This study probes self-supervised generalists alongside task-specific specialists using a mechanistically focused paradigm centered on expert-curated single-nucleotide substitutions within the splicing regions of the exon-rich OPA1 gene, demonstrating that foundation models not only match task-specific networks, but also generalize across mutation types.
Abstract
As benchmarks saturate, traditional validation methods fail to capture the probabilistic representations of genomic foundation models. This study probes self-supervised generalists (NTv3 650M, Evo2 40B, Genos 10B) alongside task-specific specialists (SpliceAI, AlphaGenome, Borzoi) using a mechanistically focused paradigm centered on expert-curated single-nucleotide substitutions within the splicing regions of the exon-rich OPA1 gene. Extending this analysis across the broader corpus of OPA1 mutations, variants of uncertain significance (VUS), and an independent 65-gene dataset of spliceogenic mutations demonstrates that foundation models not only match task-specific networks, but also generalize across mutation types. Training data scale, genetic diversity, and evolutionary context are key for capturing RNA splicing syntax, whereas sparse routing may constrain performance. Mechanistically, foundation models exhibit a compute asymmetry: Evo2 40B frontloads compute into massive parameter scale to resolve single-nucleotide variants within a narrow 425-bp context window, whereas the bidirectional NTv3 650M model and the Genos 10B mixture-of-experts (MoE) architecture backload compute to inference, requiring spatial logit aggregation and vast context windows of up to 94 kb. Furthermore, accuracy may benefit from locus-specific thresholding rather than mutation-class calibration, exposing representation boundaries of the self-supervised models.
This work shows that redundancy among partially overlapping deep mutational scanning experiments, often treated as noise, yields a consensus landscape of mutational constraint, and introduces "variant soundness", an overlap-aware rank-fusion score that uses cross-experiment consistency to derive an assay-agnostic order...
Riccardo Arnese, G. Gambardella· Molecular Systems Biology· 0 citations
Short-read RNA sequencing has revolutionized transcriptome analysis, but accurate isoform-level quantification faces substantial challenges from overlapping sequences among complex isoforms and inherent sequencing biases, leading to uneven coverage. Current alignment-based and alignment-free methods, while valuable, la...
Qiang Su, Yi Long, De-Ming Gou et al.· bioRxiv· 0 citations
PHACTn (Phylogeny-Aware Computing of Tolerance for nucleotide variants), a training-free, parameter-minimal method that infers nucleotide variant tolerability by traversing the mammalian phylogenetic tree and explicitly modeling the evolutionary independence of observed substitutions and their distance from the query s...
Ceren Yıldırım, Nurdan Kuru, Ogun Adebali· bioRxiv· 0 citations
Long-context sequence-to-function models enable nucleotide-resolution prediction of regulatory variant effects, motivating comprehensive mutational interrogation across broad genomic contexts. Yet conventional in silico saturation mutagenesis (ISM) scores mutations one at a time, requiring millions of model evaluations...
Han Yuan, Xing-Fan Huang, Benjamin J. Auerbach et al.· bioRxiv· 0 citations
RNA splicing is frequently disrupted by genetic variation, contributing to a wide range of human diseases. Identifying the variants responsible for altered splicing is therefore crucial but has long been challenging. Deep learning models that predict splicing from genomic sequence have substantially advanced their dete...
K. Jaganathan, Jian-Bin Chen, Xin Liu et al.· bioRxiv· 1 citation