PathGen

Fusing a digitized H&E slide with bulk RNA measurements demonstrably improves cancer grading and survival risk estimation — and almost never happens in practice. Histopathology is the diagnostic gold standard and is collected for essentially every resected tumor, while transcriptomic assays are expensive enough that public health systems restrict them to a small minority of patients. A multimodal model that requires both modalities is therefore untranslatable in exactly the settings that would benefit most.

PathGen closes that gap by generating the missing modality rather than demanding it. It is a conditional diffusion model whose denoiser is conditioned on whole-slide-image patch embeddings: starting from Gaussian noise in gene-expression space, it iteratively denoises toward a transcriptomic vector for the patient, with gene embeddings attending to the slide's patch embeddings through a genomic-guided co-attention block so that the sampled expression stays tied to the tissue actually on the slide. The synthesized transcriptome is then consumed by a downstream grading-and-survival head exactly as a real one would be. The framing matters: the authors position this as a screening step to decide which patients justify the cost of real sequencing, not as a substitute for it.

The model was developed at the Indian Association for the Cultivation of Science with the Alan Turing Institute and clinical collaborators in London, first posted in February 2025 under the name PathoGen and published in Nature Communications as PathGen. It is unrelated to the similarly named PathGen-1.6M pathology vision-language dataset family. Its closest catalog analogue is Path2Omics, which regresses expression and methylation directly rather than sampling them generatively.

#Key Features

• Generative rather than regressive: expression is sampled through a 1,000-step denoising trajectory, producing a transcriptomic vector drawn from a learned distribution instead of a single conditional mean.
• Genomic-guided co-attention: inside the denoiser, gene-group embeddings co-attend to patch embeddings three times over, borrowing the co-attention formulation from MCAT to keep generated genes anchored to slide morphology.
• Six curated gene groups: outputs cover tumor suppressors, oncogenes, protein kinases, cell differentiation markers, transcription factors, and cytokines/growth factors — the signature sets multimodal prognostic models already use.
• Cross-dataset transfer: checkpoints trained on TCGA cohorts were applied to the corresponding CPTAC cohorts with no further training or fine-tuning.
• Calibrated, inspectable downstream predictions: conformal prediction supplies patient-level grade sets and survival bounds with a coverage guarantee, and co-attention maps project distributed grade and risk back onto the slide.

#Technical Details

Slides are tiled into 224 × 224 patches at a randomly chosen magnification, background-filtered, and embedded to 1,024 dimensions with the frozen UNI pathology encoder; no stain normalization is applied. A four-layer gene encoder with ELU activations lifts each gene group's expression levels to a matching 1,024-dimensional embedding, and the PathGen transformer — co-attention plus transformer encoder layers — predicts the noise at each step, with per-gene-group decoders mapping back to expression levels. Training used T = 1,000 timesteps at a learning rate of 1 × 10⁻⁴ on an NVIDIA A100; synthesis takes about 21 seconds per slide.

Four TCGA cohorts were used: glioma-glioblastoma (912 WSIs, 745 cases), renal clear cell (485/462), uterine endometrial (294/267), and breast (1,010/946). Spearman correlation between synthesized and real expression reached 0.713, 0.717, 0.436 and 0.642 respectively, with normalized MAE of 0.141–0.173. On the external CPTAC cohorts the same checkpoints gave 0.662 (GBM) and 0.669 (UCEC). Downstream, adding synthesized transcriptomes to slide features lifted TCGA-GBMLGG from AUC 0.823 / C-index 0.842 to 0.890 / 0.861, statistically indistinguishable from the 0.907 / 0.866 obtained with real expression. Performance is uneven across cohorts: CPTAC-UCEC reached only AUC 0.593 and C-index 0.530.

#Applications

The intended use is triage. A pathology service holding slides but no sequencing budget can run PathGen to estimate whether transcriptomic data would meaningfully change a patient's grade or risk assessment, and spend the assay only where the conformal bounds say it matters. The same route makes retrospective slide archives without matched RNA-seq usable for multimodal prognostic modeling, and the co-attention heat maps give pathologists a way to check that the model's attention falls on regions they would consider relevant.

#Impact

PathGen reframes crossmodal generation as a cost-allocation tool rather than a claim to replace molecular testing, which is a more defensible position than the literature usually takes. Its limits are correspondingly concrete: the six gene groups do not span the transcriptome, checkpoints are cohort-specific with transfer shown only from a TCGA cohort to the same cancer type in CPTAC, and the weakest cohort's numbers are close to chance. Code and trained checkpoints are released, though under a non-commercial CC-BY-NC-ND license and with weights distributed through cloud-drive folders rather than a versioned registry.