betaFeedback

WorksComputer Science

Biology-Guided Context-Aware Residual Propagation for Spatial Gene Expression Prediction from Histology

Abstract

Spatial transcriptomics provides spatially resolved molecular profiling, but its cost, limited throughput and experimental complexity hinder routine large-scale use. Predicting spatial gene expression from hematoxylin-and-eosin-stained histology images offers a scalable alternative, yet remains challenging because histology provides only an indirect morphological view of molecular activity. Existing methods typically decode expression from visual features in a single stage, even when enhanced with spatial aggregation, retrieval or multi-resolution fusion. This formulation is restrictive because spot-level expression reflects not only local morphology but also cellular identity, neighborhood composition and tissue-level organization. We propose BCARP, a Biology-guided Context-Aware Residual Propagation framework for histology-guided spatial transcriptomics prediction. Instead of directly predicting gene expression from histology embeddings, BCARP treats them as morphology visual priors and progressively refines spot representations in a transcriptomics-oriented latent space. Each refinement step integrates nuclei-guided local cellular context with prototype-guided global context propagation, enabling information exchange across both neighboring and spatially distant spots. A pattern-scale decoder further separates spatial expression into pattern prediction from raw expression with scale calibration. Evaluated on three ST datasets and ten HEST-1k cohorts, BCARP achieves the highest mean correlation-based performance against other baselines. Numerical results support biology-guided local-global residual propagation as an effective alternative to visual decoding for histology-to-transcriptomics prediction.

Introduction

Spatial transcriptomics (ST) enables spatially resolved molecular profiling by measuring gene expression while preserving tissue coordinates [18, 17]. Unlike bulk RNA sequencing, which averages signals over tissue regions, ST quantifies mRNA expression across spatial “spots” and reveals how expression patterns are organized within histological architecture. This perspective is critical for biomedical studies, where disease progression, immune response, tumor microenvironment and cell-cell interactions are often spatially structured [16], and it bridges conventional histopathology and transcriptomic profiling [15].

Despite its value, ST remains costly, experimentally demanding and limited in coverage and throughput [14], whereas hematoxylin-and-eosin (H&E)-stained whole-slide images (WSIs) are routinely collected in clinical workflows [13]. This gap motivates histology-guided ST prediction, which estimates spatial gene expression directly from H&E images. Beyond patch-based regression such as ST-Net [1], recent methods enrich the visual representation in complementary ways: graph-based models such as MERGE [28] connect spots through spatial and morphological relations; multi-resolution models such as TRIPLEX [23] and M2OST [32] fuse target-spot, neighborhood and slide-level views; mclSTExp [33] performs contrastive cross-modal alignment for retrieval-based inference; EGN [21] refines targets with globally similar exemplars; and STFlow [31] models the whole-slide expression distribution with flow matching.

Despite this diversity, these approaches largely share a one-stage visual decoding paradigm: contextual information is aggregated into a histology representation, which is then mapped to expression in a single prediction step. This implicitly assumes that the extracted representation is already aligned with the transcriptomic factors underlying spot-level expression. In practice, expression is not determined solely by the appearance of the central patch; it is also shaped by local cellular composition, neighborhood microenvironment and broader tissue organization [22, 30, 17]. Histology-guided prediction should therefore go beyond direct visual regression and progressively refine morphology-derived features into transcriptomics-oriented representations.

To this end, we propose BCARP, a Biology-guided Context-Aware Residual Propagation framework. BCARP projects frozen pathology foundation embeddings into a task-oriented latent space and treats them as histology anchors. Instead of decoding these anchors in one step, it iteratively refines them through a local-global residual propagation loop: local cellular context encoders use nuclei-derived supervision to inject cell-type and neighborhood-composition information, while a prototype-guided global module combines learnable global priors with spot-level self-attention to model long-range tissue organization. A pattern-scale decoder then separates spatial expression-pattern estimation from raw-expression scale calibration. We evaluate BCARP on HER2ST [22], Skin [30] and ST-Net [1] under a unified spot-level protocol, and on ten HEST-1k cohorts [2] following the STFlow [31] evaluation suite, spanning multiple organs and cancer types. BCARP outperforms representative baselines, reduces poorly predicted genes, and is further validated by component and refinement-depth ablations. Our contributions are summarized as follows:

  • We propose BCARP, a residual propagation framework for multi-step context-aware refinement in histology-guided spatial transcriptomics prediction.

  • We design a local-global refinement mechanism that combines nuclei-derived local cellular supervision, learnable global context priors and spot-level self-attention to transform histology anchors into transcriptomics-oriented representations.

  • We conduct comprehensive experiments on three widely used ST datasets and ten curated HEST-1k cohorts, showing that BCARP improves overall cross-cohort performance, achieving the highest mean PCC-M and best mean rank while exhibiting cohort-dependent gains.

Deep learning has become central to WSI analysis, enabling histology images to be represented at cellular, patch and slide levels. Early specialized models focus on local biological structures; for example, HoVer-Net [24] performs nuclear instance segmentation and classification by predicting horizontal and vertical distance maps, providing nuclei-level cues for downstream tissue analysis. Beyond individual cells, graph-based approaches such as HEAT [8] model WSIs as heterogeneous graphs with nucleus-type node attributes and semantic edge relations, capturing contextual interactions among cellular components. In parallel, self-supervised pathology encoders have been developed to learn transferable patch representations from large unannotated WSI collections. CTransPath [9] combines Convolutional Neural Network and Swin Transformer architectures with semantically relevant contrastive learning, while recent large-scale models such as UNI [25], Phikon-v2 [10] and Virchow2 [11] scale DINO-style pretraining to massive pathology datasets and provide strong visual features for diverse diagnostic and biomarker prediction tasks. Other methods further improve WSI-level representation by aggregating patch embeddings into compact morphological prototypes, as in PANTHER [12] or by aligning histology images with pathology text, as in CONCH [26].

Predicting spatial gene expression from H&E-stained histology images offers a scalable alternative to experimental spatial transcriptomics. ST-Net [1] formulates the task as patch-level regression, ignoring inter-spot dependencies and broader tissue organization; subsequent work introduces spatial modeling, retrieval, graph construction, multi-scale encoding, multimodal alignment and distribution-level prediction. THItoGene [3] models dependencies among spatial locations, while BG-TRIPLEX [4] adds boundary, edge and nuclei cues through cross-attention. Among the methods we adopt as baselines, MERGE [28] constructs multi-faceted hierarchical graphs that connect spots by both spatial proximity and morphological similarity; TRIPLEX [23] fuses target-spot, neighborhood and whole-slide views; M2OST [32] regresses a central spot from multi-magnification patches through a many-to-one Transformer, and therefore operates on image crops rather than pre-extracted features; mclSTExp [33] aligns H&E and spatial-transcriptomic representations by contrastive learning and infers expression through retrieval; EGN [21] refines target representations using globally similar exemplars; and STFlow [31] models the whole-slide joint distribution with flow matching under a zero-inflated negative-binomial prior. Beyond these, MMAP [5] summarizes slide-level context with a prototype bank, ASIGN [6] extends imputation to 3D tissue, and GHIST [27] and HistoCell [7] predict at cellular resolution.

Figure 1. Pipeline overview of BCARP. Frozen histology foundation embeddings are first projected into a shared latent space, then iteratively refined through context-conditioned modulation, local biological encoding and prototype-guided global context propagation. The final representation is passed to a pattern-scale gene expression decoder to generate calibrated spatial gene expression predictions.

Methodology

Figure 1 illustrates the overall pipeline of BCARP. Let a tissue section contain NN spatial transcriptomics spots. We denote the frozen histology embeddings as E∈RN×de\mathbf{E}\in\mathbb{R}^{N\times d_e}, the spatial coordinates as P∈RN×2\mathbf{P}\in\mathbb{R}^{N\times 2} and the normalized coordinates as P~\widetilde{\mathbf{P}}, obtained by normalizing each coordinate axis by its maximum value. BCARP predicts GG target genes, with Yproc∈RN×G\mathbf{Y}^{\mathrm{proc}}\in\mathbb{R}^{N\times G} denoting the spatially smoothed target for expression-pattern supervision and Yraw∈RN×G\mathbf{Y}^{\mathrm{raw}}\in\mathbb{R}^{N\times G} denoting the raw-count target.

Histology Embedding Projection and Biology-Guided Local-Global Context Refinement

Histology embedding projection

Frozen pathology foundation embeddings provide strong morphology-aware features but are not directly optimized for spatial gene expression prediction. BCARP therefore maps them into a transcriptomics-oriented latent space using a lightweight MLP projector fproj(⋅)f_{\mathrm{proj}}(\cdot) as H0=fproj(E)\mathbf{H}^{0} = f_{\mathrm{proj}} \left( \mathbf{E} \right), then H0∈RN×d\mathbf{H}^{0}\in\mathbb{R}^{N\times d} where dd is the hidden dimension. The projected state H0\mathbf{H}^{0} serves as a stable histology-derived anchor, from which BCARP repeatedly constructs context-conditioned views and refines them with local cellular and global context priors before decoding gene expression.

Context-aware histology anchor refinement

Spatial gene expression is influenced by not only local morphology but also cellular identity, neighborhood composition, and tissue-level organization [27]. To capture these dependencies, BCARP performs RR context-conditioned refinement steps over H0\mathbf{H}^{0} using context-conditioned anchor modulation. Let Cr∈RN×d\mathbf{C}^{r}\in\mathbb{R}^{N\times d} denote the context state after step rr, with C0=0\mathbf{C}^{0}=\mathbf{0}. At step rr, we first aggregate the previous context into a section-level summary: cˉr−1=1N∑i=1NCir−1,cˉr−1∈Rd. \bar{\mathbf{c}}^{r-1} = \frac{1}{N} \sum_{i=1}^{N} \mathbf{C}^{r-1}_{i}, \qquad \bar{\mathbf{c}}^{r-1}\in\mathbb{R}^{d}. The summary is transformed into feature-wise modulations: [βr,γr,ar]=fmod(LN(cˉr−1)),βr,γr,ar∈Rd, \left[ \boldsymbol{\beta}^{r}, \boldsymbol{\gamma}^{r}, \mathbf{a}^{r} \right] = f_{\mathrm{mod}} \left( \mathrm{LN} \left( \bar{\mathbf{c}}^{r-1} \right) \right), \boldsymbol{\beta}^{r}, \boldsymbol{\gamma}^{r}, \mathbf{a}^{r} \in\mathbb{R}^{d}, where βr\boldsymbol{\beta}^{r}, γr\boldsymbol{\gamma}^{r} and ar\mathbf{a}^{r} denote shift, scale and gate parameters, respectively, LN\mathrm{LN} denotes LayerNorm, and fmod(⋅)f_{\mathrm{mod}}(\cdot) is a MLP. These parameters are broadcast to all spots and used to construct a context-modulated histology state: H~r=H0+tanh⁡(ar)⊙[LN(H0)⊙(1+γr)+βr], \widetilde{\mathbf{H}}^{r} = \mathbf{H}^{0} + \tanh \left( \mathbf{a}^{r} \right) \odot \left[ \mathrm{LN} \left( \mathbf{H}^{0} \right) \odot \left( 1+\boldsymbol{\gamma}^{r} \right) + \boldsymbol{\beta}^{r} \right], where ⊙\odot denotes element-wise multiplication. This gated modulation preserves the original morphology-derived representation while allowing accumulated context to adaptively reshape it at each refinement step. The resulting state H~r\widetilde{\mathbf{H}}^{r} is then passed to the local and global refinement modules, yielding the update path: Cr−1→H~r→Lr→Cr,\mathbf{C}^{r-1}\rightarrow\widetilde{\mathbf{H}}^{r}\rightarrow\mathbf{L}^{r}\rightarrow\mathbf{C}^{r}, where Lr\mathbf{L}^{r} is the locally refined representation obtained from nuclei-guided cellular context encoding and Cr\mathbf{C}^{r} is updated after prototype-guided global context propagation.

Local cellular context encoding

The local stage injects cellular and microenvironmental cues into the context-modulated spot representation. BCARP first constructs a fixed spatial adjacency from normalized spot coordinates by connecting each spot to its kk nearest spatial neighbors: As=kNN(P~;k),As∈{0,1}N×N. \mathbf{A}_{s} = \mathrm{kNN}( \widetilde{\mathbf{P}};k ), \qquad \mathbf{A}_{s}\in\{0,1\}^{N\times N}. The adjacency As\mathbf{A}_{s} is computed once and shared across all refinement steps. It only specifies the local propagation structure, while the node features are dynamically given by the context-modulated states H~r\widetilde{\mathbf{H}}^{r} at step rr.

To provide biological supervision, BCARP derives two auxiliary local targets from H&E-based nuclei analysis. We apply HoVer-Net [24] to detect nuclei and assign each nucleus a cell-type pseudo-label. For each spot ii, we collect nuclei whose centroids fall within a fixed radius around its coordinate pi\mathbf{p}_{i}. Let TT denote the number of cell-type categories and let nitn_{it} be the number of nuclei of type tt in the neighborhood of spot ii. We define the neighborhood-composition target and dominant cell-type target as

qi=[nit∑t′=1Tnit′+ϵ]t=1T,ci=arg⁡max⁡tnit.\begin{equation} \mathbf{q}_{i} = \left[ \frac{ n_{it} }{ \sum_{t'=1}^{T} n_{it'}+\epsilon } \right]_{t=1}^{T}, \qquad c_i = \arg\max_{t} n_{it}. \tag*{(6)} \end{equation}

Here qi\mathbf{q}_{i} represents the local cell-type composition, while cic_i summarizes the dominant local phenotype. These targets are used only for training-time auxiliary supervision. Given H~r\widetilde{\mathbf{H}}^{r} and As\mathbf{A}_{s}, BCARP uses two GNN-based local context encoders, indexed by b∈{ct,nc}b\in\{\mathrm{ct},\mathrm{nc}\}, where ct\mathrm{ct} is cell-type and nc\mathrm{nc} denotes neighborhood-composition context:

Ubr=Rb(H~r,As),Obr=fhead,b(Ubr),\begin{equation} \mathbf{U}_{b}^{r} = \mathcal{R}_{b} \big( \widetilde{\mathbf{H}}^{r}, \mathbf{A}_{s} \big), \qquad \mathbf{O}_{b}^{r} = f_{\mathrm{head},b} \left( \mathbf{U}_{b}^{r} \right), \tag*{(7)} \end{equation}

where Ubr∈RN×dr\mathbf{U}_{b}^{r}\in\mathbb{R}^{N\times d_r} denotes graph-updated local features, Obr∈RN×T\mathbf{O}_{b}^{r}\in\mathbb{R}^{N\times T} denotes auxiliary logits and fhead,bf_{\mathrm{head},b} is an MLP. The cell-type branch and neighborhood-composition branch are supervised by cic_i and qi\mathbf{q}_{i}, respectively. The corresponding predictions are

Y^ctr=softmax(Octr),Y^ncr=softmax(Oncr).\begin{equation} \widehat{\mathbf{Y}}_{\mathrm{ct}}^{r} = \mathrm{softmax} \left( \mathbf{O}_{\mathrm{ct}}^{r} \right), \qquad \widehat{\mathbf{Y}}^{r}_{\mathrm{nc}} = \mathrm{softmax} \left( \mathbf{O}_{\mathrm{nc}}^{r} \right). \tag*{(8)} \end{equation}

Hidden features from both branches are projected back to the BCARP hidden dimension via an MLP projector fbf_{b} and injected into the main refinement trajectory:

δbr=fb(Ubr),δbr∈RN×d,b∈{ct,nc}.\begin{equation} \boldsymbol{\delta}_{b}^{r} = f_{b} \left( \mathbf{U}_{b}^{r} \right), \qquad \boldsymbol{\delta}_{b}^{r}\in\mathbb{R}^{N\times d}, \qquad b\in\{\mathrm{ct},\mathrm{nc}\}. \tag*{(9)} \end{equation}

We obtain locally refined state by gated residual injection:

Lr=H~r+∑b∈{ct,nc}γbδbr,γb=tanh⁡(αb),\begin{equation} \mathbf{L}^{r} = \widetilde{\mathbf{H}}^{r} + \sum_{b\in\{\mathrm{ct},\mathrm{nc}\}} \gamma_{b} \boldsymbol{\delta}_{b}^{r}, \quad \gamma_{b}=\tanh \left( \alpha_{b} \right), \tag*{(10)} \end{equation}

where αb\alpha_{b} is a learnable residual-gate parameter for branch bb. This design allows BCARP to preserve context-modulated histology representation while injecting the cellular phenotype and the local microenvironmental mixture.

Prototype-guided global context propagation

Local graph encoding captures proximal biological context, but spatial gene expression may depend on broader tissue-level organization. BCARP introduces a small set of learnable global-context prototypes to provide a shared global-context interface. Each spot is allowed to retrieve global prior token information and then exchange the enriched information with other spots through self-attention. We first ground the locally refined representation in tissue geometry. Let fposf_{\mathrm{pos}} be a spatial positional encoder that maps normalized spot coordinates to dd-dimensional positional features:

Xr=Lr+fpos(P~),Xr∈RN×d,\begin{equation} \mathbf{X}^{r} = \mathbf{L}^{r} + f_{\mathrm{pos}} ( \widetilde{\mathbf{P}} ), \qquad \mathbf{X}^{r}\in\mathbb{R}^{N\times d}, \tag*{(11)} \end{equation}

where Xr\mathbf{X}^{r} is the spatially grounded spot representation at refinement step rr. To present the attention operations compactly, we define a multi-head cross-attention operator A(M,N)\mathcal{A}(\mathbf{M},\mathbf{N}), where M\mathbf{M} provides queries and N\mathbf{N} provides keys and values. For the mm-th attention head, we compute

Am(M,N)=softmax(MWQm(NWKm)⊤dh)NWVm,\begin{equation} \mathcal{A}_{m}(\mathbf{M},\mathbf{N}) = \mathrm{softmax} \left( \frac{ \mathbf{M}\mathbf{W}^{m}_{Q} \big(\mathbf{N}\mathbf{W}^{m}_{K}\big)^{\top} }{ \sqrt{d_h} } \right) \mathbf{N}\mathbf{W}^{m}_{V}, \tag*{(12)} \end{equation}

where HH is the number of attention heads and dh=d/Hd_h=d/H is the dimension of each head. The outputs of all attention heads are concatenated to the dd-dimensional feature:

A(M,N)=Concat(A1(M,N),…,AH(M,N)).\begin{equation} \mathcal{A}(\mathbf{M},\mathbf{N}) = \mathrm{Concat} \big( \mathcal{A}_{1}(\mathbf{M},\mathbf{N}), \ldots, \mathcal{A}_{H}(\mathbf{M},\mathbf{N}) \big). \tag*{(13)} \end{equation}

When M≡N\mathbf{M}\equiv\mathbf{N}, this reduces to standard multi-head self-attention. BCARP maintains KctxK_{\mathrm{ctx}} learnable global prototype tokens, denoted as T∈RKctx×d\mathbf{T}\in\mathbb{R}^{K_{\mathrm{ctx}}\times d}, which serve as trainable global context priors without requiring predefined biological classes or physical tissue centroids. At each refinement step, every spot first attends to these tokens, then fuses the retrieved global context through residual normalization, using global-token-to-spot attention module:

Octxr=A(Xr,T),Xˉctxr=LN(Xr+Octxr).\begin{equation} \begin{aligned} \mathbf{O}_{\mathrm{ctx}}^{r} = \mathcal{A} \left( \mathbf{X}^{r}, \mathbf{T} \right), \quad \bar{\mathbf{X}}_{\mathrm{ctx}}^{r} = \mathrm{LN} \left( \mathbf{X}^{r} + \mathbf{O}_{\mathrm{ctx}}^{r} \right). \end{aligned} \tag*{(14)} \end{equation}

This attention provides each spot with a weighted combination of learned global prototype-context anchors, while the residual normalization injects this global prior into the spatially grounded spot representation before spot-level interaction. After receiving prototype-guided global priors, the enriched spot representations exchange and propagate information through spot-level self-attention module:

Gr=A(Xˉctxr,Xˉctxr),Gr∈RN×d.\begin{equation} \mathbf{G}^{r} = \mathcal{A} \left( \bar{\mathbf{X}}_{\mathrm{ctx}}^{r}, \bar{\mathbf{X}}_{\mathrm{ctx}}^{r} \right), \qquad \mathbf{G}^{r}\in\mathbb{R}^{N\times d}. \tag*{(15)} \end{equation}

We obtain next context state by a residual feed-forward as:

Cr=LN(Xr+fFFN(Xˉctxr+Gr)).\begin{equation} \mathbf{C}^{r} = \mathrm{LN} \big( \mathbf{X}^{r} + f_{\mathrm{FFN}} \big( \bar{\mathbf{X}}_{\mathrm{ctx}}^{r} + \mathbf{G}^{r} \big) \big). \tag*{(16)} \end{equation}

The residual path preserves the spatially grounded local state, while the feed-forward branch incorporates both token-routed global context and spot-level attention output.

Pattern-Scale Gene Expression Decoder

BCARP decodes the final refined spot representation Z=CR∈RN×d\mathbf{Z}=\mathbf{C}^{R}\in\mathbb{R}^{N\times d} with a pattern-scale gene expression decoder. The decoder factorizes prediction into two complementary components including a pattern branch that estimates relative spatial gene expression structure and a scale branch that calibrates this structure to the raw-count expression space.

For each spot ii, let zi∈Rd\mathbf{z}_{i}\in\mathbb{R}^{d} denote the ii-th row of Z\mathbf{Z}. We first map zi\mathbf{z}_{i} into the decoder space and expand it into a short sequence of SS hierarchy tokens by adding learnable scale embeddings: ti(s)=fdec(zi)+e(s),s=1,…,S. \mathbf{t}_{i}^{(s)} = f_{\mathrm{dec}} \left( \mathbf{z}_{i} \right) + \mathbf{e}^{(s)}, \qquad s=1,\ldots,S. The scale embeddings induce a compact coarse-to-fine hierarchy over gene groups. A causal transformer decoder processes this sequence, allowing later tokens to condition on earlier decoding states: ui(1),…,ui(S)=Decoder(ti(1),…,ti(S)), \mathbf{u}_{i}^{(1)},\ldots,\mathbf{u}_{i}^{(S)} = \mathrm{Decoder} \big( \mathbf{t}_{i}^{(1)},\ldots,\mathbf{t}_{i}^{(S)} \big), where hi=ui(S)\mathbf{h}_{i}=\mathbf{u}_{i}^{(S)} is used for the spot-level decoding state.

Pattern branch

Given the decoding state hi\mathbf{h}_{i}, the pattern branch predicts a non-negative processed expression vector: y^iproc=p^i=softplus⁡(fpat(hi)),p^i∈R+G. \widehat{\mathbf{y}}_{i}^{\mathrm{proc}}=\widehat{\mathbf{p}}_{i} = \operatorname{softplus} \left( f_{\mathrm{pat}} \left( \mathbf{h}_{i} \right) \right), \qquad \widehat{\mathbf{p}}_{i}\in\mathbb{R}_{+}^{G}. Stacking the predictions over all spots gives the processed pattern prediction Y^proc∈R+N×G\widehat{\mathbf{Y}}^{\mathrm{proc}}\in\mathbb{R}_{+}^{N\times G}. This branch is supervised by the spatially smoothed expression target Yproc\mathbf{Y}^{\mathrm{proc}}.

Scale calibration branch

To calibrate the predicted pattern to the raw-count scale, we normalize p^i\widehat{\mathbf{p}}_{i} into a gene-wise composition π^i=p^i/(1⊤p^i+ϵ)\widehat{\boldsymbol{\pi}}_{i}=\widehat{\mathbf{p}}_{i}/(\mathbf{1}^{\top}\widehat{\mathbf{p}}_{i}+\epsilon). Here, π^i\widehat{\boldsymbol{\pi}}_{i} represents the predicted proportion of the total expression assigned to each gene at spot ii. In parallel, a library-size head predicts a positive spot-level raw-count scale: ℓ^i=exp⁡(softplus⁡(flib(hi)))−1,ℓ^i∈R+. \widehat{\ell}_{i} = \exp \left( \operatorname{softplus} \left( f_{\mathrm{lib}} \left( \mathbf{h}_{i} \right) \right) \right) - 1, \qquad \widehat{\ell}_{i}\in\mathbb{R}_{+}. The calibrated raw-count prediction is obtained by rescaling the gene-wise composition with the predicted library size: y^iraw=ℓ^iπ^i∈R+G. \widehat{\mathbf{y}}_{i}^{\mathrm{raw}} = \widehat{\ell}_{i} \widehat{\boldsymbol{\pi}}_{i}\in\mathbb{R}_{+}^{G}. Stacking over all spots gives the predicted raw-count matrix Y^raw∈R+N×G\widehat{\mathbf{Y}}^{\mathrm{raw}}\in\mathbb{R}_{+}^{N\times G}, where each entry Y^igraw\widehat{Y}^{\mathrm{raw}}_{ig} denotes the predicted raw expression count of gene gg at spot ii.

This factorization separates relative gene expression pattern learning from spot-specific raw-count calibration. The pattern branch predicts how expression is distributed across genes, while the scale branch estimates the total count level of each spot and their product yields the final gene-wise raw-count prediction.

Objective

Figure 2. ECDF curves show the distribution of fold-averaged per-gene PCC on HER2ST, Skin and ST-Net datasets. Dashed vertical lines indicate the average per-gene PCC of BCARP and the highest competing baseline.

BCARP is trained with a multi-task objective that combines processed pattern supervision, raw-count calibration and local biological guidance. The processed pattern loss is Lproc=LPCC+λhuberLHuber\mathcal{L}_{\mathrm{proc}} = \mathcal{L}_{\mathrm{PCC}} + \lambda_{\mathrm{huber}} \mathcal{L}_{\mathrm{Huber}}, where

LPCC=1−1G∑g=1GCorr⁡(Y^:,gproc,Y:,gproc),LHuber=Huber⁡δ(Y^proc,Yproc).\begin{align} &\mathcal{L}_{\mathrm{PCC}} = 1 - \frac{1}{G} \sum_{g=1}^{G} \operatorname{Corr} \big( \widehat{\mathbf{Y}}^{\mathrm{proc}}_{:,g}, \mathbf{Y}^{\mathrm{proc}}_{:,g} \big), \tag*{(22)}\\ &\mathcal{L}_{\mathrm{Huber}} = \operatorname{Huber}_{\delta} \big( \widehat{\mathbf{Y}}^{\mathrm{proc}}, \mathbf{Y}^{\mathrm{proc}} \big). \tag*{(23)} \end{align}

For raw-count calibration, Y^raw\widehat{\mathbf{Y}}^{\mathrm{raw}} is supervised by the raw UMI target Yraw\mathbf{Y}^{\mathrm{raw}} using a negative-binomial likelihood:

LNB=−1NG∑i=1N∑g=1Glog⁡pNB(Yigraw∣Y^igraw,θg),\begin{equation} \mathcal{L}_{\mathrm{NB}} = - \frac{1}{NG} \sum_{i=1}^{N} \sum_{g=1}^{G} \log p_{\mathrm{NB}} \big( Y^{\mathrm{raw}}_{ig} \mid \widehat{Y}^{\mathrm{raw}}_{ig}, \theta_g \big), \tag*{(24)} \end{equation}

where Y^igraw\widehat{Y}^{\mathrm{raw}}_{ig} is used as the predicted count-rate parameter and θg\theta_g is the gene-wise dispersion. We further apply a log-space MSE loss to align the calibrated prediction with the raw-count scale:

LlogMSE=MSE⁡(log⁡(1+Y^raw),log⁡(1+Yraw)).\begin{equation} \mathcal{L}_{\mathrm{logMSE}} = \operatorname{MSE} \big( \log \big( 1+\widehat{\mathbf{Y}}^{\mathrm{raw}} \big), \log \big( 1+\mathbf{Y}^{\mathrm{raw}} \big) \big). \tag*{(25)} \end{equation}

The decoder loss is

Ldec=Lproc+λNBLNB+λlogMSELlogMSE\begin{equation} \mathcal{L}_{\mathrm{dec}} = \mathcal{L}_{\mathrm{proc}} + \lambda_{\mathrm{NB}} \mathcal{L}_{\mathrm{NB}} + \lambda_{\mathrm{logMSE}} \mathcal{L}_{\mathrm{logMSE}} \tag*{(26)} \end{equation}

The local biological branches are supervised by HoVer-Net-derived pseudo-labels. The cell-type loss is averaged over refinement steps:

Lct=1R∑r=1RCE⁡(Y^ctr,c),\begin{equation} \mathcal{L}_{\mathrm{ct}} = \frac{1}{R} \sum_{r=1}^{R} \operatorname{CE} \big( \widehat{\mathbf{Y}}_{\mathrm{ct}}^{r}, \mathbf{c} \big), \tag*{(27)} \end{equation}

where c={ci}i=1N\mathbf{c}=\{c_i\}_{i=1}^{N} denotes the dominant cell-type pseudo-labels. The neighborhood-composition loss is the average KL divergence between the soft composition target qi\mathbf{q}_i and the predicted distribution Y^nc,ir\widehat{\mathbf{Y}}_{nc,i}^{r}:

Lnc=1R∑r=1R1N∑i=1N∑t=1Tqit[log⁡qit−log⁡Y^nc,itr].\begin{equation} \mathcal{L}_{\mathrm{nc}} = \frac{1}{R} \sum_{r=1}^{R} \frac{1}{N} \sum_{i=1}^{N} \sum_{t=1}^{T} q_{it} \left[ \log q_{it} - \log \widehat{\mathbf{Y}}^{r}_{nc,it} \right]. \tag*{(28)} \end{equation}

The final training objective is

L=Ldec+λctLct+λncLnc.\begin{equation} \mathcal{L} = \mathcal{L}_{\mathrm{dec}} + \lambda_{\mathrm{ct}} \mathcal{L}_{\mathrm{ct}} + \lambda_{\mathrm{nc}} \mathcal{L}_{\mathrm{nc}}. \tag*{(29)} \end{equation}

Experiments

HER2STSkinST-Net
ModelPCC-H↑\uparrowPCC-M↑\uparrowMSE↓\downarrowMAE↓\downarrowRVD↓\downarrowPCC-H↑\uparrowPCC-M↑\uparrowMSE↓\downarrowMAE↓\downarrowRVD↓\downarrowPCC-H↑\uparrowPCC-M↑\uparrowMSE↓\downarrowMAE↓\downarrowRVD↓\downarrow
Frozen-feature references
ResNet50+Ridge0.22490.22911.07820.82240.90390.26950.32601.35650.90500.85810.09850.08900.97620.78620.9125
UNI+Ridge0.25830.25981.05790.81440.88400.28950.32621.33820.89680.88120.10960.09360.96660.78210.9223
UNI+MLP0.34740.33161.00300.78900.84520.30290.36041.25470.86700.83680.18430.16330.91950.76090.8644
UNI+Decoder0.42490.37210.94310.77240.81880.53920.49301.44550.95381.45110.23190.18680.82830.73510.7685
Method baselines
M2OST0.31210.27560.98530.76962.14420.29330.27391.38180.92100.57390.15440.12390.76670.69781.5675
MERGE0.29130.28001.08850.82740.89100.28270.30841.35980.90520.90740.12140.11080.95290.77760.9234
EGN0.38840.28741.21410.99530.88570.40390.37040.95050.87340.89880.10980.11110.98260.77230.9673
TRIPLEX0.35810.34271.07240.81560.89900.33370.37491.33330.89470.91370.16020.14500.96940.78060.9263
mclSTExp0.38560.33560.95280.75890.81430.52120.47221.49830.98620.44450.12690.10970.86590.73750.8548
STFlow0.38880.34100.84640.72520.85550.51370.48321.19830.86540.49510.19710.15650.74040.68970.7245
BCARP (Ours)0.45670.40320.83810.71910.70840.58520.53281.18180.86200.26560.30460.24230.67610.65900.7500

Table 1. Performance comparison on HER2ST, Skin and ST-Net datasets. PCC-H and PCC-M denote average per-gene PCC over highly variable genes and all target genes, respectively. MSE and MAE measure prediction error in expression space, and RVD measures the relative deviation in expression variance between predictions and ground truth. Among the method baselines, best results are shown in bold and second-best results are underlined; the frozen-feature references are reported for context and are excluded from this ranking.

Settings

Datasets and preprocessing

We evaluate BCARP on three established whole-slide spatial transcriptomics benchmarks, HER2ST [22], Skin [30], and ST-Net [1], as well as ten HEST-1k [2] cohorts, spanning nine organs and eight cancer types. Each sample consists of an H&E-stained histology image with spatially resolved gene expression, where each spot is associated with a tissue coordinate, a spot-centered image patch, and an expression vector. For HER2ST, Skin, and ST-Net, we follow the unified spot-level protocol of MERGE [28]: all methods use the same patient-level splits, 224×224224\times224 spot-centered patches, tissue coordinates, and 250-gene target panel after gene filtering. For HEST-1k, we use the predefined patient-grouped splits for each cohort and select up to 250 highly variable genes, resulting in 250 genes for eight cohorts, 197 genes for LUNG, and 98 genes for PAAD. We apply SPCS smoothing [29] to obtain processed expression targets for pattern-oriented supervision, while retaining raw-count targets for calibration-oriented objectives. All evaluation metrics are computed in the same log⁡(1+x)\log(1 + x)-transformed raw expression space.

Baselines and implementation details

We compare BCARP with representative histology-guided expression prediction methods covering graph modeling, multi-resolution fusion, contrastive retrieval, exemplar refinement, whole-slide generative modeling, and frozen-feature prediction. The method baselines include MERGE [28], TRIPLEX [23], mclSTExp [33], EGN [21], STFlow [31], and M2OST [32]. For fairness, baselines that support pre-extracted features are evaluated with the same frozen UNI [25] embeddings used by BCARP, replacing their original image encoders. STFlow is already based on UNI features, while M2OST is not encoder-matched because its patch-tokenized multi-field-of-view design requires image crops as input. We further include frozen-feature reference baselines to isolate the effect of the visual backbone and prediction head. On the main datasets, UNI+Ridge, UNI+MLP, and UNI+Decoder evaluate frozen UNI embeddings with increasingly expressive predictors. On HEST-1k, we follow the standard ridge-probe protocol and compare UNI, ResNet50 [35], Virchow2 [11], GigaPath [34] and CONCH [26] embeddings under the same ridge-regression head. Source code is publicly available at https://github.com/anvuv/BCARP.

ModelSKCMLUNGPAADREADCOADLYMPHCCRCCIDCPRADHCCMeanMean
Rank↓\downarrow
Frozen-feature ridge probes
ResNet50+Ridge0.55610.53100.39630.07380.34630.26420.17290.55640.20050.06760.316510.8
CONCH+Ridge0.61380.58190.45850.19410.42140.31310.18160.59950.30310.07800.37457.7
UNI+Ridge0.63400.58330.48630.19870.42890.30200.20770.62120.23080.08680.37806.3
GigaPath+Ridge0.60250.58760.48230.20840.43240.29580.19670.61220.28860.09480.38016.2
Virchow2+Ridge0.63340.60050.48050.20510.42200.30290.22500.64110.27530.11180.38985.4
Method baselines
M2OST0.56510.55720.36230.06660.01440.19520.11490.49440.25840.11550.274410.4
MERGE0.53610.45940.37510.14110.32420.31200.22200.52090.27590.11400.32819.0
EGN0.64510.57740.45860.15900.43550.24200.19220.60280.22990.09450.36377.9
mclSTExp0.59980.59010.46030.23040.44110.32430.21410.59650.32960.11410.39004.9
TRIPLEX0.67660.60420.46900.16640.44150.31450.23370.61570.32790.07230.39224.3
STFlow0.69680.63150.49400.26290.43430.31340.22660.61180.31530.17900.41662.6
BCARP (Ours)0.66040.61620.48570.25720.44910.38200.24710.59490.34870.17540.42172.5

Table 2. PCC-M performance across the ten HEST-1k cohorts. Mean denotes the average PCC-M across cohorts, and Mean Rank denotes the average within-cohort rank, where lower is better. Best results are shown in bold, and second-best results are underlined.

Comparative Results

Table 1 reports that BCARP obtains the highest PCC-H and PCC-M on all three datasets. On HER2ST, BCARP reaches 0.4567/0.40320.4567/0.4032, compared with 0.3888/0.34100.3888/0.3410 for STFlow and 0.3581/0.34270.3581/0.3427 for TRIPLEX, the strongest method baselines on this dataset. On Skin, BCARP reaches 0.5852/0.53280.5852/0.5328 against 0.5137/0.48320.5137/0.4832 for STFlow. The largest improvement is on ST-Net, where BCARP obtains 0.3046/0.24230.3046/0.2423 while no competing method exceeds 0.1971/0.15650.1971/0.1565. This indicates that BCARP recovers the spatial layout of expression more faithfully, rather than only reducing pointwise numerical error. BCARP achieves the lowest MAE across all three datasets, the lowest RVD on HER2ST and Skin, and the lowest MSE on HER2ST and ST-Net. UNI+Ridge and UNI+MLP remain well below all refinement-based methods, but UNI+Decoder is a strong reference: it reaches 0.4249/0.37210.4249/0.3721 on HER2ST and 0.5392/0.49300.5392/0.4930 on Skin, which is competitive with, and on Skin better than, every method baseline. This indicates that a substantial part of the achievable correlation is already carried by frozen pathology embeddings with decoder.

Table 2 extends the comparison to ten HEST-1k cohorts [2], which cover a wider range of organs, cancer types and acquisition platforms. BCARP obtains the highest mean PCC-M (0.42170.4217) and the best mean within-cohort rank (2.52.5), followed by STFlow (0.41660.4166, 2.62.6). The two methods are complementary across cohorts: BCARP ranks first on 4/10 cohorts. The largest single-cohort gain is on LYMPH, where BCARP reaches 0.38200.3820 against 0.32430.3243 for the next best method, a cohort in which tissue organization is strongly structured but the local morphological signal is comparatively weak. We also observe that the ridge probes are competitive in this setting and that Virchow2+Ridge ranks first on IDC (0.64110.6411), where BCARP reaches 0.59490.5949. This is consistent with the small size of several HEST cohorts, in which a high-capacity refinement model has fewer spots per section to exploit and a linear probe over strong frozen features is a robust alternative. Overall, the HEST results indicate that the advantage of BCARP transfers beyond the three widely used benchmarks, though the per-cohort margins over STFlow are small and cohort-dependent.

Gene-level distribution analysis

Average correlation can be driven by a small subset of well-predicted genes, so we further examine the full distribution of fold-averaged per-gene PCC using empirical cumulative distribution functions (ECDFs) in Figure 2. A rightward shift of the ECDF indicates that a larger fraction of genes is predicted above any given correlation level. On all three datasets the BCARP curve lies to the right of every baseline over essentially the entire range, and the shift is most pronounced in the lower half of the distribution: the fraction of genes falling below weak correlation values is markedly reduced.

Figure 3

Figure 3. Qualitative comparison on HER2ST for FN1 on 3 slides.

Qualitative analysis

Figure 3 compares predicted and measured FN1 expression on three HER2ST test slides spanning the range of prediction difficulty. On the easiest slide, all methods localise the expressed region and differ only marginally, with STFlow slightly ahead of BCARP. On a moderately difficult slide, BCARP follows the boundary of the expressed region most closely, though all methods overestimate expression across the section and the margin between them is small. On the hardest slide, every method fails, returning nearly inverted or unstructured maps.

VariantPCC-H↑\uparrowPCC-M↑\uparrowMSE↓\downarrowMAE↓\downarrowRVD↓\downarrow
Full0.45670.40320.83810.71910.7084
Main components
w/o Proj.0.42830.37720.90290.75120.8592
w/o CMod0.43790.38640.96080.77510.7870
w/o Local0.44210.38820.88730.74980.8138
w/o Global0.43300.37990.92640.76460.8092
w/o Pos.0.44240.38840.89880.75070.8285
w/o Dec.0.45710.40370.94840.77361.6684
Local-branch ablations
w/o NC0.45910.40410.85310.73380.7552
w/o CT0.45850.40130.89410.75340.7483
Global-module ablations
w/o Prototype0.45860.40610.85380.72820.8123
w/o Self-Attn.0.45540.40320.86790.73230.7724
w/o Global Attn.0.43190.38030.88780.74480.9665

Ablation Analysis

Component ablation

Table tab:component_ablation evaluates the contribution of each BCARP component on HER2ST. We denote the projector, context modulation, local context encoder, global propagation stage, positional encoding, and pattern-scale decoder as Proj., CMod, Local, Global, Pos., and Dec., respectively. For global-module ablations, w/o Global removes the full global stage; w/o Global Attn. removes both prototype attention and spot self-attention but keeps the residual update; w/o Prototype and w/o Self-Attn. remove only the corresponding attention component. The w/o Dec. variant replaces the pattern-scale decoder with a standard MLP head. Removing Proj., CMod, Local, Global, or Pos. consistently weakens performance, showing that BCARP benefits from task-oriented projection, iterative anchor modulation, local biological guidance, long-range context propagation, and explicit tissue geometry. Removing Dec. preserves correlation but substantially worsens MSE, MAE, and RVD, suggesting that the decoder mainly improves expression-scale calibration. For the local branches, w/o CT retains only NC, while w/o NC retains only CT. These results show that neither branch uniformly improves PCC individually, while the complete local module yields better expression-error and variance fidelity. The global-module ablations similarly indicate complementary effects of prototype attention and spot-level self-attention.

Figure 4. Relative percentage change of each metric under different refinement depths compared with R=1R=1 on ST-Net dataset.

Effect of refinement depth

Figure 4 analyzes the effect of refinement depth RR, with all metrics reported relative to the reference R=1R=1 on ST-Net dataset. Relatively, R=5R=5 improves PCC-H/PCC-M by 0.8%/0.3%0.8\%/0.3\% and reduces MSE, MAE and RVD by 2.8%2.8\%, 1.5%1.5\% and 4.4%4.4\%, respectively. This indicates that a small number of refinement steps can already provide strong performance, while moderate additional refinement improves calibration and maintains competitive correlation. We therefore use R=5R=5 as default.

Conclusion

This paper introduced BCARP, a biology-guided context-aware residual propagation framework for histology-guided spatial transcriptomics prediction. BCARP treats frozen pathology embeddings as morphology-derived anchors and progressively refines them with nuclei-guided local cellular context, prototype-guided global propagation, and a pattern-scale decoder. Across HER2ST, Skin, ST-Net, and ten HEST-1k cohorts, BCARP achieves strong correlation-based performance, including the highest mean PCC-M across the HEST-1k cohorts. Ablations show that anchor conditioning and local–global propagation support correlation, while the decoder and nuclei-guided branches contribute more strongly to scale fidelity. However, limitations remain: local supervision relies on HoVer-Net pseudo-labels and may be affected by nuclei detection/classification errors and cross-tissue domain shift, while morphology-only evidence limits prediction for genes weakly associated with H&E appearance. Future work will study label-noise robustness, complementary molecular supervision, adaptive refinement, and efficient attention.

APPENDICES

Problem Formulation

As illustrated in Figure 5, given a whole-slide H&E image, we represent its spatial-transcriptomics (ST) annotation as a set of spots S={(xi,pi,yi)}i=1N\mathcal{S}=\{(\mathbf{x}_i,\mathbf{p}_i,\mathbf{y}_i)\}_{i=1}^{N}, where xi\mathbf{x}_i is the spot-centered histology patch, pi ⁣∈ ⁣R2\mathbf{p}_i\!\in\!\mathbb{R}^{2} is its tissue coordinate and yi ⁣∈ ⁣RG\mathbf{y}_i\!\in\!\mathbb{R}^{G} is the measured expression vector over GG target genes of spot ii. Histology-guided spatial transcriptomics prediction seeks a mapping:

F:{(xj,pj)}j=1N⟼Y^∈RN×G,y^i=Y^i,:.\begin{equation} F: \{ \big( \mathbf{x}_{j}, \mathbf{p}_{j} \big)\}_{j=1}^{N} \longmapsto \widehat{\mathbf{Y}}\in\mathbb{R}^{N\times G}, \qquad \widehat{\mathbf{y}}_{i} = \widehat{\mathbf{Y}}_{i,:}. \tag*{(30)} \end{equation}

Here, FF predicts the section-level expression matrix Y^\widehat{\mathbf{Y}}, and its ii-th row y^i\widehat{\mathbf{y}}_{i} approximates the ground-truth expression profile yi\mathbf{y}_{i} of spot ii.

Predicting gene expression strictly from highly localized morphology is inherently ill-posed. Transcriptomic profiling is highly susceptible to technical artifacts, noise and dropout events (where expressed genes are falsely recorded as zeros). Furthermore, the expression profile of a single spot is heavily dependent not only on its own cellular composition but also on the broader tissue microenvironment and intercellular signaling. To address this, BCARP treats frozen pathology-foundation embeddings as morphology-derived anchors. These anchors are progressively refined through a dual mechanism: biology-guided local context encoding and prototype-guided global context propagation, with the aim of improving both local biological consistency and whole-section spatial context.

Figure 5. Spatial Transcriptomics Prediction from Histology

Dataset Statistics

Table 7 reports, for every fold, the held-out partition, the number of training and test slides and the corresponding number of spatial spots. In total, the evaluation spans 289,413289{,}413 annotated spots across thirteen datasets and diverse tissue morphologies.

To prevent patient-level information leakage, all experiments use patient-disjoint cross-validation, in which every tissue section from a given patient is assigned exclusively to either the training or the test partition. HER2ST [22] contains eight patients and is split into eight folds with one held-out patient per fold; Skin [30] contains four patients and is split into four folds, again one patient per fold; ST-Net [1] contains twenty-three patients, which we group into eight folds so that each fold holds out two to three patients. Together these three datasets contribute 52,91852{,}918 annotated spots.

We additionally evaluate on ten cohorts of the HEST-1k [2], spanning skin, lung, pancreas, bowel, lymph node, kidney, breast, prostate and liver, and acquired with either Visium or Xenium. For these cohorts we adopt the official patient-grouped folds released, which adds a further 236,495236{,}495 spots and 2929 folds.

Target genes.

For HER2ST, Skin, and ST-Net, every method is trained and evaluated on the same fixed 250-gene panel following the MERGE [28] protocol. This panel corresponds to the 250250 most-expressed genes per dataset introduced by TRIPLEX [23] and reused by MERGE [28], and all metrics are computed in the shared log⁡(1+x)\log(1+x) space. For the highly variable gene (HVG) analysis, we additionally report results on a 5050-gene subset per dataset. The HVG subset is selected by a different criterion, gene-wise variability computed from the training spots only. For completeness, we list in Figure 6 the 250250-gene subsets used for each ST dataset. For the HEST-1k cohorts we select up to 250250 highly variable genes per cohort from the training slides only, following the same criterion as above. The resulting panel contains 250250 genes for the eight Visium cohorts, but is capped by the probe set on the two Xenium cohorts, giving 197197 genes for LUNG and 9898 for PAAD.

Figure 6. List of the 250 genes used in all experiments on the three ST datasets.

Baseline Reproduction

Two reproduction protocols.

The scores reported in this section are not identical to those in Table 1 of the main paper, and the difference is deliberate. In the main paper, every baseline whose architecture consumes pre-extracted patch features is re-run with the same frozen UNI [25] embeddings as BCARP, so that the comparison isolates the prediction head from the choice of visual backbone. In this supplement, we instead retain each baseline’s native architecture and visual encoder: every method keeps the patch encoder, feature dimensionality and architectural choices prescribed by its original paper or public code, and only the data splits, the spot-centered patches, the target gene panel and the evaluation metrics are standardised. The main-paper protocol asks how the prediction architectures compare under a common foundation encoder, whereas the supplementary protocol evaluates the corresponding native architectures under the same data and metric setting. The resulting per-gene correlations differ accordingly, in both directions and by method-dependent amounts. The two tables should therefore be read side by side.

To ensure a fair comparison, every competing method is reproduced under a single experimental protocol that follows MERGE [28]: identical patient-level splits, the same spot-centred patches, the same 250250-gene target panel and the same evaluation metrics. Architectures are re-implemented from their official publications or public source code whenever available and hyperparameters follow the original recommendations while sharing a common optimization protocol. Because BCARP uses frozen UNI embeddings, we additionally include UNI+MLP and UNI+Decoder probes to separate the contribution of the refinement loop from that of the pattern-scale decoder. The reproduced baselines are:

  • UNI [25]+MLP – frozen UNI [25] embeddings are mapped directly to gene expression using a standalone MLP predictor.

  • UNI+Decoder - frozen UNI embeddings decoded directly by the pattern-scale decoder, without iterative refinement, local biological modelling or global context propagation.

  • ST-Net [1] - a CNN regression framework (DenseNet-121 backbone with a linear head) that predicts per-spot expression from histology patches without explicit spatial context.

  • HisToGene [19] - a transformer that models dependencies among neighbouring spots through spatial attention.

  • BLEEP [20] - a contrastive multimodal framework that aligns histology and transcriptomics in a shared embedding space; following MERGE [28], we use a ResNet18 patch encoder, the same backbone adopted in MERGE [28] and TRIPLEX [23].

  • TRIPLEX [23] - a multi-resolution architecture fusing target-patch, neighborhood and whole-slide visual information.

  • EGN [21] - an exemplar-guided framework that retrieves similar transcriptomic patterns from reference samples to refine predictions.

  • MERGE [28] - a hierarchical graph neural network with spatial graph reasoning and multi-scale tissue aggregation, reproduced from its official source code.

  • M2OST [32] - a many-to-one regression transformer that jointly consumes patches from several magnification levels of the WSI pyramid to predict the expression of their common spot.

  • STFlow [31] - a flow-matching model that generates the section-level expression matrix through a small number of iterative transport steps.

  • mclSTExp [33] - a contrastive method that retrieves expression profiles from a training bank of spot embeddings.

We standardize data splits, patches, target genes and evaluation metrics across all methods. Training settings are kept as consistent as possible while respecting method-specific architectural requirements.

Frozen-feature ridge probes.

On the HEST-1k cohorts the frozen-feature references reported in the main paper follow the ridge-probe protocol released with the benchmark [2]: patch embeddings are extracted once per spot with the corresponding foundation encoder and regressed onto the target panel with an ℓ2\ell_2-regularised linear head, using the benchmark’s default regularisation strength rather than a per-cohort tuned value. All encoders are compared under exactly this protocol, so differences between them reflect the embeddings and not the head.

On the reproduction of MERGE.

The MERGE results in our evaluation are not directly comparable to its published numbers because the protocols differ along several axes. The original MERGE experiments train and evaluate on SPCS-smoothed expression using eight-fold cross-validation across tissue sections. In contrast, our evaluation uses patient-disjoint folds and computes every reported metric against the raw expression target after the log⁡(1+x)\log(1+x) transform; SPCS is retained only as training-time pattern supervision. In addition, the encoder-controlled results in Table 1 of the main paper replace supported visual encoders with frozen UNI features. This is particularly relevant to MERGE because its hierarchical graph includes feature-space clustering, so changing the visual representation may also alter graph construction. To separate these effects, Supplementary Table 3 retains each baseline’s native architecture and encoder while using our patient-disjoint splits, target panel, and raw-log evaluation. The main-paper table additionally controls the visual backbone using UNI where the architecture permits. The resulting MERGE scores should therefore be interpreted as protocol-specific reproductions rather than direct replications or contradictions of its published results.

HER2STSkinST-Net
ModelPCC-H()PCC-M()MSE()MAE()RVD()PCC-H()PCC-M()MSE()MAE()RVD()PCC-H()PCC-M()MSE()MAE()RVD()
Frozen-feature reference models
ResNet50+Ridge0.22490.22911.07820.82240.90390.26950.32601.35650.90500.85810.09850.08900.97620.78620.9125
UNI+Ridge0.25830.25981.05790.81440.88400.28950.32621.33820.89680.88120.10960.09360.96660.78210.9223
UNI+MLP0.34740.33161.00300.78900.84520.30290.36041.25470.86700.83680.18430.16330.91950.76090.8644
UNI+Decoder0.42490.37210.94310.77240.81880.53920.49301.44550.95381.45110.23190.18680.82830.73510.7685
Method baselines (native configurations)
ST-Net0.32390.24301.25871.00332.81330.37870.33521.00020.86523.34350.17300.15090.79110.71410.8978
HisToGene0.28820.20581.39421.07331.10190.30760.24290.92100.82700.75570.11490.10900.88840.75620.8821
BLEEP0.28460.21290.84230.79621.65380.36940.30440.80240.76781.87890.04300.03510.85720.73963.0336
EGN0.26980.27491.33710.88560.88570.26550.31541.54250.94330.89880.09670.08171.17140.84620.9673
TRIPLEX0.35810.34271.07240.81560.89900.33370.37491.33330.89470.91370.16020.14500.96940.78060.9263
mclSTExp0.36270.31331.02600.78674.89840.46750.42531.92571.11751.18580.14510.11770.85650.73275.6416
M2OST0.31210.27560.98530.76962.14420.29330.27391.38180.92100.57390.15440.12390.76670.69781.5675
MERGE0.16990.16181.58301.04890.80770.17500.16561.92481.18780.88230.06720.05580.84660.74310.9111
STFlow0.38880.34100.84640.72520.85550.51370.48321.19830.86540.49510.19710.15650.74040.68970.7245
BCARP (Ours)0.45670.40320.83810.71910.70840.58520.53281.18180.86200.26560.30460.24230.67610.65900.7500

Table 3. Performance comparison across the three ST datasets under their native architectural and encoder configurations. Bold indicates the best result and underlining indicates the second-best result; frozen-feature references are excluded from this ranking.

Implementation Details

Preprocessing and supervision targets

For each spot we crop a 224 ⁣× ⁣224224\!\times\!224 H&E patch centred on its tissue coordinate and extract a frozen UNI [25] embedding ei ⁣∈ ⁣Rde\mathbf{e}_i\!\in\!\mathbb{R}^{d_e} with de ⁣= ⁣1024d_e\!=\!1024. Two complementary expression targets are used: a spatially smoothed target Yproc\mathbf{Y}^{\mathrm{proc}}, obtained with the two-factor SPCS smoothing used by MERGE [28] and used here to supervise the pattern branch and the raw-count target Yraw\mathbf{Y}^{\mathrm{raw}}, retained for the calibration objectives. Spatial structure is encoded once per section as a fixed kk-nearest-neighbour graph As=kNN(P~;k)\mathbf{A}_s=\mathrm{kNN}(\widetilde{\mathbf{P}};k) built on max-normalised coordinates P~\widetilde{\mathbf{P}}; this adjacency is shared across all refinement steps and only specifies the local propagation structure.

Nuclei-derived biological supervision

Auxiliary local targets are derived from H&E nuclei analysis with HoVer-Net [24], which detects nuclei and assigns each a cell-type pseudo-label. For each spot ii we aggregate nuclei whose centroids fall within a fixed radius ρ\rho (in pixels, measured on the same patch grid as the 224 ⁣× ⁣224224\!\times\!224 crops) of pi\mathbf{p}_i. Letting nitn_{it} be the number of type-tt nuclei in this neighborhood and TT the number of categories, the neighborhood-composition target and dominant cell-type target are

qi=[nit∑t′=1Tnit′+ϵ]t=1T,ci=arg⁡max⁡tnit.\begin{equation} \mathbf{q}_{i} = \big[ \frac{ n_{it} }{ \sum_{t'=1}^{T} n_{it'}+\epsilon } \big]_{t=1}^{T}, \qquad c_i = \arg\max_{t} n_{it}. \tag*{(31)} \end{equation}

These quantities are used only as training-time auxiliary supervision and are not required at inference.

Architecture and optimization

BCARP is implemented in PyTorch. A lightweight MLP projects each frozen UNI [25] embedding from its de ⁣= ⁣1024d_e\!=\!1024-dimensional space into a d=512d=512-dimensional transcriptomics-oriented latent space, which serves as the histology anchor. The anchor is refined for R ⁣= ⁣5R\!=\!5 context-conditioned steps; each step applies context-conditioned anchor modulation, two GraphSAGE local context encoders (cell-type and neighborhood-composition, hidden width dr ⁣= ⁣256d_r\!=\!256) and prototype-guided global propagation with Kctx ⁣= ⁣50K_{\mathrm{ctx}}\!=\!50 learnable global tokens and H ⁣= ⁣8H\!=\!8 attention heads (head dimension dh ⁣= ⁣64d_h\!=\!64). The refined representation is decoded by the pattern-scale decoder using S ⁣= ⁣6S\!=\!6 hierarchy tokens. HoVer-Net [24] targets use radius ρ ⁣= ⁣128\rho\!=\!128 pixels and T ⁣= ⁣6T\!=\!6 nuclei categories. The model is trained with AdamW (learning rate 1 ⁣× ⁣10−41\!\times\!10^{-4}, weight decay 1 ⁣× ⁣10−21\!\times\!10^{-2}), batch size 11 WSI, for at most 100100 epochs, with a ReduceLROnPlateau schedule (factor 0.50.5, patience 77) on a single RTX 5070 Ti GPU. The configuration is summarised in Table 4.

Foundation Backbone
Pathology foundation encoder: UNI [25]
Spot patch size: 224
Embedding dimension ded_e: 1024
Refinement iterations RR: 5
Biology-guided Context Modeling
Spatial neighbors kk: 8
HoVer-Net [24] neighborhood radius ρ\rho: 128 px
Cell-type categories TT: 6
Local encoder: GraphSAGE
Local hidden dimension drd_r: 256
Global prototype tokens KctxK_{\mathrm{ctx}}: 50
Attention heads HH: 8
Head dimension dhd_h: 64
Decoder
Hierarchical decoder scales SS: 6
Dropout: 0.1
Training
Optimizer: AdamW
Learning rate: 1×10−41\times10^{-4}
Weight decay: 1×10−21\times10^{-2}
Batch size: 1 WSI
Epochs: 100 (max)
Learning-rate scheduler: ReduceLROnPlateau
(factor 0.5, patience 7)
Hardware: RTX 5070 Ti
Loss
Huber threshold δ\delta: 1.0
λHuber\lambda_{\mathrm{Huber}}: 0.5
λNB\lambda_{\mathrm{NB}}: 0.05
λlogMSE\lambda_{\mathrm{logMSE}}: 0.1
λct\lambda_{\mathrm{ct}}: 0.05
λnc\lambda_{\mathrm{nc}}: 0.05

Table 4. Hyperparameter configuration of BCARP.

LabelsPCC-H↑\uparrowPCC-M↑\uparrowMSE↓\downarrowMAE↓\downarrowRVD↓\downarrow
HoVer-Net0.45670.40320.83810.71910.7084
CellViT0.46340.40600.87690.74260.6998
Method / componentTrainable (M)E2E / slide (s)Peak (GB)
Baselines based on UNI embedding
UNI+Ridge0.261.280.22
mclSTExp41.461.290.75
MERGE21.531.290.84
STFlow1.351.321.03
TRIPLEX63.121.576.88
EGN77.011.593.80
BCARP (ours)21.101.371.35
Native image encoder (no UNI)
M2OST7.040.8511.75
Foundation model controls.

The HEST ridge-probe comparison uses ResNet50 [35], UNI [25], Virchow2 [11], CONCH [26], and GigaPath [34] embeddings with the same linear head and fixed regularisation setting. We also examined OmiCLIP/Loki [37] and STORM [38]. Loki (built on OmiCLIP) uses a transcriptomic reference bank at prediction time and therefore follows a different reference-assisted setting. We attempted to reproduce STORM using its released implementation but could not establish a reliable matched reproduction under our protocol, so we do not report potentially configuration-dependent numbers.

Evaluation metrics

Let Yraw,Y^raw ⁣∈ ⁣RN×G\mathbf{Y}^{\mathrm{raw}},\hat{\mathbf{Y}}^{\mathrm{raw}}\!\in\!\mathbb{R}^{N\times G} be the ground-truth and predicted raw-count matrices over NN test spots and GG genes. Unless otherwise noted, all metrics are computed after the log⁡(1+x)\log(1+x) transform; we denote the transformed matrices by

Ylog:=log⁡ ⁣(1+Yraw),Y^log:=log⁡ ⁣(1+Y^raw),\begin{equation} \mathbf{Y}^{\mathrm{log}}:=\log\!\big(1+\mathbf{Y}^{\mathrm{raw}}\big),\qquad \hat{\mathbf{Y}}^{\mathrm{log}}:=\log\!\big(1+\hat{\mathbf{Y}}^{\mathrm{raw}}\big), \tag*{(32)} \end{equation}

and write Yiglog\mathbf{Y}^{\mathrm{log}}_{ig} and Y^iglog\hat{\mathbf{Y}}^{\mathrm{log}}_{ig} for ii-th spot and gg-th gene.

For each held-out fold, metrics are computed using only the spots in that fold. Dataset-level scores in the main tables are the arithmetic mean of the corresponding fold-level scores. For PCC, we first compute each gene’s correlation within a fold and then average over the valid genes in the corresponding gene set.

Per-gene Pearson correlation (PCC-M, PCC-H). For each gene we compute the PCC across spots,

PCCg=∑i(Y^iglog−Y^ˉglog)(Yiglog−Yˉglog)∑i(Y^iglog−Y^ˉglog)2 ∑i(Yiglog−Yˉglog)2,\begin{equation} \mathrm{PCC}_g= \frac{\sum_i (\hat{\mathbf{Y}}^{\mathrm{log}}_{ig}-\bar{\hat{\mathbf{Y}}}^{\mathrm{log}}_{g})(\mathbf{Y}^{\mathrm{log}}_{ig}-\bar{\mathbf{Y}}^{\mathrm{log}}_{g})} {\sqrt{\sum_i (\hat{\mathbf{Y}}^{\mathrm{log}}_{ig}-\bar{\hat{\mathbf{Y}}}^{\mathrm{log}}_{g})^2}\, \sqrt{\sum_i (\mathbf{Y}^{\mathrm{log}}_{ig}-\bar{\mathbf{Y}}^{\mathrm{log}}_{g})^2}}, \tag*{(33)} \end{equation}

where Y^ˉglog\bar{\hat{\mathbf{Y}}}^{\mathrm{log}}_{g} and Yˉglog\bar{\mathbf{Y}}^{\mathrm{log}}_{g} are averaged across all spots. We report PCC-M as the mean of PCCg\mathrm{PCC}_g over all valid target genes (G=250G=250 for HER2ST, Skin, and ST-Net), and PCC-H as the mean over the 50-gene HVG subset for these three datasets. For HEST-1k, PCC-M is averaged over the cohort-specific target panel described in Sec. B.

Relative variance distance (RVD). Because a high PCC in log-space does not by itself guarantee faithful expression recovery, we consider RVD, which measures how well the prediction preserves the per-gene heterogeneity of the data:

RVD=1G∑g=1G(Var(Y^:,glog)−Var(Y:,glog))2(Var(Y:,glog))2,\begin{equation} \mathrm{RVD}=\frac{1}{G}\sum_{g=1}^{G} \frac{\left(\mathrm{Var}(\hat{\mathbf{Y}}^{\mathrm{log}}_{:,g})-\mathrm{Var}(\mathbf{Y}^{\mathrm{log}}_{:,g})\right)^2} {\left(\mathrm{Var}(\mathbf{Y}^{\mathrm{log}}_{:,g})\right)^2}, \tag*{(34)} \end{equation}

where Var(⋅)\mathrm{Var}(\cdot) is the across-spot variance. Lower RVD indicates that the predicted expression reproduces the spatial variance of the ground truth more faithfully. For all correlation metrics higher is better; for MSE, MAE and RVD lower is better.

Absolute-error metrics.

MSE=1NG∑i,g(Y^iglog−Yiglog)2,\begin{equation} \mathrm{MSE}=\frac{1}{NG}\sum_{i,g}\left(\hat{\mathbf{Y}}^{\mathrm{log}}_{ig}-\mathbf{Y}^{\mathrm{log}}_{ig}\right)^2, \tag*{(35)} \end{equation}
MAE=1NG∑i,g∣Y^iglog−Yiglog∣.\begin{equation} \mathrm{MAE}=\frac{1}{NG}\sum_{i,g}\left\lvert \hat{\mathbf{Y}}^{\mathrm{log}}_{ig}-\mathbf{Y}^{\mathrm{log}}_{ig}\right\rvert . \tag*{(36)} \end{equation}

Figure 7. Distribution of fold-averaged per-gene Pearson correlation coefficients (PCC-H and PCC-M) across HER2ST [22], ST-Net [1] and Skin [30]. BCARP generally shifts the per-gene correlation distributions toward higher values across both the full target panel and the HVG subset.

Figure 8. Per-fold PCC-M performance across the three ST datasets.

Figure 9. Per-fold PCC-M performance across the ten HEST-1k cohorts. Dashed lines mark each method’s cohort mean.

DatasetMetric12345678910
HER2STPCC-H↑\uparrow+0.0+5.4+6.3+8.1+7.9+7.9+7.4+8.5+8.1+7.6
PCC-M↑\uparrow+0.0+5.7+6.1+7.8+7.5+7.5+7.4+8.4+7.7+7.4
MSE↓\downarrow+0.0-6.8-2.0-5.1-5.7-2.3-2.5-7.4-3.9-3.8
MAE↓\downarrow+0.0-4.2-0.7-1.9-2.5-1.0-1.4-3.7-1.6-1.7
RVD↓\downarrow+0.0-14.3+18.4-12.6+1.7+7.1+11.6-11.7+7.9+10.2
SkinPCC-H↑\uparrow+0.0+1.2+1.2+1.6+1.0+1.1+1.5+1.8+1.1+1.1
PCC-M↑\uparrow+0.0+0.7+0.6+0.7+0.2+0.2+1.1+0.8+0.3+0.3
MSE↓\downarrow+0.0-5.5+24.8+29.4+29.0+28.2+0.8+31.9+27.3+29.4
MAE↓\downarrow+0.0-3.9+9.9+13.0+12.7+12.3-0.5+14.4+11.5+13.0
RVD↓\downarrow+0.0+14.1-8.7+9.8+3.5+11.7+12.5+15.5-0.2+10.2
ST-NetPCC-H↑\uparrow+0.0-0.8-1.1-5.8+0.8+0.7-2.2-3.6-3.6-0.8
PCC-M↑\uparrow+0.0-1.2-1.4-4.9+0.3+0.1-2.7-3.4-3.7-1.0
MSE↓\downarrow+0.0-6.2-4.4-3.1-2.8-1.0-4.9-0.8-3.5-3.3
MAE↓\downarrow+0.0-3.3-2.3-1.5-1.5-0.7-2.6-0.5-2.1-1.8
RVD↓\downarrow+0.0-1.4-0.9+0.9-4.4-2.7-0.2-0.3-2.8-0.8
AveragePCC-H↑\uparrow+0.0+2.0+2.1+1.3+3.2+3.2+2.2+2.2+1.9+2.7
PCC-M↑\uparrow+0.0+1.8+1.8+1.2+2.7+2.6+1.9+1.9+1.4+2.2
MSE↓\downarrow+0.0-6.2+6.1+7.1+6.8+8.3-2.2+7.9+6.6+7.4
MAE↓\downarrow+0.0-3.8+2.3+3.2+2.9+3.5-1.5+3.4+2.6+3.2
RVD↓\downarrow+0.0-0.6+2.9-0.6+0.3+5.4+7.9+1.2+1.6+6.5

Table 5. Refinement-depth ablation as {relative change} (%) of each metric with respect to the anchor R ⁣= ⁣1R\!=\!1.

KctxK_{\mathrm{ctx}}PCC-H↑\uparrowPCC-M↑\uparrowMSE↓\downarrowMAE↓\downarrowRVD↓\downarrowMean Imp.↑\uparrow
00.45860.40610.85380.72820.81230.00%
50.4608 (++0.5%)0.4076 (++0.4%)0.8933 (++4.6%)0.7507 (++3.1%)0.7522 (−-7.4%)+0.11%
100.4509 (−-1.7%)0.3941 (−-3.0%)0.8807 (++3.2%)0.7435 (++2.1%)0.9480 (++16.7%)−-5.32%
250.4534 (−-1.1%)0.4043 (−-0.4%)0.9277 (++8.7%)0.7657 (++5.1%)0.8649 (++6.5%)−-4.37%
500.4567 (−-0.4%)0.4032 (−-0.7%)0.8381 (−-1.8%)0.7191 (−-1.2%)0.7084 (−-12.8%)+2.95%
1000.4374 (−-4.6%)0.3891 (−-4.2%)0.8903 (++4.3%)0.7537 (++3.5%)0.7987 (−-1.7%)−-2.98%
2000.4530 (−-1.2%)0.3999 (−-1.5%)0.8591 (++0.6%)0.7357 (++1.0%)0.8464 (++4.2%)−-1.72%

Table 6. Sensitivity to the number of global prototype tokens KctxK_{\mathrm{ctx}} on HER2ST. Mean Imp. denotes the mean sign-corrected relative improvement across the five metrics, where decreases in MSE, MAE, and RVD are counted as positive improvements.

Experimental Analysis

Per-gene correlation distributions.

Figure 7 compares the fold-averaged per-gene PCC distributions over the full panel and over the highly variable subset. Beyond shifting the bulk of genes toward higher correlation, BCARP also trims the left tail. This is visible on Skin [30], the only cohort whose distribution extends below zero, where a non-negligible fraction of genes have PCC below zero under TRIPLEX [23] and EGN [21] while almost all of BCARP’s genes remain positive. STFlow is the closest competitor throughout, and the gap to it is much smaller than the gap to TRIPLEX and EGN: on Skin the two distributions overlap substantially and BCARP’s advantage is a modest rightward shift of the mode rather than a separation. The same ordering holds on HER2ST [22], again with considerable overlap. The separation is cleanest on ST-Net [1], where the BCARP histogram is essentially disjoint from all three baselines, which concentrate between 0.080.08 and 0.160.16 while BCARP peaks near 0.250.25-0.300.30.

Per-fold stability.

Figure 8 and Figure 9 report per-fold PCC-M. On the three ST datasets, BCARP is highest in every fold, and the separation is most decisive on ST-Net [1], where it leads at all eight folds, by more than a factor of two at folds 0 and 7 and by a narrower margin elsewhere. The clearest exception is the HER2ST [22] folds corresponding to patients E and F (folds 4-5), where all methods drop sharply and BCARP’s margin narrows; this points to specific tissue morphologies that remain difficult for every model rather than to instability of BCARP itself. On the HEST-1k cohorts the picture is more balanced and we report it as such: BCARP and STFlow trade the lead fold by fold, with STFlow ahead on HCC, PAAD, READ 1 and LUNG 0, and BCARP ahead on LYMPH_IDC, PRAD and most CCRCC folds. On IDC all four methods are within a narrow band and BCARP is not the strongest. BCARP’s advantage on HEST-1k is therefore reflected in its aggregate cross-cohort performance rather than uniform dominance within individual cohorts, consistent with the mean-rank comparison in Table 4 of the main paper.

Computational cost.

Table tab:supp_cost reports three quantities measured under distinct conditions. Trainable parameters count only those updated by the optimiser and exclude the frozen encoder on which most methods depend. End-to-end latency covers the full inference path, encoder forward pass included, measured on a representative HER2ST section containing N=712N=712 spots. Peak memory is the maximum allocator high-water mark observed over a single training epoch across the evaluated datasets. For methods using frozen UNI embeddings, the shared encoder dominates end-to-end latency. Against the 1.28 s per-slide cost of UNI+Ridge, BCARP adds only 0.09 s, while the most expensive UNI-based heads add approximately 0.3 s. Since UNI embeddings can be extracted once and cached, this shared cost is amortized during training. M2OST does not use a separate frozen UNI encoder and therefore has the lowest measured end-to-end latency, but it incurs the largest peak memory footprint. BCARP remains within the same memory range as the other embedding-based heads while training fewer parameters than mclSTExp, TRIPLEX, and EGN.

Sensitivity to the number of global prototype tokens.

Table 6 varies KctxK_{\mathrm{ctx}}, taking Kctx=0K_{\mathrm{ctx}}=0, a prototype-free global stage that retains spot-level self-attention, as the reference. The response is not monotonic. Correlation peaks at a very small number of tokens: Kctx ⁣= ⁣5K_{\mathrm{ctx}}\!=\!5 gives the best PCC-H and PCC-M, and both are marginally above the prototype-free baseline, which indicates that the prototype-guided component contributes only modestly to correlation on HER2ST. Its contribution is concentrated instead in variance fidelity, where Kctx ⁣= ⁣50K_{\mathrm{ctx}}\!=\!50 reduces RVD by 12.8%12.8\% relative to the baseline, the largest single improvement in the table, while also being the only setting that improves both absolute-error metrics. Aggregated over the five metrics, Kctx ⁣= ⁣50K_{\mathrm{ctx}}\!=\!50 is the only configuration with a clearly positive mean improvement (+2.95%+2.95\%) and we adopt it as the default. We note that four of the six non-zero token budgets score below the prototype-free baseline, so the benefit of prototype-guided propagation is token-budget dependent rather than monotonic in KctxK_{\mathrm{ctx}}.

Sensitivity to the nuclei-analysis backend.

The neighborhood-composition and cell-type targets of Equation (31) depend on an off-the-shelf nuclei segmentation model. To check our conclusions, we regenerate both targets with CellViT [36] and retrain BCARP on HER2ST (Table tab:segmentation_ablation). All other settings are identical and only the pseudo-label generator changes. The two sources give closely comparable results, and the differences do not point uniformly in one direction: CellViT is slightly better on the correlation metrics and on RVD, while HoVer-Net [24] is better on both absolute-error metrics. The CellViT experiment is therefore used as a post-hoc robustness check rather than for model selection. Since the two backends yield comparable downstream performance without a consistent winner across metrics, we retain HoVer-Net for consistency with all main experiments. These results indicate that BCARP’s HER2ST performance is not highly sensitive to the choice between the two tested nuclei-analysis backends, although broader cross-tissue robustness remains to be studied.

DatasetFoldTest partitionTrain slidesTest slidesTrain spots / Test spots
ST datasets
HER2ST0A30611,555 / 2,065
1B30611,908 / 1,712
2C30612,534 / 1,086
3D30611,787 / 1,833
4E33311,891 / 1,729
5F33311,522 / 2,098
6G33312,249 / 1,371
7H33311,894 / 1,726
Skin0P10_rep936,980 / 1,691
1P2_rep936,721 / 1,950
2P5_rep937,039 / 1,632
3P9_rep935,273 / 3,398
ST-Net0BC23209 C/D, BC23377 C/D, BC23901 C/D60827,215 / 3,412
1BC23268 C/D, BC23450 D/E, BC23903 C/D59926,896 / 3,731
2BC23269 C/D, BC23506 C/D, BC23944 D/E59926,616 / 4,011
3BC23270 D/E, BC23508 D/E, BC24044 D/E59926,598 / 4,029
4BC23272 D/E, BC23567 D/E, BC24105 C/D59926,248 / 4,379
5BC23277 D/E, BC23803 D/E, BC24220 D/E59926,587 / 4,040
6BC23287 C/D, BC23810 D/E, BC24223 D/E59926,490 / 4,137
7BC23288 D/E, BC23895 C/D62627,739 / 2,888
HEST-1k cohorts
SKCM0TENX117111,741 / 1,293
1TENX115111,293 / 1,741
LUNG0TENX141111,944 / 3,262
1TENX118113,262 / 1,944
PAAD0TENX116215,561 / 2,010
1TENX140213,961 / 3,610
2TENX126215,620 / 1,951
READ0ZEN48, ZEN49223,819 / 4,588
1ZEN36, ZEN40224,588 / 3,819
COAD0TENX147-TENX149135,079 / 10,572
1TENX1113110,572 / 5,079
LYMPH_IDC0NCBI6843114,972 / 4,992
1NCBI6833114,972 / 4,992
2NCBI6823114,976 / 4,988
3NCBI6813114,972 / 4,992
CCRCC0INT13-INT15, INT2420455,849 / 18,371
1INT9-INT1220467,651 / 6,569
2INT6-INT8, INT2320463,464 / 10,756
3INT2-INT520467,098 / 7,122
4INT1, INT20-INT2220459,472 / 14,748
5INT16-INT1920457,566 / 16,654
IDC0TENX993114,775 / 20,761
1TENX953127,776 / 7,760
2NCBI7853131,526 / 4,010
3NCBI7833132,531 / 3,005
PRAD0MEND154, MEND156-MEND16215836,819 / 25,891
1MEND139-MEND15381525,891 / 36,819
HCC0NCBI643111,941 / 2,255
1NCBI642112,255 / 1,941

Table 7. Detailed cross-validation splits by fold. For HER2ST, Skin and ST-Net, test partitions are identified by patient; all sections from a patient are assigned exclusively to the training or the test partition. For the HEST-1k cohorts we adopt the official patient-grouped folds released with the benchmark and identify test partitions by HEST slide ID.

Figure 10. Qualitative comparison of predicted expression for three marker genes on Skin sample P2_rep3.

Figure 11. Qualitative comparison of predicted expression for three marker genes on Skin sample P5_rep2.

Figure 12. Qualitative comparison of predicted expression for three marker genes on Skin sample P10_rep2

References

  1. [1]He, B., Bergenstråhle, L., Stenbeck, L., Abid, A., Andersson, A., Borg, Å., Maaskola, J., Lundeberg, J., & Zou, J. (2020). Integrating spatial gene expression and breast tumour morphology via deep learning. Nature Biomedical Engineering, 4(8), 827–834.
  2. [2]Jaume, G., Doucet, P., Song, A. H., Lu, M. Y., Almagro-Perez, C., Wagner, S. J., Vaidya, A. J., Chen, R. J., Williamson, D. F. K., Kim, A., & Mahmood, F. (2024). HEST-1k: A Dataset for Spatial Transcriptomics and Histology Image Analysis. Advances in Neural Information Processing Systems.
  3. [3]Jia, Y., Liu, J., Chen, L., Zhao, T., & Wang, Y. (2023). THItoGene: a deep learning method for predicting spatial transcriptomics from histological images. Briefings in Bioinformatics, 25(1).
  4. [4]Qu, M., Wu, Y., Di, D., Su, A., Su, T., Song, Y., & Fan, L. (2024). Boundary-Guided Learning for Gene Expression Prediction in Spatial Transcriptomics. 2024 IEEE International Conference on Bioinformatics and Biomedicine (BIBM), 445–450.
  5. [5]Nguyen, H. D., Pham, N. D. H., Nguyen, T. M. D., Nguyen, D. T., Nguyen, H. T., & Nguyen, D. M. (2026). MMAP: A Multi-Magnification and Prototype-Aware Architecture for Predicting Spatial Gene Expression. In PRICAI 2025: Trends in Artificial Intelligence (pp. 381–395). Springer Nature Singapore.
  6. [6]Zhu, J., Deng, R., Yao, T., Xiong, J., Qu, C., Guo, J., Lu, S., Yin, M., Wang, Y., Zhao, S., et al. (2025). Asign: an anatomy-aware spatial imputation graphic network for 3d spatial transcriptomics. Proceedings of the Computer Vision and Pattern Recognition Conference, 30829–30838.
  7. [7]Zhang, P., Gao, C., Zhang, Z., Yuan, Z., Zhang, Q., Zhang, P., Du, S., Zhou, W., Li, Y., & Li, S. (2025). Systematic inference of super-resolution cell spatial profiles from histology images. Nature Communications, 16(1).
  8. [8]Chan, T. H., Cendra, F. J., Ma, L., Yin, G., & Yu, L. (2023). Histopathology Whole Slide Image Analysis with Heterogeneous Graph Representation Learning. 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 15661–15670.
  9. [9]Wang, X., Yang, S., Zhang, J., Wang, M., Zhang, J., Yang, W., Huang, J., & Han, X. (2022). Transformer-based unsupervised contrastive learning for histopathological image classification. Medical Image Analysis, 81, 102559.
  10. [10]Filiot, A., Jacob, P., Mac Kain, A., & Saillard, C. (2024). Phikon-v2, a large and public feature extractor for biomarker prediction. https://arxiv.org/abs/2409.09173
  11. [11]Zimmermann, E., Vorontsov, E., Viret, J., Casson, A., Zelechowski, M., Shaikovski, G., Tenenholtz, N., Hall, J., Klimstra, D., Yousfi, R., et al. (2024). Virchow2: Scaling self-supervised mixed magnification models in pathology. https://arxiv.org/abs/2408.00738
  12. [12]Song, A. H., Chen, R. J., Ding, T., Williamson, D. F. K., Jaume, G., & Mahmood, F. (2024). Morphological Prototyping for Unsupervised Slide Representation Learning in Computational Pathology. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 11566–11578.
  13. [13]Zhang, Y., Liu, S., Chen, J., Chen, R., Yang, Z., Sheng, R., Li, X., Wang, T., Liu, H., Yang, F., et al. (2025). Deep learning-based histomorphological subtyping and risk stratification of small cell lung cancer from hematoxylin and eosin-stained whole slide images. Genome Medicine, 17(1), 98.
  14. [14]Pentimalli, T. M., Karaiskos, N., & Rajewsky, N. (2025). Challenges and opportunities in the clinical translation of high-resolution spatial transcriptomics. Annual Review of Pathology: Mechanisms of Disease, 20(1), 405–432.
  15. [15]Kunhoth, S., Al-maadeed, S., Akbari, Y., & Al Saady, R. M. (2026). Computational methods for breast cancer molecular profiling using routine histopathology: a review. Archives of Computational Methods in Engineering, 33(2), 2299–2321.
  16. [16]Jain, S., & Eadon, M. T. (2024). Spatial transcriptomics in health and disease. Nature Reviews Nephrology, 20(10), 659–671.
  17. [17]Gulati, G. S., D’Silva, J. P., Liu, Y., Wang, L., & Newman, A. M. (2025). Profiling cell identity and tissue architecture with single-cell and spatial transcriptomics. Nature Reviews Molecular Cell Biology, 26(1), 11–31.
  18. [18]Williams, C. G., Lee, H. J., Asatsuma, T., Vento-Tormo, R., & Haque, A. (2022). An introduction to spatial transcriptomics for biomedical research. Genome Medicine, 14(1), 68.
  19. [19]Pang, M., Su, K., & Li, M. (2021). Leveraging information in spatial transcriptomics to predict super-resolution gene expression from histology images in tumors. bioRxiv.
  20. [20]Xie, R., Pang, K., Chung, S., Perciani, C., MacParland, S., Wang, B., & Bader, G. (2023). Spatially Resolved Gene Expression Prediction from Histology Images via Bi-modal Contrastive Learning. Advances in Neural Information Processing Systems, 36.
  21. [21]Yang, Y., Hossain, M. Z., Stone, E. A., & Rahman, S. (2023). Exemplar Guided Deep Neural Network for Spatial Transcriptomics Analysis of Gene Expression Prediction. 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 5028–5037.
  22. [22]Andersson, A., Larsson, L., Stenbeck, L., Salmén, F., Ehinger, A., Wu, S. Z., Al-Eryani, G., Roden, D., Swarbrick, A., Borg, Å., et al. (2021). Spatial deconvolution of HER2-positive breast cancer delineates tumor-associated cell type interactions. Nature Communications, 12(1), 6012.
  23. [23]Chung, Y., Ha, J. H., Im, K. C., & Lee, J. S. (2024). Accurate Spatial Gene Expression Prediction by Integrating Multi-Resolution Features. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 11591–11601.
  24. [24]Graham, S., Vu, Q. D., Raza, S. E. A., Azam, A., Tsang, Y. W., Kwak, J. T., & Rajpoot, N. (2019). HoVer-Net: Simultaneous Segmentation and Classification of Nuclei in Multi-Tissue Histology Images. Medical Image Analysis, 58, 101563.
  25. [25]Chen, R. J., Ding, T., Lu, M. Y., Williamson, D. F. K., Jaume, G., Song, A. H., Chen, B., Zhang, A., Shao, D., Shaban, M., Williams, M., Oldenburg, L., Weishaupt, L. L., Wang, J. J., Vaidya, A., Le, L. P., Gerber, G., Sahai, S., Williams, W., & Mahmood, F. (2024). Towards a general-purpose foundation model for computational pathology. Nature Medicine, 30(3), 850–862.
  26. [26]Lu, M. Y., Chen, B., Williamson, D. F. K., Chen, R. J., Liang, I., Ding, T., Jaume, G., Odintsov, I., Le, L. P., Gerber, G., Parwani, A. V., Zhang, A., & Mahmood, F. (2024). A visual-language foundation model for computational pathology. Nature Medicine, 30(3), 863–874.
  27. [27]Fu, X., Cao, Y., Bian, B., Wang, C., Graham, D., Pathmanathan, N., Patrick, E., Kim, J., & Yang, J. Y. H. (2025). Spatial gene expression at single-cell resolution from histology using deep learning with GHIST. Nature Methods, 22(9), 1900–1910.
  28. [28]Ganguly, A., Chatterjee, D., Huang, W., Zhang, J., Yurovsky, A., Johnson, T. S., & Chen, C. (2025). MERGE: Multi-faceted Hierarchical Graph-based GNN for Gene Expression Prediction from Whole Slide Histopathology Images. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 15611–15620.
  29. [29]Liu, Y., Wang, T., Duggan, B., Sharpnack, M., Huang, K., Zhang, J., Ye, X., & Johnson, T. S. (2022). SPCS: a spatial and pattern combined smoothing method for spatial transcriptomic expression. Briefings in Bioinformatics, 23(3), bbac116.
  30. [30]Ji, A., Rubin, A., Thrane, K., Jiang, S., Reynolds, D., et al. (2020). Multimodal Analysis of Composition and Spatial Architecture in Human Squamous Cell Carcinoma. Cell, 182.
  31. [31]Huang, T., Liu, T., Babadi, M., Jin, W., & Ying, R. (2025). Scalable Generation of Spatial Transcriptomics from Histology Images via Whole-Slide Flow Matching. Forty-second International Conference on Machine Learning.
  32. [32]Wang, H., Du, X., Liu, J., Ouyang, S., Chen, Y.-W., & Lin, L. (2025). M2OST: Many-to-one Regression for Predicting Spatial Transcriptomics from Digital Pathology Images. Proceedings of the AAAI Conference on Artificial Intelligence, 39(7), 7709–7717.
  33. [33]Min, W., Shi, Z., Zhang, J., Wan, J., & Wang, C. (2024). Multimodal contrastive learning for spatial gene expression prediction using histology images. Briefings in Bioinformatics, 25(6), bbae551.
  34. [34]Xu, H., Usuyama, N., Bagga, J., Zhang, S., Rao, R., Naumann, T., Wong, C., Gero, Z., González, J., Gu, Y., Xu, Y., Wei, M., Wang, W., Ma, S., Wei, F., Yang, J., Li, C., Gao, J., Rosemon, J., … Poon, H. (2024). A whole-slide foundation model for digital pathology from real-world data. Nature, 630(8015), 181–188.
  35. [35]He, K., Zhang, X., Ren, S., & Sun, J. (2016). Deep Residual Learning for Image Recognition. 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 770–778.
  36. [36]Hörst, F., Rempe, M., Heine, L., Seibold, C., Keyl, J., Baldini, G., Ugurel, S., Siveke, J., Grünwald, B., Egger, J., & Kleesiek, J. (2024). CellViT: Vision Transformers for precise cell segmentation and classification. Medical Image Analysis, 94, 103143. https://doi.org/10.1016/j.media.2024.103143
  37. [37]Chen, W., Zhang, P., Tran, T. N., Xiao, Y., Li, S., Shah, V. V., Cheng, H., Brannan, K. W., Youker, K., Lai, L., et al. (2025). A visual--omics foundation model to bridge histopathology with spatial transcriptomics. Nature Methods, 1–15.
  38. [38]Gurarslan, D., Camargo, O., Zeyveli, O., Almalioglu, Y., Li, Y., Turan, M., & Kahveci, T. (2026). STORM: spatial transcriptomics optimization by resolution via matrix factorization. Briefings in Bioinformatics, 27(3).

Paper details

vlnc:W_wb9q4rf4
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 4

downloads · 0

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

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents