Abstract PR-03 Inferring genomic properties and histologic subtypes of solid tumors from H&E whole-slide images
Bibliographic record
Abstract
Abstract Background: The impact of somatic mutations on tumor morphology remains incompletely understood. In investigating this, prior works have aggregated subtypes with different prior probabilities of genomic features (e.g., identifying TP53 oncogenic variants in “ovarian cancer” amounts to identifying high-grade serous histology), confounding the analysis. Methods: We developed Paladin, a collection of machine learning tools to infer genomic properties from hematoxylin and eosin (H&E)-stained whole-slide images (WSIs). We assembled a pan-cancer cohort of 72,725 sequenced tumor samples matched to 379,937 H&E WSIs from 63,932 patients, the largest to date. Using these data, we developed transformer models to infer 172 OncoKB-annotated genomic biomarkers of treatment response across 52 histologies using the most granular histology available. External validation was performed on the TCGA. We also developed and validated a companion model, AEON (adaptive embedding ontology network) to infer the most likely cancer subtypes, formalized as OncoTree codes, directly from H&E WSIs using an ontology-smoothed loss function based on knowledge graphs. Results: As part of Paladin, we trained over 4,000 models to mine phenotype-genotype associations, finding that 5% of tested target-histology pairs were strongly associated (test AUROC ≥ 0.8). We highlight two examples. First, FGFR3 mutational status, identified with internal test AUROC 0.90 (95% CI: 0.85 - 0.94, external test 0.89 (95% CI 0.84-0.94), prior state of the art (SOTA) 0.73) in urothelial carcinoma (BLCA), indicates erdafitinib. BLCA specimens with FGFR3 variants have a well described-phenotype, including papillary architecture and raisinoid nuclei. Second, STK11 mutational status, inferred with test AUROC 0.92 (95% CI: 0.89 - 0.95, external test 0.88 (95% CI 0.83-0.91), prior SOTA 0.62) in lung adenocarcinoma (LUAD), indicates a poor prognosis (log-rank p < 0.01 for overall survival (OS) by Paladin-inferred STK11 status). Turning to inference of OncoTree codes from H&E WSIs, we found that AEON attained a micro average one-vs-one AUROC of 0.99 with 69% top-1 accuracy (c.f. GDD-ENS, a genomic classifier, which attained AUROC 0.96 and 65% top-1 accuracy). Cancers of unknown primary (CUPs) were reclassified using AEON, recapitulating expected subtype-specific genomic feature enrichment (Fisher’s exact q ≥ 0.01 except KEAP1 and STK11 in LUAD and KRAS in pancreatic adenocarcinoma) and overall survival for patients with metastatic specimens of that histology (log-rank p ≥ 0.01 in 11/15 AEON-reclassified subtypes of CUP compared to ground-truth metastatic specimens of same histology). Conclusions: Our work uncovers phenotype-genotype relationships for granular subtypes which have been previously lumped together. Such approaches advance our knowledge of how somatic mutations affect tissue morphology and will improve access to precision oncology. AEON performs well in identifying cancer subtypes, supporting the role of machine learning on H&E images to aid with difficult diagnoses and refine subtyping of CUPs. Citation Format: Kevin M. Boehm, Madison Darmofal, Arfath Pasha, Andrew Aukerman, Armaan Kohli, Raymond Lim, Tom Pollard, Darin Moore, Anika Begum, Natasha Rekhtman, Hikmat Al-Ahmadie, Jason Chang, Klaus Busam, Daniel Gomez, Nancy Lee, Luke R.G. Pike, Himanshu Nagar, James Janopaul-Naylor, Lior Z. Braunstein, Justin Jee, Nikolaus Schultz, Sohrab P. Shah, Francisco Sanchez-Vega. Inferring genomic properties and histologic subtypes of solid tumors from H&E whole-slide images [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr PR-03.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.003 |
| Meta-epidemiology (narrow) | 0.001 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.000 |
| Open science | 0.001 | 0.001 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.005 | 0.002 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".