Path2Omics

Molecular profiling increasingly drives treatment decisions in oncology, but bulk RNA sequencing and methylation arrays are expensive, slow, and frequently impossible on the small biopsies that reach a routine pathology service. A hematoxylin-and-eosin-stained section, by contrast, is generated for essentially every resected solid tumor. Path2Omics, developed by the Cancer Data Science Laboratory at the National Cancer Institute, narrows that gap by inferring bulk gene expression and DNA methylation directly from the digitized slide.

Its central design decision concerns tissue preparation. Formalin-fixed, paraffin-embedded (FFPE) sections are what clinical practice produces, but fresh-frozen (FF) sections preserve nucleic acids far better and therefore carry cleaner molecular labels for training. Instead of choosing one, the authors trained a separate model on each preparation and combined their outputs at inference, so a query FFPE slide benefits from signal learned on frozen material. Across 23 cancer cohorts from The Cancer Genome Atlas, this integrated predictor recovers more genes accurately than either component model on its own.

Path2Omics succeeds the same group's earlier DeepPT model. It was posted as a preprint in February 2025 and published in Cancer Research later that year.

#Key Features

• Two omics layers from one image: The same pipeline predicts transcript abundance and CpG-level DNA methylation, giving both a functional and an epigenetic readout from a single stained section.
• FFPE and fresh-frozen integration: Two preparation-specific models are trained independently and their predictions merged, letting the cleaner frozen-tissue signal improve inference on the FFPE slides clinicians actually have.
• Pan-cancer training, single checkpoint: One model is trained across all 23 cohorts rather than refit per tumor type, and the same trained model is applied unchanged to every evaluation cohort.
• External validation without retraining: Seven independent cohorts spanning glioma, breast, kidney, lung, and colorectal cancer were scored with the TCGA-trained model, testing transfer across scanners, laboratories, and staining protocols.
• Clinically anchored downstream evaluation: Rather than stopping at correlation, the authors show that inferred expression supports survival modeling and chemotherapy response prediction close to what measured expression achieves.

#Technical Details

The pipeline has three stages. Each whole-slide image is cut into 512 x 512 pixel tiles at 20x magnification, Sobel edge detection discards background tiles, and Macenko normalization reduces staining variation. A frozen CTransPath pathology encoder then converts each retained tile into a 768-dimensional feature vector; this is the same frozen-encoder-plus-lightweight-head recipe used to adapt slide-level foundation models such as UNI and Prov-GigaPath to downstream tasks. Prediction is handled by a three-layer perceptron: a 768-node input layer, a 512-node hidden layer, and an output layer with one node per target gene or CpG probe. Training used 20,497 slides — 9,456 FFPE and 11,041 FF — from 8,007 patients.

Counting genes predicted at Pearson correlation above 0.4, the FFPE model reached an average of 3,503 genes on FFPE slides and the FF model 7,299 genes on FF slides, while the integrated model reached 5,417 genes on FFPE slides. Methylation followed the same pattern: 9,223, 13,194, and 12,272 well-predicted CpG probes respectively. On the 15 cancer types shared with DeepPT, the integrated model reached 4,903 well-predicted genes, roughly five times its predecessor. External evaluation covered 1,323 slides across NCI-LGG, five CPTAC cohorts, and TransNeo-Breast, averaging 4,391 well-predicted genes — a 30% gain over the FFPE model alone. Inferred expression yielded a mean high C-index of 0.69 for survival across 12 cohorts, and predicted chemotherapy response at AUC 0.82 against 0.87 for measured expression.

#Applications

The immediate use is molecular triage in settings where sequencing is unavailable, unaffordable, or precluded by tissue quantity: estimating expression-based prognostic signatures, screening cohorts for candidates worth confirmatory assays, or retrospectively enriching archival slide collections that were never profiled. Because inference runs on scanned FFPE slides already produced for diagnosis, it fits existing digital pathology workflows without additional tissue. Researchers also gain a way to attach molecular labels to large historical image cohorts for epidemiological and outcomes studies.

#Impact

Path2Omics extends a growing line of work turning routine histology into a molecular proxy, and its contribution is methodological as much as empirical: it shows that fresh-frozen material, usually discarded as clinically irrelevant, is a useful training resource for models deployed on FFPE. The seven-cohort external evaluation is the strongest evidence offered, since transfer across institutions is where slide-based predictors most often fail. Limits are real. Predictions are cohort-level correlations, not measurements, and accuracy varies widely by gene and by tumor type. The public release provides code under the MIT license together with extracted tile features, metadata, and training scripts on Zenodo, but no trained checkpoint, so reproducing the model requires rerunning training; the accompanying documentation further restricts model use to academic research.