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

A new diagnostic method separates genuine regulatory signal from statistical predictability in genomic foundation models, revealing that much of what these models report about the noncoding genome may reflect sequence memorization rather than biology, according to a preprint posted June 5 on arXiv [1]. The study, which examined 30,448 dark genome elements across 92 glioma-relevant loci, tackles a core problem in computational genomics: likelihood-based scoring in language models is tautologically coupled to local sequence predictability, meaning elements that are simply easier for the model to predict can appear regulatory even when they are not [1][3]. The authors introduce a residualization-and-permutation diagnostic that controls for four nuisance covariates — k-mer entropy, GC content, log element length, and log TSS distance — and evaluates reported overlaps against a per-gene permutation null [3][4]. When applied across three architecturally distinct models — Caduceus-Ph, HyenaDNA, and Enformer — the diagnostic produced a clean decomposition. The two language models, Caduceus-Ph and HyenaDNA, share a sequence-predictability layer that co-ranks long, well-predicted transposable elements, while Enformer alone retains residual cCRE-discriminative signal once predictability is controlled [3][4]. The top-100 lists from the two layers showed zero overlap [1][4]. A six-feature linear baseline matched Caduceus top-decile membership with an AUC of 0.985, indicating that the class hierarchy in raw regulatory impact scores is largely a length-and-distance hierarchy with a regulatory veneer [1][4]. The study also found that a sharp 10-kilobase proximal-regulatory horizon survived every control applied, including scoring windows, perturbation schemes, and residualization [1][3]. Cross-checks against conservation data, brain cis-eQTLs, and STRING protein-protein interaction networks anchored what biological signal remains. Top-100 elements across all three models showed a 3.3-fold enrichment per model for matching brain eQTLs, with an empirical p-value below 5×10⁻³ [1][4]. However, a proposed transposable-element regulatory layer and a NRXN1-NLGN1 protein-pair convergence both failed proper permutation tests once those tests were constructed [3][4]. The authors describe the trans-synaptic adhesion narrative as post-hoc storytelling on two data points that does not survive a proper PPI null [4]. The findings arrive amid broader scrutiny of genomic language models. A separate preprint introduced the Mechanistic Invariance Test, a 650-sequence benchmark across eight classes, and found that models including Caduceus score regulatory elements at incorrect positions higher than correct positions, driven by AT-content correlation rather than positional regulatory logic [5][6]. That study reported compositional effects dominating positional effects by 46-fold, with a 100-parameter position-aware PWM achieving perfect performance where billion-parameter networks failed [5][6]. The authors of the dark regulome study deliver their diagnostic as a general methodological tool for any ISM-based regulatory study [1][2].