 1. We also used these tools to correct grammar and improve sentence structure in selected passages. They were not used to generate research ideas, propose or refine hypotheses, design the methodology or experiments, or conduct or interpret analyses. The manuscript’s content, structure, and arguments are the authors’ own. We take full responsibility for the final text, code, and claims.
Ethics Statement
This work develops computational methods for accelerating biomolecular structure prediction and does not involve human subjects, personal data, or clinical decision-making. AlphaFold 3 experiments were conducted in a non-commercial academic setting using model parameters obtained through the official access process and subject to the applicable terms of use. We do not redistribute AlphaFold 3 model parameters or AF3-derived Blitz checkpoints; users of the released AF3-related code must obtain authorized parameters independently. Boltz2-Blitz parameters and the Blitz training and inference code will be released openly. The methods are intended for scientific research rather than clinical or other consequential decision-making.
Training and sampling details
Training corpora and optimization
The AF3 student is initialized from the teacher and distilled on the publicly available structures from the Protein Data Bank [3], re-featurized through the AF3 pipeline. A 640-token crop is taken before featurization. The Boltz-2 student uses the Boltz data pipeline, sampling RCSB and OpenFold-processed targets at equal probability with cluster-level sampling and the Boltz cropper. Evaluation targets are not used for gradient updates.
The eight-segment training recipe uses an optimization budget of 3,000 steps on 32 H100 GPUs. The fine teacher grid is refined from 200 to 350 edges over training; the student partition remains fixed. The log-normal weighting has , , and . The terminal anchor assigns the final teacher edge to zero noise a proposal probability of , where is the number of teacher-grid edges, and rescales the remaining probabilities.
| Parameter | Setting |
| Trainable parameters | Diffusion module only |
| Initialization | Pretrained teacher weights |
| Student segments | 8 |
| Teacher-grid refinement | 200 to 350 edges, log-linear |
| Optimization budget | 3,000 steps |
| Optimizer | Adam/AdamW, no weight decay |
| Peak learning rate | |
| Warmup | 500 steps |
| Adam coefficients | , |
| Effective batch | 128 |
| Diffusion multiplicity | 32, synchronized noise level |
| Recycle updates | 3 |
| Coordinate objective | Weighted alignment and pseudo-Huber |
| Noise weighting | Square-root Karras weighting |
| Terminal anchor | Enabled |
| Dropout | Disabled |
| AF3 relative | 0.01 (4 steps); 0.004 (8 and 16 steps) |
| AF3 absolute | 0.16 / 0.064 / 0.064 Å (4 / 8 / 16 steps) |
Table 4. Eight-segment Blitz training recipe and AF3 sampling minima.
In Algorithm 1, is a sampled adjacent teacher-grid edge, and satisfies . Edge sampling uses the EDM proposal above, not uniform sampling over segments. Edge starts coinciding with an interior student boundary are removed. The teacher destination is clipped to the lower segment boundary . TEACHERSTEP uses Heun when and the frozen teacher’s denoised prediction when . STEP applies (8) with the current student at . When the teacher reaches , its endpoint is used directly. The complete target is stopped-gradient, and neither branch evaluates a denoiser at zero noise.
For AF3, Align follows the weighted rigid-alignment procedure of AlphaFold 3 [2]. Our distillation loss uses a vector pseudo-Huber penalty and square-root Karras weight:
Here indexes atoms, , and . The reported eight-step AF3 run uses unit weights on resolved atoms, without chain-type upweighting. The vector penalty is not divided by the atom count. Losses are averaged over augmented copies and the effective batch before an Adam update; the copies of each structure share one sampled edge but have independent augmentation and Gaussian noise.
Positive nodes and deployment updates
The terminal zero is a projection endpoint, not a positive log-noise interval.
Both deployed students use deterministic frozen-denoiser exponential updates. For a positive-noise interval, let . Holding the denoiser fixed while integrating in log noise gives
We tuned the steering strengths and noise minima on a 2025 PDB validation set constructed using the same protocol as the Boltz-2 benchmark [17]. The step scale is for AF3 and for Boltz-2. At , the limiting update is the denoiser projection. This is a closed-form update for the frozen-denoiser subproblem, not an exact integration of the full neural ODE. The controlled numerical experiment instead uses Euler updates. The matched default and geometric AF3 students use the same deployed update, while the teachers retain their own deterministic or stochastic 200-step recipes.
Reference-free candidate ranking. For candidate , let , , and be bond, angle, and clash violation counts from the model-only checker, and let be the model’s highest-confidence candidate. We use
The lexicographic order breaks ties by fewer bond-plus-angle violations, then lower sample index. If no candidate has fewer total violations than , we retain . All counts use predicted coordinates and chemical identities, without a reference structure.
Datasets and evaluation conventions
Dataset differences. We use a reconstruction of the Boltz-2 benchmark ( targets) [17]. Boltz-2 is excluded from the Runs N’ Poses structural comparison because its training extends beyond that benchmark’s cutoff [19]. The RecentPDB384 results use the Protenix release [2, 7]; differing model training cutoffs also preclude treating every comparison as zero-shot.
Scoring choices. We use OpenStructure’s structural and chemical scores [5 OpenStructure developers], with stereochemical checks enabled for complex IDDT unless marked unchecked [14]. PXM is the separate Protenix-native evaluator [7]. Table 1 pairs checked and unchecked scores on matched targets. PXM is evaluated with stereochemical checks disabled. The broad tables average interface scores over the full evaluated set; the paired DockQ analysis uses reference-interface-applicable targets and penalizes missing predicted interfaces.
RF-valid is the percentage of retained predictions passing Boltz’s reference-free physical-validity checks: inter-chain clashes, ligand bond lengths and angles, internal clashes, stereochemistry, and planarity [17].
Ligand RMSD and lDDT-PLI average assigned ligands within each target, then weight targets equally. PB-valid and Success use a fixed primary ligand under the DeCAF-compatible policy [18]; Success cannot be reconstructed from the adjacent aggregate RMSD and PB-valid columns. Primary-ligand accuracy in Table 3 uses PB-evaluable targets.
Benchmark composition
Molecular composition
Table 5 separates molecular compositions on the Boltz-2 benchmark and RecentPDB384, with matched targets within each comparison. Protein and protein–ligand complexes dominate the aggregate; nucleic-acid strata expose differences that it can hide.
| Model | P | P+L | P+NA | P+NA+L |
| AlphaFold 3 teacher | 0.8510 | 0.8841 | 0.7686 | 0.7190 |
| AlphaFold 3 () | 0.8506 | 0.8835 | 0.7571 | 0.7040 |
| AlphaFold 3-Blitz [4] | 0.8447 | 0.8820 | 0.7454 | 0.6922 |
| AlphaFold 3-Blitz [8] | 0.8537 | 0.8864 | 0.7647 | 0.7134 |
| AlphaFold 3-Blitz [16] | 0.8544 | 0.8852 | 0.7696 | 0.7165 |
| Boltz-2 teacher | 0.8432 | 0.8698 | 0.7563 | 0.7066 |
| Boltz-2-Blitz [8] | 0.8405 | 0.8728 | 0.7305 | 0.6840 |
| Boltz-2-Blitz [16] | 0.8464 | 0.8746 | 0.7473 | 0.6989 |
| Boltz-1x teacher | 0.7989 | 0.8344 | 0.6099 | 0.5901 |
| DeCAF default | 0.6580 | 0.7126 | 0.4833 | 0.4259 |
| DeCAF SDE | 0.6429 | 0.7023 | 0.4685 | 0.4138 |
| DeCAF FK | 0.6525 | 0.7051 | 0.4656 | 0.4142 |
| Protenix | 0.8426 | 0.8778 | 0.7806 | 0.7248 |
| Protenix ODE | 0.5259 | 0.5888 | 0.3673 | 0.2887 |
| Protenix Mini | 0.6384 | 0.6905 | 0.5187 | 0.4865 |
Table 5. Stereochecked complex lDDT () by molecular composition in protein-containing complexes. Boltz-2 benchmark (2,300 targets).
| Model | P | P+L | P+NA | P+NA+L |
| AlphaFold 3 teacher | 0.8220 | 0.8760 | 0.7786 | 0.7933 |
| AlphaFold 3-Blitz [8] | 0.8232 | 0.8776 | 0.7844 | 0.7927 |
| AlphaFold 3-Blitz [16] | 0.8232 | 0.8766 | 0.7866 | 0.7923 |
| Boltz-1x teacher | 0.7651 | 0.8144 | 0.6560 | 0.6844 |
| DeCAF default | 0.6270 | 0.6934 | 0.5146 | 0.5598 |
| DeCAF FK | 0.6255 | 0.6855 | 0.5109 | 0.5433 |
| DeCAF SDE | 0.6249 | 0.6799 | 0.5001 | 0.5317 |
| Protenix | 0.8073 | 0.8619 | 0.7856 | 0.8060 |
| Protenix ODE | 0.4669 | 0.5590 | 0.3777 | 0.4025 |
| Protenix Mini | 0.5937 | 0.6710 | 0.5117 | 0.5381 |
Table 5 (PDF p. 15). Stereochecked complex lDDT () by molecular composition in protein-containing complexes. RecentPDB384.
P: protein; L: ligand; NA: DNA or RNA, including complexes containing both. Nucleic-acid strata are pooled with equal weight per target. Complexes without protein remain in the overall results.
Runs N’ Poses results
Both AF3 students retain structural accuracy; the sixteen-step student reduces the eight-step geometry errors but retains more than the teacher.
| Metric | AF3 Teacher | AF3 Blitz [8] | AF3 Blitz [16] | Boltz 1x | DeCAF Def. | DeCAF FK | Protenix Def. | Protenix Mini |
| cLDDT | .8320 | .8304 | .8345 | .8087 | .6914 | .6955 | .8298 | .6451 |
| bbLDDT | .8685 | .8722 | .8712 | .8790 | .8712 | .8640 | .8704 | .8538 |
| TM | .8756 | .8830 | .8777 | .8881 | .8850 | .8826 | .8781 | .8635 |
| QS | .5748 | .5755 | .5753 | .6060 | .6058 | .6055 | .5995 | .5617 |
| DockQ | .1486 | .1495 | .1484 | .1354 | .1323 | .1303 | .1302 | .1385 |
| ICS | .1258 | .1260 | .1265 | .1185 | .1182 | .1176 | .1121 | .1150 |
| IPS | .1357 | .1312 | .1341 | .1257 | .1249 | .1226 | .1248 | .1249 |
| lDDT-PLI | .6752 | .6852 | .6783 | .7231 | .7608 | .7278 | .7634 | .7337 |
| RMSD (Å) | 4.620 | 4.279 | 4.704 | 3.276 | 2.656 | 3.252 | 2.964 | 3.00 |
| Assigned < 2Å (%) | 62.14 | 63.17 | 62.48 | 70.98 | 76.44 | 70.81 | 74.49 | 72.05 |
| PB-valid (%) | 72.37 | 74.47 | 72.60 | 95.32 | 54.33 | 91.33 | 72.37 | 67.45 |
| 95% CI | [68.1, 76.6] | [70.3, 78.5] | [68.1, 76.8] | [93.2, 97.2] | [49.6, 59.0] | [88.8, 93.9] | [68.1, 76.6] | [63.0, 71.7] |
| Primary < 2Å (%) | 83.84 | 83.37 | 83.37 | 78.92 | 79.86 | 80.33 | 81.97 | 77.05 |
| 95% CI | [80.3, 87.1] | [79.9, 86.9] | [79.9, 86.9] | [74.9, 82.7] | [76.1, 83.6] | [76.6, 84.1] | [78.2, 85.5] | [73.1, 81.0] |
| Success (%) | 61.36 | 64.17 | 62.06 | 75.88 | 47.54 | 73.30 | 60.42 | 54.80 |
| 95% CI | [56.7, 66.0] | [59.5, 68.6] | [57.4, 66.7] | [71.9, 79.9] | [42.9, 52.2] | [69.1, 77.5] | [56.0, 65.1] | [50.1, 59.5] |
| RF-valid (%) | 13.37 | 11.94 | 13.07 | 47.47 | 4.25 | 26.44 | 11.09 | 5.45 |
| Bonds | 0.01 | 4.04 | 0.68 | 9.96 | 317.15 | 326.62 | 0.13 | 452.59 |
| Angles | 0.01 | 0.65 | 0.11 | 3.90 | 142.02 | 155.58 | 0.26 | 149.52 |
| Clashes | 0.36 | 19.84 | 2.74 | 54.03 | 415.79 | 336.86 | 4.30 | 308.85 |
| Violations | 0.38 | 24.53 | 3.54 | 67.89 | 874.96 | 819.05 | 4.70 | 910.97 |
| OST-clean (%) | 91.73 | 85.55 | 88.39 | 3.09 | 0.00 | 0.00 | 50.38 | 0.00 |
Table 6. Runs N’ Poses: structural accuracy, ligand placement, and geometry.
PB-valid, primary accuracy, and Success all use the same 427 targets (17% of the eligible set). Ligand-assignment and scoring failures restrict this subset, so these rates are not full-benchmark estimates. Brackets give 95% bootstrap intervals. DeCAF FK uses four particles and twelve steps.
RecentPDB384 results
The sixteen-step student reduces the eight-step geometry errors while retaining similar structural accuracy. We exclude Boltz-2 and its students because this benchmark predates its June 2023 training cutoff [7, 17].
| Metric | AF3 Teacher | AF3 Blitz [8] | AF3 Blitz [16] | Boltz 1x | DeCAF Def. | DeCAF FK | DeCAF SDE | Protenix Def. | Protenix ODE | Protenix Mini |
| cLDDT | .8286 | .8273 | .8286 | .7296 | .6051 | .5993 | .5954 | .8231 | .4680 | .5979 |
| bbLDDT | .8948 | .8961 | .8958 | .8673 | .8637 | .8573 | .8573 | .8866 | .8822 | .8683 |
| TM | .8572 | .8590 | .8567 | .8287 | .8240 | .8191 | .8226 | .8500 | .8486 | .8277 |
| QS | .7482 | .7502 | .7466 | .6986 | .6922 | .6959 | .6891 | .7257 | .7167 | .6901 |
| Metric | AF3 Teacher | AF3 Blitz [8] | AF3 Blitz [16] | Boltz 1x | Boltz Def. | DeCAF FK | DeCAF SDE | DeCAF Def. | Protenix ODE | Protenix Mini |
| DockQ | .3428 | .3431 | .3395 | .3036 | .2969 | .2936 | .2957 | .3258 | .3275 | .2986 |
| ICS | .3704 | .3715 | .3692 | .3207 | .3163 | .3182 | .3137 | .3430 | .3371 | .3220 |
| IPS | .4138 | .4145 | .4147 | .3763 | .3760 | .3724 | .3723 | .3962 | .3930 | .3817 |
| lDDT-PLI | .5363 | .5342 | .5364 | .4948 | .5039 | .4977 | .4898 | .5257 | .4844 | .5055 |
| RMSD (Å) | 9.147 | 8.480 | 8.473 | 9.159 | 7.982 | 8.405 | 8.751 | 8.086 | 7.967 | 8.08 |
| Assigned < 2Å (%) | 49.28 | 47.55 | 47.34 | 44.72 | 44.71 | 42.84 | 42.16 | 48.10 | 42.79 | 42.90 |
| PB-valid (%) | 51.14 | 47.73 | 45.45 | 77.27 | 22.73 | 71.59 | 19.32 | 55.68 | 28.41 | 43.18 |
| Primary < 2Å (%) | 61.36 | 59.09 | 60.23 | 57.95 | 56.82 | 56.82 | 56.82 | 63.64 | 60.23 | 57.95 |
| Success (%) | 30.68 | 32.95 | 29.55 | 46.59 | 17.05 | 45.45 | 14.77 | 39.77 | 28.41 | 37.50 |
| RF-valid (%) | 40.36 | 39.32 | 37.50 | 92.45 | 25.78 | 61.98 | 25.78 | 34.11 | 28.91 | 32.55 |
| Bonds | 0.52 | 2.29 | 0.80 | 15.76 | 254.24 | 263.73 | 256.59 | 0.62 | 729.90 | 382.14 |
| Angles | 0.47 | 1.10 | 0.54 | 8.47 | 119.82 | 132.35 | 128.16 | 0.67 | 326.12 | 136.47 |
| Clashes | 1.61 | 4.35 | 2.80 | 19.96 | 254.65 | 192.72 | 253.66 | 5.27 | 1001.85 | 195.19 |
| Violations | 2.61 | 7.74 | 4.14 | 44.18 | 628.71 | 588.81 | 638.41 | 6.55 | 2057.87 | 713.80 |
| OST-clean (%) | 68.53 | 67.20 | 70.93 | 2.67 | 0.00 | 0.00 | 0.00 | 38.93 | 0.00 | 0.00 |
| PXM | .8260 | .8286 | .8278 | .7884 | .7872 | .7804 | .7808 | .8206 | .8192 | .7942 |
Table 7. RecentPDB384: structural accuracy, ligand placement, and geometry.
Size and tail sensitivity in the matched AF3 comparison
A small subset of targets contributes much of the remaining clash burden. For the sixteen-step AF3 student, 8oq5 alone accounts for 61.6% of counted clashes on the matched comparison (Figure 6). We retain this target in every primary result. The influence analysis shows why mean counts should be read alongside defect incidence and the full distribution, rather than used as a description of a typical prediction.
Figure 6. A small subset carries much of the clash burden. Left: cumulative clashes ranked by target. Middle: counts for 8oq5. Right: means with all targets and after omitting 8oq5 as an influence diagnostic. Every primary result retains this target.
Geometry and candidate selection
We compare the canonical stochastic teacher with the four-, eight-, and sixteen-step students on matched Boltz-2 benchmark targets. The eight-step schedule comparison is shown in Figure 5.
How often a defect occurs and how severe it becomes
A small number of heavily damaged predictions can dominate a mean count. Figure 7 separates the fraction of affected targets from the mean number of violations among those targets. The sixteen-step student has
Figure 7. Severity and incidence answer different questions. Top: targets with a counted defect. Bottom: mean count among affected targets.
fewer defect-affected predictions than the teacher in this comparison, but its remaining defects are more concentrated. A low incidence therefore does not imply low severity when a failure occurs. These checks cover polymer bonds, angles, and clashes; passing them does not establish full chemical validity or replace ligand PoseBusters checks.
Distribution tails and target-size sensitivity
Figure 8 and Figure 9 show how much of the mean comes from the tail and whether the pattern persists across target sizes. Each target has equal weight within its size bin.
Figure 8. The full violation distribution. Curves show the fraction of matched targets above each count. The value at zero is incidence; no extreme target is removed.
Figure 9. Geometry across target sizes. Mean total count and defect incidence within predefined reference-residue bins, using identical targets for all four methods.
Effect of candidate selection
Table 8 uses each model’s native confidence ranking. Structural accuracy remains competitive, and some ligand and interface scores improve relative to our reranker, but geometry violations are more frequent. On identical targets, reranking reduces the affected fraction from 13.9% to 6.9% for AF3-Blitz [16] and from 42.6% to 26.9% for Boltz-2-Blitz [16]. Checked lDDT also improves in these paired comparisons. We therefore retain the reranker: the lower defect incidence than the teacher reported for the sixteen-step students depends on this selection rule.
| Metric | AlphaFold3 | Boltz-2 | ||||||
| Teacher | Blitz [4] | Blitz [8] | Blitz [16] | Teacher | Blitz [8] | Blitz [16] | ||
| NFE | 200 | 8 | 4 | 8 | 16 | 600 | 8 | 16 |
| cLDDT | 0.8615 | 0.8550 | 0.8512 | 0.8615 | 0.8624 | 0.8503 | 0.8476 | 0.8525 |
| bbLDDT | 0.9136 | 0.9147 | 0.9128 | 0.9144 | 0.9150 | 0.9047 | 0.9061 | 0.9074 |
| TM | 0.8898 | 0.8902 | 0.8898 | 0.8901 | 0.8909 | 0.8758 | 0.8789 | 0.8792 |
| DockQ | 0.4198 | 0.4163 | 0.4220 | 0.4216 | 0.4168 | 0.3927 | 0.3929 | 0.3948 |
| QS | 0.7784 | 0.7763 | 0.7793 | 0.7767 | 0.7782 | 0.7378 | 0.7381 | 0.7381 |
| ICS | 0.4263 | 0.4239 | 0.4274 | 0.4258 | 0.4245 | 0.3883 | 0.3880 | 0.3878 |
| IPS | 0.4560 | 0.4567 | 0.4578 | 0.4572 | 0.4581 | 0.4223 | 0.4256 | 0.4252 |
| lDDT-PLI | 0.6437 | 0.6435 | 0.6386 | 0.6448 | 0.6471 | 0.5596 | 0.5572 | 0.5598 |
| RMSD () | 6.63 | 6.51 | 6.43 | 6.43 | 6.38 | 9.00 | 8.59 | 8.67 |
| Assigned < 2 (%) | 61.3 | 61.4 | 60.8 | 61.6 | 61.8 | 52.2 | 52.0 | 52.7 |
| PB-valid (%) | 45.1 | 38.4 | 35.3 | 42.9 | 45.7 | 60.9 | 53.1 | 55.5 |
| Primary < 2 (%) | 70.8 | 71.4 | 71.2 | 71.8 | 71.9 | 61.5 | 61.2 | 61.9 |
| Success (%) | 34.7 | 32.5 | 31.2 | 35.7 | 36.2 | 36.5 | 31.1 | 34.6 |
| RF-valid (%) | 41.5 | 37.4 | 36.7 | 39.5 | 41.9 | 98.5 | 93.6 | 93.7 |
| Bonds | 0.06 | 6.46 | 23.36 | 2.21 | 0.21 | 0.82 | 7.45 | 2.58 |
| Angles | 0.09 | 3.04 | 8.58 | 0.67 | 0.17 | 0.61 | 4.81 | 1.86 |
| Clashes | 0.54 | 7.87 | 49.80 | 7.71 | 2.89 | 2.52 | 9.79 | 4.64 |
| Any violation (%) | 9.8 | 80.2 | 83.9 | 27.9 | 13.8 | 30.2 | 62.4 | 42.5 |
Table 8. Native-confidence selection on the Boltz-2 benchmark.
Geometry counts are per target. Comparisons use matched evaluable targets within each family; ligand subsets can differ from Table 2.
Uncertainty estimates
Table 9 reports pointwise 95% intervals from 10,000 PDB-cluster bootstrap draws, measuring target-sampling uncertainty rather than variation across training runs.
| Metric | Teacher | Blitz [4] | Blitz [8] | Blitz [16] | |
| cLDDT | [0.857, 0.866] | [0.855, 0.865] | [0.850, 0.861] | [0.859, 0.868] | [0.859, 0.868] |
| bbLDDT | [0.909, 0.918] | [0.911, 0.919] | [0.908, 0.917] | [0.910, 0.919] | [0.911, 0.919] |
| TM | [0.882, 0.897] | [0.881, 0.896] | [0.881, 0.896] | [0.881, 0.896] | [0.882, 0.897] |
| DockQ | [0.403, 0.437] | [0.397, 0.430] | [0.402, 0.435] | [0.404, 0.438] | [0.400, 0.433] |
| QS | [0.765, 0.791] | [0.763, 0.788] | [0.767, 0.794] | [0.763, 0.790] | [0.765, 0.791] |
| ICS | [0.410, 0.443] | [0.405, 0.438] | [0.412, 0.446] | [0.410, 0.443] | [0.408, 0.441] |
| IPS | [0.439, 0.472] | [0.438, 0.470] | [0.441, 0.474] | [0.441, 0.474] | [0.440, 0.473] |
| lDDT-PLI | [0.625, 0.662] | [0.625, 0.662] | [0.619, 0.656] | [0.624, 0.661] | [0.625, 0.662] |
| RMSD (Å) | [6.06, 7.21] | [5.97, 7.10] | [5.90, 7.00] | [5.90, 7.00] | [5.96, 7.08] |
| Assigned < 2Å | [58.9, 63.7] | [58.9, 63.8] | [58.6, 63.4] | [58.9, 63.7] | [58.9, 63.8] |
| PB-valid | [42.0, 48.0] | [32.3, 38.1] | [31.8, 37.5] | [40.0, 46.1] | [42.5, 48.7] |
| Primary < 2Å | [65.2, 70.6] | [64.6, 70.0] | [64.8, 70.2] | [65.2, 70.7] | [65.6, 71.0] |
| Success | [31.8, 37.6] | [26.7, 32.4] | [27.5, 33.2] | [32.8, 38.7] | [32.9, 38.8] |
| RF-valid | [39.4, 43.6] | [34.4, 38.3] | [34.9, 39.0] | [38.5, 42.5] | [40.3, 44.3] |
| Bonds | [0.05, 0.07] | [2.39, 3.39] | [7.75, 19.76] | [0.29, 1.57] | [0.08, 0.13] |
| Angles | [0.07, 0.11] | [0.89, 1.16] | [2.53, 6.30] | [0.15, 0.34] | [0.06, 0.10] |
| Clashes | [0.41, 0.69] | [1.97, 7.55] | [11.02, 44.46] | [1.61, 7.27] | [0.49, 3.69] |
| Any violation | [8.6, 11.1] | [48.5, 52.7] | [55.4, 59.5] | [10.0, 12.7] | [5.8, 7.9] |
Table 9. Uncertainty in mean scores and paired comparisons. A. Mean uncertainty: AF3.
| Metric | Teacher | Blitz [8] | Blitz [16] |
| cLDDT | [0.845, 0.855] | [0.844, 0.854] | [0.848, 0.858] |
| bbLDDT | [0.900, 0.909] | [0.901, 0.910] | [0.902, 0.911] |
| TM | [0.867, 0.883] | [0.869, 0.885] | [0.870, 0.885] |
| DockQ | [0.376, 0.409] | [0.373, 0.406] | [0.377, 0.410] |
| QS | [0.724, 0.751] | [0.721, 0.749] | [0.722, 0.749] |
| ICS | [0.372, 0.404] | [0.369, 0.401] | [0.370, 0.402] |
| IPS | [0.406, 0.438] | [0.406, 0.438] | [0.407, 0.438] |
| lDDT-PLI | [0.541, 0.578] | [0.536, 0.573] | [0.538, 0.576] |
| RMSD (Å) | [8.32, 9.70] | [8.11, 9.39] | [8.14, 9.41] |
| Assigned < 2Å | [49.7, 54.6] | [49.2, 54.1] | [49.7, 54.6] |
| PB-valid | [58.0, 63.8] | [49.0, 55.1] | [52.9, 59.0] |
| Primary < 2Å | [58.0, 63.8] | [57.3, 63.2] | [57.8, 63.6] |
| Success | [33.6, 39.5] | [27.8, 33.4] | [31.1, 36.8] |
| RF-valid | [98.0, 99.0] | [92.6, 94.6] | [93.2, 95.1] |
| Bonds | [0.74, 1.11] | [3.28, 7.02] | [1.18, 2.67] |
| Angles | [0.54, 0.97] | [1.96, 4.61] | [0.77, 1.84] |
| Clashes | [2.03, 3.77] | [4.58, 8.48] | [2.58, 4.84] |
| Any violation | [28.4, 32.2] | [39.8, 44.0] | [24.9, 28.6] |
Table 9. Uncertainty in mean scores and paired comparisons. B. Mean uncertainty: Boltz-2.
Intervals use 10,000 PDB-cluster bootstrap draws. Rates are percentages; geometry counts are per target.
| Metric | Boltz-1x | DeCAF | DeCAF SDE | DeCAF FK | Protenix | Protenix ODE | Protenix Mini |
| cLDDT | [0.799, 0.811] | [0.666, 0.682] | [0.653, 0.668] | [0.658, 0.673] | [0.848, 0.859] | [0.538, 0.554] | [0.651, 0.663] |
| bbLDDT | [0.888, 0.897] | [0.882, 0.891] | [0.874, 0.883] | [0.874, 0.884] | [0.902, 0.912] | [0.899, 0.909] | [0.889, 0.899] |
| TM | [0.863, 0.878] | [0.854, 0.870] | [0.850, 0.866] | [0.849, 0.866] | [0.873, 0.888] | [0.870, 0.886] | [0.860, 0.876] |
| DockQ | [0.372, 0.403] | [0.355, 0.385] | [0.348, 0.377] | [0.349, 0.379] | [0.391, 0.424] | [0.387, 0.419] | [0.353, 0.384] |
| QS | [0.719, 0.746] | [0.710, 0.738] | [0.705, 0.733] | [0.712, 0.739] | [0.743, 0.769] | [0.735, 0.762] | [0.704, 0.732] |
| ICS | [0.369, 0.400] | [0.359, 0.389] | [0.355, 0.385] | [0.360, 0.391] | [0.391, 0.424] | [0.384, 0.416] | [0.353, 0.384] |
| IPS | [0.407, 0.438] | [0.402, 0.433] | [0.399, 0.429] | [0.402, 0.433] | [0.424, 0.455] | [0.424, 0.456] | [0.404, 0.436] |
| lDDT-PLI | [0.575, 0.612] | [0.569, 0.606] | [0.563, 0.599] | [0.569, 0.606] | [0.606, 0.643] | [0.558, 0.598] | [0.559, 0.597] |
| RMSD (Å) | [6.47, 7.56] | [6.44, 7.53] | [6.58, 7.66] | [6.50, 7.58] | [6.44, 7.56] | [6.19, 7.19] | [6.88, 8.02] |
| Assigned < 2Å | [54.9, 59.8] | [53.5, 58.5] | [53.7, 58.5] | [54.1, 59.0] | [57.0, 61.9] | [52.4, 57.4] | [51.9, 56.8] |
| PB-valid | [55.2, 61.2] | [22.1, 27.3] | [24.1, 29.4] | [51.0, 56.9] | [42.8, 48.8] | [24.0, 29.4] | [25.5, 31.0] |
| Primary < 2Å | [63.3, 69.2] | [62.8, 68.6] | [62.9, 68.7] | [62.6, 68.5] | [66.5, 72.2] | [62.8, 68.6] | [60.9, 66.6] |
| Success | [36.4, 42.3] | [18.6, 23.6] | [19.2, 24.3] | [35.4, 41.3] | [31.2, 36.9] | [22.4, 27.6] | [21.5, 26.6] |
| RF-valid | [98.6, 99.4] | [24.2, 27.7] | [25.7, 29.4] | [64.9, 68.8] | [30.8, 34.7] | [29.0, 32.8] | [29.5, 33.4] |
| Bonds | [9.47, 14.45] | [252.86, 284.68] | [256.20, 287.62] | [263.42, 297.15] | [0.16, 0.21] | [637.36, 696.26] | [357.66, 390.61] |
| Angles | [4.10, 7.30] | [115.68, 132.17] | [123.16, 140.39] | [129.07, 147.97] | [0.50, 0.67] | [269.23, 299.08] | [123.68, 134.82] |
| Clashes | [16.22, 21.68] | [258.64, 319.85] | [250.01, 304.12] | [206.79, 251.70] | [2.77, 5.04] | [544.28, 742.67] | [164.69, 197.33] |
| Any violation | [94.1, 95.9] | [96.3, 97.7] | [96.8, 98.1] | [97.0, 98.3] | [47.0, 51.2] | [100.0, 100.0] | [99.9, 100.0] |
Table 9 (PDF p. 21). Uncertainty in mean scores and paired comparisons. C. Mean uncertainty: other methods.
Intervals use 10,000 PDB-cluster bootstrap draws. Rates are percentages; geometry counts are per target.
| Metric | AF3 [4] | AF3 [8] | AF3 [16] | Boltz-2 [8] | Boltz-2 [16] |
| cLDDT | -.0056 [-.0071, -.0042] | +.0020 [.0012, .0028] | +.0019 [.0013, .0026] | -.0007 [-.0020, .0005] | +.0034 [.0025, .0044] |
| bbLDDT | -.0008 [-.0017, .0002] | +.0014 [.0007, .0021] | +.0015 [.0009, .0021] | +.0012 [.0001, .0022] | +.0025 [.0016, .0034] |
| TM | -.0009 [-.0026, .0007] | -.0010 [-.0027, .0005] | -.0004 [-.0020, .0011] | +.0016 [-.0003, .0034] | +.0024 [.0006, .0042] |
| DockQ | -.0013 [-.0037, .0012] | +.0009 [-.0017, .0033] | -.0031 [-.0057, -.0006] | -.0031 [-.0060, -.0002] | +.0003 [-.0024, .0030] |
| QS | +.0021 [-.0003, .0045] | -.0016 [-.0041, .0009] | -.0000 [-.0024, .0023] | -.0027 [-.0059, .0003] | -.0019 [-.0049, .0010] |
| Metric | AF3 [4] | AF3 [8] | AF3 [16] | Boltz-2 [8] | Boltz-2 [16] |
| ICS ↑ | +.0026 [.0003, .0048] | .0000 [-.0023, .0023] | -.0016 [-.0037, .0006] | -.0027 [-.0056, -.0001] | -.0020 [-.0046, .0005] |
| IPS ↑ | +.0014 [-.0009, .0037] | +.0014 [-.0008, .0036] | +.0010 [-.0011, .0031] | -.0001 [-.0025, .0022] | +.0004 [-.0019, .0027] |
| lDDT-PLI ↑ | -.0065 [-.0120, -.0013] | -.0016 [-.0064, .0030] | .0000 [-.0054, .0050] | -.0051 [-.0098, -.0003] | -.0023 [-.0069, .0022] |
| RMSD (Å) ↓ | -0.19 [-0.42, 0.04] | -0.17 [-0.39, 0.05] | -0.12 [-0.34, 0.10] | -0.26 [-0.55, 0.04] | -0.22 [-0.52, 0.07] |
| Assigned < 2Å ↑ | -0.3 [-1.4, 0.9] | +0.0 [-1.0, 1.0] | +0.1 [-1.0, 1.1] | -0.5 [-1.3, 0.3] | +0.1 [-0.6, 0.7] |
| PB-valid ↑ | -10.4 [-13.0, -7.7] | -2.0 [-4.4, 0.2] | +0.8 [-1.3, 2.9] | -8.9 [-10.9, -6.9] | -5.0 [-6.8, -3.4] |
| Primary < 2Å ↑ | -0.4 [-1.9, 1.1] | +0.1 [-1.2, 1.4] | +0.3 [-0.9, 1.7] | -0.6 [-1.9, 0.8] | -0.2 [-1.4, 1.0] |
| Success ↑ | -4.3 [-6.4, -2.3] | +1.1 [-0.6, 2.7] | +1.2 [-0.3, 2.6] | -5.9 [-7.6, -4.4] | -2.6 [-3.9, -1.5] |
| RF-valid ↑ | -4.6 [-6.1, -3.1] | -1.0 [-2.4, 0.5] | +0.8 [-0.7, 2.3] | -4.8 [-5.9, -3.9] | -4.4 [-5.3, -3.4] |
| Bonds ↓ | +12.79 [7.69, 19.71] | +0.73 [0.23, 1.50] | +0.04 [0.02, 0.06] | +3.94 [2.46, 6.10] | +0.88 [0.36, 1.69] |
| Angles ↓ | +3.92 [2.45, 6.22] | +0.15 [0.07, 0.25] | -0.01 [-0.02, 0.01] | +2.34 [1.32, 3.86] | +0.52 [0.17, 1.02] |
| Clashes ↓ | +23.52 [10.49, 43.90] | +3.25 [1.07, 6.65] | +1.06 [-0.02, 3.11] | +3.48 [2.17, 5.25] | +0.81 [0.26, 1.53] |
| Any violation ↓ | +47.6 [45.5, 49.7] | +1.5 [0.4, 2.6] | -3.0 [-4.0, -2.1] | +11.6 [9.7, 13.5] | -3.6 [-5.3, -1.9] |
Table 9 (PDF p. 21, second). Uncertainty in mean scores and paired comparisons. D. Student minus teacher.
Each comparison uses its own metric-specific teacher–student intersection and joint PDB-cluster resampling. These supports can differ from the multi-model tables. Differences in rates are percentage points. Intervals are pointwise and do not measure variation across training runs.
Numerical analysis of the molecular sampler
We test whether the schedule improvement survives changes in minimum noise, then examine how individual sampling intervals affect the final structure. The schedule sweeps use the Boltz-2 benchmark; the interval diagnostics use selected targets with paired candidates.
Controlled sampler and minimum-noise sensitivity
The AF3 field centers valid atomic coordinates using , where subtracts the valid-atom centroid. Conditioning, masks, initialization, and the coordinate frame are fixed. These deterministic teacher experiments use scale-one Euler updates without churn, random frame augmentation, or steering.
The budget and minimum-noise sweeps cover 44 configurations with five candidates per target on the Boltz-2 benchmark ( targets). Accuracy uses the complete-case intersection described in Appendix B. We average candidates within each target, then weight targets equally. All reported lDDT includes stereochemical checks.
Changing the minimum noise alone does not reproduce the geometric schedule’s combination of accuracy and geometry (Figure 10). The schedule advantage also depends on the step budget, as shown in Figure 2.
Figure 10. Noise grids and minimum-noise sensitivity. A–B: eight positive nodes and seven relative intervals, with maximum noise 2,560 Å and minimum 0.064 Å; terminal zero is separate. C–F: accuracy and geometry at eight evaluations as the minimum noise changes. Bands are target-bootstrap intervals. Bond and angle rates are per 100 parameterized residues; clashes are per 100 atoms.
Same-input local defects
Each selected diagnostic target contributes two paired candidates. We evaluate intervals starting from both coarse and refined trajectories. From each starting state, Euler, Heun, and the refined calculation use the same input and coordinate frame. Local defects use component RMSE, for valid atoms; endpoint displacement uses unaligned atom RMSD, .
The reference-refinement differences are much smaller than the single-step defects for both starting-state constructions (Figure 11). This checks sensitivity to the numerical reference without treating it as an exact solution.
Figure 11. Local defects and reference sensitivity. Top: refined starting states. Bottom: coarse starting states. Columns show single-step Euler and Heun defects against a 128-subdivision reference, followed by the 64-to-128-subdivision reference difference. Curves average candidates within targets; bands resample targets. The horizontal axis gives interval midpoints in noise.
Single-interval interventions and final geometry
For Equation (6), the control follows the all-refined path. The branch shares its prefix, substitutes one coarse Euler step, and then follows the same refined continuation. We score the final structures.
For the default grid, the last positive interval causes the strongest mean loss of stereochemically checked lDDT, although an earlier interval produces a larger endpoint displacement (Figure 12). Bond and angle violations also rise at low noise. The geometric grid changes this damage profile, showing why coordinate displacement alone is insufficient to judge the quality of an update. The grids cover different noise windows, so equal interval indices are not directly matched. Nor can the individual effects be added: an earlier change would alter the starting states of later intervals.
Figure 12. How one coarse interval affects the final structure. Solid and dashed curves use 200 and 100 subdivisions in the refined calculation. A and C–E show branch minus control; B shows unaligned endpoint atom RMSD. Negative accuracy changes and positive violation changes are worse. Bond and angle rates are per 100 parameterized residues; clashes are per 100 atoms. Comparisons pair targets and candidates; bands resample targets.
References
- [2]Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf Ronneberger, Lindsay Willmore, Andrew J. Ballard, Joshua Bambrick, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature, 630:493–500, 2024. doi: 10.1038/s41586-024-07487-w.DOI
- [3]Helen M. Berman, John Westbrook, Zukang Feng, Gary Gilliland, T. N. Bhat, Helge Weissig, Ilya N. Shindyalov, and Philip E. Bourne. The protein data bank. Nucleic Acids Research, 28(1):235–242, 2000. doi: 10.1093/nar/28.1.235.DOI
- [4]David Berthelot, Arnaud Autef, Jierui Lin, Dian Ang Yap, Shuangfei Zhai, Siyuan Hu, Daniel Zheng, Walter Talbott, and Eric Gu. TRACT: Denoising diffusion models with transitive closure time-distillation. arXiv preprint arXiv:2303.04248, 2023.arxiv.org/abs/2303.04248
- [5]Marco Biasini, Tobias Schmidt, Stefan Bienert, Valerio Mariani, Gabriel Studer, Jürgen Haas, Niklaus Johner, Andreas Daniel Schenk, Ansgar Philippisen, and Torsten Schwede. OpenStructure: An integrated software framework for computational structural biology. Acta Crystallographica Section D: Biological Crystallography, 69(5):701–709, 2013. doi: 10.1107/S0907444913007051.DOI
- [6]Martin Buttenschoen, Garrett M. Morris, and Charlotte M. Deane. PoseBusters: AI-based docking methods fail to generate physically valid poses or generalise to novel sequences. Chemical Science, 15:3130–3139, 2024. doi: 10.1039/D3SC04185A.DOI
- [7]ByteDance AML AI4Science Team, Xinshi Chen, Yuxuan Zhang, Chan Lu, Wenzhi Ma, Jiaqi Guan, Chengyue Gong, Jincai Yang, Hanyu Zhang, Ke Zhang, Shenghao Wu, Kuangqi Zhou, Yanping Yang, Zhenyu Liu, Lan Wang, Bo Shi, Shaochen Shi, and Wenzhi Xiao. Protenix: Advancing structure prediction through a comprehensive AlphaFold 3 reproduction. bioRxiv, 2025. doi: 10.1101/2025.01.08.631967.
- [8]Chai Discovery, Jacques Boitreaud, Jack Dent, Matthew McPartlon, Joshua Meier, Vinicius Reis, Alex Rogozhnikov, and Kevin Wu. Chai-1: Decoding the molecular interactions of life. bioRxiv, 2024. doi: 10.1101/2024.10.10.615955.
- [9]A. M. Collette, S. A. Hassan, S. I. Schmidt, A. J. Lara, W. Yang, and N. L. Samara. An unusual dual sugar-binding lectin domain controls the substrate specificity of a mucin-type O-glycosyltransferase. Science Advances, 10: eadj8829, 2024. doi: 10.1126/sciadv.adj8829.DOI
- [10]Chengyue Gong, Xinshi Chen, Yuxuan Zhang, Yuxuan Song, Hao Zhou, and Wenzhi Xiao. Protenix-Mini: Efficient structure predictor via compact architecture, few-step diffusion and switchable pLM. arXiv preprint arXiv:2507.11839, 2025. doi: 10.48550/arXiv.2507.11839.DOI
- [11]Jonathan Heek, Emiel Hoogeboom, and Tim Salimans. Multistep consistency models. arXiv preprint arXiv:2403.06807, 2024.arxiv.org/abs/2403.06807
- [12]Felix Kallenborn, Alejandro Chacon, Christian Hundt, Hassan Sirelkhatim, Kieran Didi, Sooyoung Cha, Christian Dallago, Milot Mirdita, Bertil Schmidt, and Martin Steinegger. GPU-accelerated homology search with MMseqs2. Nature Methods, 22(10):2024–2027, 2025. doi: 10.1038/s41592-025-02819-8.DOI
- [13]Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, volume 35, 2022.arxiv.org/abs/2206.00364
- [14]Valerio Mariani, Marco Biasini, Alessandro Barbato, and Torsten Schwede. IDDT: A local superposition-free score for comparing protein structures and models using distance difference tests. Bioinformatics, 29(21):2722–2728, 2013. doi: 10.1093/bioinformatics/btt473.DOI
- [17]Saro Passaro, Gabriele Corso, Jeremy Wohlwend, Mateo Reveiz, Stephan Thaler, Vignesh Ram Somnath, Noah Getz, Tally Portnoi, Julien Roy, Hannes Stark, David Kwabi-Addo, Dominique Beaini, Tommi Jaakkola, and Regina Barzilay. Boltz-2: Towards accurate and efficient binding affinity prediction. bioRxiv, 2025. doi: 10.1101/2025.06.14.659707.DOI
- [18]Gianluca Scarpellini, Ron Shprints, Peter Holderrieth, Juno Nam, Pranav Murugan, Rafael Gómez-Bombarelli, Tommi Jaakkola, Maruan Al-Shedivat, Nicholas Matthew Boffi, and Avishek Joey Bose. Few-step cofolding with all-atom flow maps. arXiv preprint arXiv:2606.08375, 2026. doi: 10.48550/arXiv.2606.08375.DOI
- [19]Peter Škrinjar, Jérôme Eberhardt, Gerardo Tauriello, Torsten Schwede, and Janani Dura iraj. Have protein-ligand cofolding methods moved beyond memorisation? bioRxiv, 2025. doi: 10.1101/2025.02.03.636309. Benchmark and code at https://github.com/plinder-org/runs-n-poses.DOI
- [20]Yang Song and Prafulla Dhariwal. Improved techniques for training consistency models. In International Conference on Learning Representations, 2024.arxiv.org/abs/2310.14189
- [21]Yang Song, Prafulla Dhariwal, Mark Chen, and Ilya Sutskever. Consistency models. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pp. 32211–32252. PMLR, 2023.arxiv.org/abs/2303.01469
- [24]Zhe Zhang, Yuanning Feng, Yuxuan Song, Keyue Qiu, Hao Zhou, and Wei-Ying Ma. DCFold: Efficient protein structure generation with single forward pass. arXiv preprint arXiv:2605.17899, 2026. doi: 10.48550/arXiv.2605.17899.DOI











