The Dark Regulome: Disentangling Predictability from Regulation in Genomic Foundation Models

2026-06-08 · Source: cs.CL updates on arXiv.org · Field: Science & Research — Life Sciences & Biology, Health & Medical Research, Mathematics & Computational Sciences · Depth: Expert, extended

Summary

This study introduces a novel residualization-and-permutation diagnostic to accurately interpret genomic foundation model outputs for the "dark regulome" in high-grade gliomas. Applied to Caduceus-Ph, HyenaDNA, and Enformer across 30,448 dark genome elements at 92 glioma-relevant loci, the diagnostic disentangles sequence predictability from true regulatory influence. It reveals a consistent 10 kb proximal-regulatory horizon and a clear architectural split: language models (Caduceus-Ph, HyenaDNA) share a predictability layer ranking long transposable elements, while Enformer uniquely identifies a regulatory-output layer of short proximal cCREs, with zero top-100 overlap. The analysis also confirms a 3.3x enrichment (p_emp<5x10^-3) for brain cis-eQTLs in top-100 elements, providing validated synaptogenic-locus candidates.

Key takeaway

For research scientists interpreting genomic foundation model outputs, you must apply the residualization-and-permutation diagnostic to avoid misinterpreting sequence predictability as true regulatory function. This diagnostic helps you identify genuinely regulatory elements, especially those within the robust 10 kb proximal-regulatory horizon. Focus your experimental validation efforts on candidates that show regulatory-output layer signals, like those uniquely identified by Enformer, to ensure biological relevance.

Key insights

A diagnostic separates sequence predictability from true regulatory influence in genomic foundation models, revealing distinct functional layers.

Principles

• ISM likelihood scores conflate regulatory function with sequence predictability.
• Cross-architecture agreement alone does not isolate true regulatory signals.
• A 10 kb proximal-regulatory horizon is a robust signal in genomic models.

Method

The residualization-and-permutation diagnostic separates predictability-driven from regulation-driven RIS variance by controlling for nuisance covariates (k-mer entropy, GC content, log element length, log TSS distance) and evaluating against permutation nulls.

In practice

• Apply the diagnostic to separate predictability from regulation.
• Prioritize 10 kb proximal elements for experimental validation.

Topics

• Genomic Foundation Models
• In-silico Mutagenesis
• Dark Regulome
• Glioma Biology
• Gene Regulation
• Noncoding DNA

Best for: AI Scientist, Research Scientist

Related on AIssential

• Knowledge Graphs and Reasoning LLMs for Finding Simple Yet Effective Transcriptomic Perturbation Predictors — Simple K-nearest neighbour models leveraging knowledge graphs, enhanced by RL-trained LLMs, offer competitive transcriptomic perturbation prediction.
• Stringology-Based Motif Discovery from EEG Signals: an ADHD Case Study — Stringology-based motif discovery offers a novel, complementary approach to analyze EEG dynamics in neurodevelopmental disorders like ADHD.
• Mitosis Detection in the Wild: Multi-Tumor and Context-Aware Generalization in the MIDOG 2025 Challenge — "In the wild" mitosis detection demands models robust to diverse biological and contextual variance.
• PromoterAtlas: decoding regulatory sequences across Gammaproteobacteria using a transformer model — PromoterAtlas is a transformer model that decodes bacterial regulatory sequences across thousands of Gammaproteobacteria species.
• Rethinking Genomic Modeling Through Optical Character Recognition — OpticalDNA reframes genomic modeling as OCR-style document understanding, achieving efficiency and high-fidelity compression via visual layouts.
• Uncovering Physical Drivers of Dark Matter Halo Structures with Auxiliary-Variable-Guided Generative Models — DL-CFM disentangles physical factors in generative models using auxiliary variables, enhancing interpretability and control for scientific data.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.