A representation-accessibility analysis of frozen genomic language models across regulatory, epigenetic, promoter, splice-site, and variant-effect prediction tasks shows that local biological signal is partially present in frozen representations, but is not always accessible through final pooled embeddings.
Abstract
Genomic foundation models are increasingly reused as frozen feature extractors for downstream sequence prediction, offering a compute-efficient alternative to full fine-tuning. However, it remains unclear when biological information encoded by these models is accessible without task-specific adaptation. We present a representation-accessibility analysis of frozen genomic language models across regulatory, epigenetic, promoter, splice-site, and variant-effect prediction tasks. We evaluate DNABERT-2, Nucleotide Transformer, HyenaDNA, GENERATOR-v2, and Omni-DNA under unified frozen-probing protocols, while separating diagnostic readout analyses from validation-selected checks. Our results reveal a consistent task-dependent pattern: frozen probes recover 95-100 % of fine-tuned performance on promoter tasks, but average splice-site recovery drops to 60-88 %. Frozen embeddings are also competitive on broad Genomic Benchmark tasks such as coding-region and species-discrimination classification, but show larger gaps on some regulatory and OCR tasks. Layer-wise probing, in-silico mutagenesis, variant-effect prediction, and embedding geometry show that local biological signal is partially present in frozen representations, but is not always accessible through final pooled embeddings.
Genomic language models achieve strong performance across regulatory-genomics tasks, yet what these models internally represent remains opaque, and the field lacks a principled procedure for verifying that an apparent ``concept''inside a model is real rather than an artifact of sequence composition. We introduce a framework that combines sparse dictionary learning with causal intervention to extract, validate, and causally test interpretable features in genomic foundation models. Training top-$k$ sparse autoencoders on the hidden activations of two architecturally distinct models, Nucleotide Transformer ($6$-mer tokenization) and DNABERT-2 (byte-pair encoding), we recover thousands of monosemantic features that map to transcription-factor (TF) sequence motifs. We show that the naive validation of such features against position weight matrices is severely confounded by GC composition and repetitive elements, producing hundreds of spurious ``TF features'', and we develop a composition-matched, binding-resolved protocol that removes these confounds. Critically, we move beyond correlation: by ablating individual dictionary directions during the model's forward pass and measuring the induced shift in the model's own predictive distribution, we establish that specific features are \emph{causally} used to represent cell-type-specific TF binding, not merely motif presence. Across three transcription factors (CTCF, GATA1, REST) and both architectures, causally validated binding features emerge reproducibly ($7$--$14$ of $15$ tested features per condition), while two classes of negative control, scrambled binding labels and randomly selected features, yield no detectable signal. The framework is purely computational, uses only public data, and provides a reusable standard for interpretability claims in genomic deep learning.
Transcriptomic foundation models are increasingly used as reusable cell and gene representations, but validating them on new data under weak supervision and distribution shift is hard: standard comparisons conflate genuine representation signal with model capacity, row-identity artifacts, gains over strong task-specific baselines, and outcome rules chosen after seeing the test set. We introduce a pre-registered, final-test-once evaluation framework that locks the outcome rule, seeds, and target-gene-grouped splits before any test data are seen, and scores each frozen representation against a strong expression baseline, a matched-capacity Gaussian control, and a within-split row-identity (shuffle) control; only the per-cell embedding-extraction step is model-specific. Applying it to three architecturally distinct models-Geneformer, scGPT, and UCE-across two external Replogle Perturb-seq datasets (RPE1 and K562), all three clear the capacity and row-identity controls by a wide margin, yet none reliably beats the expression baseline: the strongest (Geneformer) exceeds it by at most about $0.03$ test $R^2$ and clears the pre-registered four-of-five-seed threshold in neither dataset, while scGPT and UCE fall below it. All three therefore land in the same pre-registered partial-replication category-a consistent cross-architecture outcome, even though the baseline-relative gap differs in sign and magnitude across models. These representations carry real structure beyond trivial controls but, under this weak magnitude label, do not transfer past a simple strong baseline; the locked framework is reusable for any frozen transcriptomic representation by swapping only the extraction step.
Pretrained genome language models provide reusable representations for DNA sequence analysis, but turning them into reliable downstream predictors remains non-trivial. Their practical performance depends strongly on fine-tuning recipes, and default recipes reported in prior studies may be suboptimal for new tasks or model backbones, making weak downstream results difficult to interpret. These requirements place a substantial operational burden on many intended users, whose expertise is often centered on biological questions and interpretation rather than machine-learning engineering. Reliable use of genome language models therefore requires more than conventional AutoML-style tuning: it requires a systematic, budget-aware, and auditable procedure that lowers the barrier to downstream adaptation. We present GenomeHarness, an agentic harness for adapting genome language models through controlled search over fine-tuning recipes. GenomeHarness combines an AI agent for proposing and repairing recipe edits, a harness for protocol-constrained execution, resource management, and test isolation, and a Monte Carlo tree search controller for allocating search effort across recipe lineages. We evaluate GenomeHarness on DNABERT2 and NTv2-100M-Multi across the NT Benchmark and Genomic Benchmarks. Final evaluation is performed using three random seeds after recipe freezing. Across 52 model-task settings, GenomeHarness improves mean test MCC in 47 settings, including 24 of 26 DNABERT2 settings and 23 of 26 NTv2-100M-Multi settings. The gains are especially pronounced on Genomic Benchmarks and on tasks where the root recipe is unstable or poorly matched, such as human ocr ensembl task. Search traces further show that GenomeHarness progressively identifies stronger recipes, turning downstream adaptation into a controlled and auditable workflow rather than a manual tuning process.
Weicai Long, Yusen Hou, Houcheng Su et al.· 0 citations
Across the included studies, stronger evidence for gLM utility was generally associated with biologically informed or task-aligned model design, including evolutionary alignments, motif-aware objectives, long-context architectures, RNA structural priors, population-aware representations, and domain-specific pretraining.
Mahinaz A. Mashhour, Manal Abdel Wahed, Mai S. Mabrouk· Biochemical and Biophysical...· 0 citations
Large-scale genomic language models (gLMs) hold promise for modeling gene regulation, yet their ability to capture personal gene expression variations remains unresolved. We developed xDecoder, a unified decoding framework that utilizes gLMs and sequence-to-function (S2F) embeddings to learn how personal genetic variation shapes gene expression from paired genome-transcriptome data. Compared to the pretrained genomic models, xDecoder with personalized DNA-RNA training makes cross-individual prediction tractable for seen genes in a few-shot setting. However, zero-shot prediction at unseen loci remains unreliable and gene-dependent, revealing a cross-locus transfer bottleneck of current sequence models. Experiments incorporating individual-level chromatin accessibility suggested that regulatory-state information important for unseen-locus prediction is not fully captured by current DNA-only models. Overall, these results highlight the potential utility of the few-shot setting, the limitations of DNA-only models, and point toward multi-omic, variant-aware frameworks as a promising direction for building personalized regulatory models.
GB.GeneUnet, an 837M-parameter transformer-based U-Net pretrained on 6 trillion tokens from multi-species genomes in OpenGenome2 is introduced, extending genomic context to 1 Mb with up to 100× inference speedup over GeneMoE, a preliminary MoE transformer baseline of similar model size pretrained on the same data.
Ning Sun, William de Vazelhes, Pan Li et al.· bioRxiv· 0 citations