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

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

Problem Statement
High-grade gliomas integrate into neural circuits by forming functional synapses with neurons, a process driven by specific gene expression programs. While the protein machinery is characterized, the upstream noncoding regulatory programs (the "dark regulome") activating these synaptogenic genes remain unknown. Experimental dissection of the ~98% noncoding genome is intractable at scale. Consequently, researchers have turned to genomic foundation models (GFMs) and in-silico mutagenesis (ISM) to identify regulatory elements.

However, a critical methodological confound exists: likelihood-based RIS (Regulatory Influence Score) in masked or causal language models is tautologically coupled to local sequence predictability. Removing any sequence with high mutual information to its neighborhood (e.g., repetitive transposable elements memorized during pretraining) lowers regional likelihood, regardless of whether the element possesses regulatory function. Without a diagnostic to separate "sequence predictability" from "regulatory output," cross-architecture agreements may reflect statistical artifacts of large sample sizes rather than genuine biological signal.

Methodology
The authors analyzed 30,448 dark genome elements across 92 glioma-relevant loci (categorized into synaptogenic, proliferative, and brain control tiers) using three architecturally distinct foundation models:

1. Caduceus-Ph: A bidirectional Mamba masked language model (unsupervised).
2. HyenaDNA: A causal Hyena-based language model (unsupervised).
3. Enformer: A supervised convolutional-transformer trained on epigenomic tracks (CAGE).

The study introduces a residualization-and-permutation diagnostic to disentangle the two layers of variance:

• Residualization: For each element, four nuisance covariates were computed from the unmasked sequence: 4-mer Shannon entropy, GC content, log element length, and log distance to the Transcription Start Site (TSS). An Ordinary Least Squares (OLS) regression was fitted to predict $|RIS|$ from these covariates. The residual ($\hat{\epsilon}$) represents the variance not explained by sequence predictability or basic genomic features.
• Permutation Nulls: All cross-model overlaps and top-K agreements were evaluated against marginal-preserving permutation nulls (per-gene rank shuffling) to determine if observed enrichments were statistically significant beyond chance.
• Orthogonal Validation: Residualized rankings were cross-checked against conservation (phastCons), brain tissue cis-eQTLs (GTEx), and protein-protein interactions (STRING).

Key Results

1. The 10 kb Regulatory Horizon: A sharp geometric transition was observed where elements within 10 kb of the TSS exerted a mean $|RIS|$ approximately 480 times higher than those beyond 10 kb. This horizon persisted across all three models, all gene tiers, and all perturbation schemes, surviving every control applied.
2. Decomposition of Layers:
   • Language Models (Caduceus/HyenaDNA): These models shared a "sequence-predictability layer." Their top-100 rankings were dominated by long, well-predicted transposable elements (TEs) at moderate TSS distances. The overlap between Caduceus and HyenaDNA top-100 lists was 76 elements (60× enrichment over null), but this layer showed zero overlap with Enformer's top-100.
   • Supervised Model (Enformer): Enformer retained a distinct "regulatory-output layer," ranking short, proximal cCREs (promoters and enhancers) highly. Once predictability was controlled via residualization, Enformer was the only architecture to retain a measurable cCRE-discriminative signal.
3. Re-evaluation of Biological Claims:
   • Transposable Elements: The raw RIS hierarchy suggested a TE-mediated regulatory layer. However, after residualization, the correlation between residuals and cCRE membership was negligible for language models. A simple six-feature linear baseline (including length and distance) predicted top-decile membership with AUC = 0.985, indicating the raw hierarchy was driven by length and distance, not regulatory grammar.
   • NRXN1/NLGN1 Convergence: A previously highlighted convergence of NRXN1 and NLGN1 protein pairs failed proper permutation tests against the STRING PPI graph, with observed pair counts falling below the null expectation.
   • eQTL Enrichment: Despite the predictability confound, the top-100 elements from each model showed a consistent 3.3× enrichment for matching brain cis-eQTLs ($p_{emp} < 5 \times 10^{-3}$), suggesting a genuine, albeit modest, regulatory signal survives at the top of the rankings.
   • Conservation: Residualized RIS for language models was anti-correlated with conservation, suggesting their signals lean on recently evolved repetitive sequences rather than conserved regulatory elements.

Significance and Claims
The paper claims to provide a general methodological tool for any ISM-based regulatory study. Its primary contribution is the residualization-and-permutation diagnostic, which allows researchers to:

• Separate variance attributable to sequence predictability from regulation-driven variance.
• Evaluate overlaps and top-K agreements against rigorous, marginal-preserving nulls.
• Report effect sizes alongside p-values to avoid the trap of "large n, small p" artifacts.

The authors conclude that while GFMs offer a zero-shot route to the dark regulome, the lens is miscalibrated. Cross-architecture agreement alone is insufficient; true regulatory signal is only recovered when predictability layers are explicitly controlled. The study identifies a robust 10 kb proximal regulatory horizon and a shortlist of synaptogenic-locus candidates enriched for brain eQTLs, which are presented as candidates for experimental follow-up (e.g., CRISPRi perturbation), while retiring previous claims of TE-mediated regulation and specific protein-pair convergences that failed the new diagnostic tests.