BioMorphNet: Multimodal Tissue Analysis

• BioMorphNet is a multimodal deep learning framework that integrates tissue morphology and spatial transcriptomics to classify cancer tissues at a patch-level with high precision.
• It utilizes a three-branch architecture combining a morphological graph encoder with a spatial transcriptomic MLP and pathway fusion via transformer-based cross-attention.
• Experimental results demonstrate state-of-the-art performance with balanced accuracies up to 81.9% and AUROCs exceeding 0.90 across multiple cancer datasets.

BioMorphNet is a multimodal deep learning architecture designed to integrate tissue morphological features and spatial gene expression for patch-level tissue classification and biomarker discovery in whole-slide images (WSIs) with spatial transcriptomics. It combines graph convolutional modeling of the tumor microenvironment (TME), clinical pathway and learnable pathway modules, and transformer-based multimodal fusion to achieve state-of-the-art performance in cancer tissue analysis and spatial genomics-supported biomarker identification (Liu et al., 13 Jan 2026).

1. Architectural Framework and Multimodal Integration

BioMorphNet operates at the patch level, representing each 224×224 px region (patch) of a WSI as a separate data point. The model is organized into three parallel branches:

• Morphological Graph Encoder: Each patch and its $k=8$ spatially-adjacent neighbors constitute the nodes of a TME graph. Node features ($x^c$ for the central patch, $x_i$ for neighbors) are 1024-dimensional embeddings extracted using a pretrained histology convolutional neural network (UNI v1). Edge weights $e_i$ are determined by a combined inverse Mean Squared Error (MSE) of both morphological and molecular feature similarity:

$$r_i^{\text{morph}} = \frac{1}{\text{MSE}(x^c, x_i)}, \quad r_i^{\text{bio}} = \frac{1}{\text{MSE}(g^c, g_i)}, \quad e_i = \frac{r_i^{\text{morph}} + r_i^{\text{bio}}}{2}.$$

The adjacency matrix $\mathbf{A}$ encodes these edge weights.

A two-layer Graph Convolutional Network (GCN) propagates information according to:

$$H^{(l+1)} = \sigma\left(\tilde D^{-1/2}\tilde A\,\tilde D^{-1/2} H^{(l)} W^{(l)}\right),\quad \tilde A = A + I$$

yielding a 512-dimensional morphological embedding $h$ for each central node.

• Spatial Transcriptomic Branch: Extracts a 512-dimensional embedding from each patch’s gene expression profile (20k–22k genes) via a multi-layer perceptron (MLP).
• Pathway Fusion Branch: Uses both predefined clinical pathway aggregations and a learnable pathway module.
  • Clinical Pathways: For a set of pathways $P_j \subseteq \{1,\dots,d\}$, computes activation $z_j = \sum_{p \in P_j} G_p$ and selects pathways with $\mathrm{Overlap} \geq 0.9$.
  • Learnable Pathways: Introduces trainable pathways via weight matrix $W \in \mathbb R^{a \times d}$ ($a\sim200$). Each pathway is masked to the top 5% weights, softmax-normalized, and applied to the gene vector, giving $z'_i = \sum_{j=1}^d a_{i,j} G_j$.

Fusion between morphological and pathway spaces is achieved by cross-attention transformer blocks, where morphological embeddings serve as queries/keys and pathway features as values. Gating weights $\alpha_1,\alpha_2,\alpha_3$ are adaptively assigned to the morphology, morphopathway, and gene branches, followed by a linear-softmax classifier.

2. Optimization, Loss, and Training Regimen

The primary objective is a weighted cross-entropy loss over tissue categories, with class-imbalance handled by

$w_i = \frac{N}{C N_i}$

where $N_i$ is the count of class $i$ samples and $C$ is the total number of classes. The softmax output $\hat{y}_i$ and ground-truth $y_i$ yield the loss:

$$\mathcal L_{\rm cls} = -\sum_{i=1}^C w_i y_i \log \hat{y}_i.$$

The total loss includes weight decay:

$$\mathcal L = \mathcal L_{\rm cls} + \lambda_{\rm WD}\|\Theta\|^2.$$

Training uses AdamW (weight decay $1\times10^{-4}$, learning rate $1\times10^{-4}$, batch size 32), up to 60 epochs, with early stopping on validation accuracy, repeated over five random splits per dataset.

3. Experimental Evaluation and Quantitative Results

BioMorphNet was evaluated on three cancer WSI datasets with paired transcriptomics:

Dataset	WSIs	Classes	Genes (approx.)	Patches	Balanced Acc (%)	AUROC
Prostate	7	4	20k	20,579	80.1 (+2.67)	0.97
Colorectal	6	3	22k	10,000	75.2 (+5.48)	0.932
Breast	8	2	12k	3,000	81.9 (+6.29)	0.918

All figures indicate improvements over the strongest state-of-the-art multimodal fusion baselines. AUROC consistently exceeded 0.90. Confusion matrices demonstrated >90% recall for key tumor grades (e.g., 90.4% correct for Gleason grade 4). Statistical significance was established (paired $t$-tests, $p < 0.05$) for gains across five random cross-validation splits.

4. Biomarker Discovery, Biological Insights, and Tumor Microenvironment Modeling

BioMorphNet supports high-confidence differential gene expression analysis by collecting patches with classifier confidence $\max \hat{y}_i \geq 0.95$ and conducting Wilcoxon rank-sum tests across predicted categories.

Key results include:

• Prostate: PPFIA2 (upregulated in high-grade lesions), MT1G (tumor suppressor, downregulated in tumor/stroma).
• Breast: DDX5, CD24, ERBB2 (established breast cancer markers).
• Colorectal: ITLN1, PLA2G2A (downregulated in tumors, associated with anti-angiogenic function).

Graph-based modeling of the TME leverages both morphology and gene similarity to modulate neighborhood influence, enabling more accurate context-dependent tissue classification and underpinning improved biomarker identification.

5. Pathway Modules: Clinical and Learnable Mechanisms

BioMorphNet’s dual-pathway strategy:

• Clinical pathway module maps gene expression to known biological pathways (e.g., signaling, metabolic processes), acting as a prior-encoded intermediate.
• Learnable pathway module discovers novel gene sets that may correspond to previously unannotated or context-specific signaling assemblies, using a sparsity-inducing mask (top 5%), softmax normalization, and MLP transformation.

Morphology–pathway fusion with transformer-based cross-attention aligns morphological and pathway codes, producing a morphopathway embedding refined by two transformer layers (8 heads each, dropout 0.25). A gating mechanism adaptively assigns fusion weights, optimizing representation for final classification.

6. Pareto Optimality and Network Morphogenesis: Theoretical Connections

Synthesizing principles from stochastic morphogenesis (Lucas et al., 7 Jan 2026), a minimal model based on local branching, fusion, and stopping rules can produce network morphologies that are Pareto-optimal with respect to coverage, transport, and robustness. Lucas et al. demonstrate that varying the branching probability ($\lambda$) and stopping probability ($p$) generates a spectrum from sparse tree-like (low $\lambda$, high $p$) to dense, loopy (high $\lambda$, low $p$) architectures, corresponding to real-world fungal networks.

BioMorphNet, by explicitly incorporating graph-based representation of tissue morphologies and adaptively fusing such features with biological pathway information, can be contextualized within this framework of multi-objective optimality. A plausible implication is that the hybrid tree–loop network topologies observed as optimal in biological transport systems inform the architectural biases that underpin BioMorphNet’s superior task performance.

7. Limitations and Future Research Directions

Identified limitations include:

• Dependence on supervised learning, leading to suboptimal classification for rare classes due to insufficient data (e.g., underrepresented tumor grades).
• Incomplete spatial omics, especially sparse transcriptome sampling in some regions, potentially impacts edge-weighting and neighborhood effects.
• Pathway knowledge base incompleteness, omitting emerging or context-specific molecular networks.

Proposed directions:

• Data augmentation via synthetic patches or self-supervised pretraining is suggested to ameliorate class imbalance.
• Expansion to multi-resolution graphs can better capture both local and global spatial interactions.
• Integration of unsupervised pathway discovery (e.g., community detection in gene co-expression graphs) to complement current modules.
• Application to larger-scale spatial omics datasets; extension toward outcome prediction tasks such as survival or treatment response.

BioMorphNet thus constitutes a generalizable, interpretable, and rigorously evaluated architecture at the intersection of computational pathology, network science, and spatial omics, demonstrating methodological and conceptual connections to the principles of biological network morphogenesis (Liu et al., 13 Jan 2026, Lucas et al., 7 Jan 2026).