Foundation-model-guided radiogenomic discovery linking cancer genomes to cancer scans

Frederik Hauke, Jeremias Krause, Patrick Wienholt, Christiane Kuhl, Ingo Kurth, Sikander Hayat, Jakob Nikolas Kather, Sven Nebelung, Daniel TruhnarXiv 2026 · arXiv

TL;DR

This paper introduces a zero-shot, hypothesis-free radiogenomics framework that pairs the genomic foundation model Evo 2 with routine clinical imaging to uncover gene-phenotype associations at a genome-wide scale. By correlating AI-predicted mutation severity with radiomic features from tumor scans, the method successfully recovers known cancer drivers and discovers 46 novel imaging-associated genes linked to structural cell biology.

What Problem Does This Paper Address?

Traditional methods for identifying cancer-driving genes rely heavily on recurrence statistics—how often a gene is mutated across a large population. This approach requires massive patient cohorts and completely misses the "long tail" of rarely mutated genes whose functional impacts remain unknown. Furthermore, understanding how these mutations manifest physically in a tumor typically requires expensive, invasive experimental follow-ups.

The authors address this bottleneck by asking: Can we use a general-purpose genomic language model combined with standard, non-invasive clinical imaging (like CT or MRI) to discover gene-phenotype links across the entire genome without needing massive mutation counts?

Methods

The authors built a cross-modal analysis pipeline that operates in four main stages:

  1. Zero-Shot Mutation Scoring: Instead of training a model on specific cancer data, they used Evo 2, a biological foundation model trained on 9.3 trillion base pairs of DNA. For every somatic single-nucleotide variant in a patient, Evo 2 computes a "likelihood drop" ($|\Delta LL|$) between the reference and mutant alleles, serving as a proxy for functional disruption.
  2. Per-Gene Aggregation: These variant-level scores are aggregated into gene-level metrics (e.g., maximum severity, mean severity, mutation count) for any gene with at least five carriers.
  3. Radiomic Feature Extraction: Using automated deep learning segmentation tools (such as nnU-Net and TotalSegmentator), they extracted quantitative features from clinical scans, including tumor volume ($V_{tumor}$), tumor-to-organ ratio, intensity heterogeneity, and necrotic fraction.
  4. TMB-Adjusted Correlation: To ensure they weren't just detecting highly mutated tumors, they used a partial Spearman correlation adjusted for Total Mutation Burden (TMB). This isolates the specific effect of a gene's mutation severity on the tumor's physical appearance. Benjamini-Hochberg correction was applied to control the False Discovery Rate (FDR < 0.05).
Overview of the cross-modal analysis pipeline

Datasets

The framework was evaluated using paired whole-exome sequencing and diagnostic imaging from three cohorts in The Cancer Genome Atlas (TCGA) and The Cancer Imaging Archive (TCIA):

  • TCGA-cRCC (Clear cell renal cell carcinoma): $n = 162$ patients
  • TCGA-HCC (Hepatocellular carcinoma): $n = 57$ patients
  • TCGA-BC (Breast cancer): $n = 121$ patients

Total dataset size was 340 patients. No new benchmarks were introduced; the study leverages existing public multi-omics registries.

Key Insights and Ideas

  • Reframing Radiomics from Confirmation to Discovery: Traditional radiogenomics takes a known cancer gene and checks its imaging footprint. This inverse approach scans every mutated gene genome-wide without pre-selection, shifting the paradigm toward true hypothesis-free discovery.
  • Connecting Mendelian Genetics to Somatic Oncology: Many of the newly discovered genes associated with tumor volume are well-known in medical genetics for causing inherited structural disorders (like ciliopathies and cytoskeletal diseases). This suggests that the same cellular programs governing tissue structure during development also shape tumor morphology when disrupted somatically.
  • Bypassing the "Long-Gene" Confound: Traditional recurrence methods often flag exceptionally long genes (like CSMD3) as false positives because their large size naturally accumulates more random mutations. By evaluating the severity of mutations via Evo 2 rather than just counting them, this framework mitigates that bias.

Key Results

  • In the largest cohort (TCGA-cRCC), the pipeline identified 65 unique genes with FDR-significant gene-imaging associations.
  • 19 of these hits were established cancer drivers (e.g., PBRM1, SETD2, MTOR, PTEN), validating that the unsupervised framework successfully captures real cancer biology.
  • The remaining 46 hits were novel genes completely absent from curated cancer panels. Strong correlations with tumor volume were found for genes like ABCA13, MAST4, and ALMS1.
  • The FDR-significant set showed a 2.22-fold enrichment for known cancer drivers compared to non-significant tested genes ($29.2\%$ vs $13.2\%$).
  • While the smaller HCC and BC cohorts lacked the statistical power to clear strict FDR thresholds, their top nominally significant genes still successfully recovered canonical disease drivers (like CTNNB1 and TP53 for liver cancer; PIK3CA and CDH1 for breast cancer).
Rediscovery of established cancer drivers and cohort comparison

Limitations and Future Work

  • Sample Size & Power: Only the cRCC cohort was large enough to yield FDR-significant results; larger, more uniformly imaged cohorts are required to empower discovery in other cancer types.
  • Correlation vs. Causation: The framework identifies statistical associations between evolutionary sequence constraints and macroscopic imaging features, but functional perturbation experiments (e.g., CRISPR screens) or spatial transcriptomics are needed to prove causality.
  • Exome Restrictions: The analysis was limited to coding regions via whole-exome sequencing, missing out on Evo 2's capability to score non-coding regulatory variants.
  • Imaging Noise: Using historical data from TCGA introduces noise due to highly heterogeneous imaging protocols across different clinical sites. Future work will focus on applying this to larger, standardized registries and expanding to whole-genome sequencing data.

Summary updated

Abstract

The function of many genes is still unknown, and conventional driver-discovery methods, which rely on how frequently a gene is mutated, cannot assess genes that are only rarely affected. Here we pair Evo~2-based genome analysis with routine clinical imaging to identify gene--phenotype associations at genome-wide scale. For every somatic mutation across three TCGA cohorts (cRCC=clear cell renal cell carcinoma, HCC=hepatocellular carcinoma, and BC=breast cancer; $n = 340$ total), Evo~2 predicts a severity score, with no task-specific training. Per-gene severity summaries are then correlated with radiomic features extracted from paired tumor segmentations, controlling for total mutation burden. In TCGA-cRCC ($n = 162$), this sweep recovers established renal-cancer drivers and identifies 46 additional genes reaching false discovery rate (FDR) significance absent from curated cancer-gene panels, several of which are Mendelian ciliopathy and cytoskeletal-disease genes. These results demonstrate that pairing a genomic language model with widely available clinical imaging can serve as a hypothesis-free discovery tool for gene--imaging associations invisible to conventional approaches.
BibTeX
@misc{hauke2026foundation,
  author = {Frederik Hauke and Jeremias Krause and Patrick Wienholt and Christiane Kuhl and Ingo Kurth and Sikander Hayat and Jakob Nikolas Kather and Sven Nebelung and Daniel Truhn},
  title = {{Foundation-model-guided radiogenomic discovery linking cancer genomes to cancer scans}},
  year = {2026},
  eprint = {2607.20583},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2607.20583},
}
Figure