Sequence-to-Function Genomic Models

Summary¶

Sequence-to-function models take a stretch of DNA as input and predict genome tracks — per-base-pair values matching what an experimental assay (RNA-seq, DNase-seq, ChIP-seq, Hi-C) would measure in a given cell type. Because the input is sequence alone, the same model scores a variant by predicting tracks for the reference and alternate alleles and taking the difference, which makes it applicable to any variant regardless of allele frequency, population, or whether it was ever observed. Since more than 98% of human genetic variation is non-coding, this model class is currently the main computational route to mechanistic interpretation of the variants that GWAS actually finds.[1]

Why this model class exists¶

Non-coding variants have no protein-level consequence to read off, and they act through several distinct molecular routes: chromatin accessibility, histone modification, three-dimensional conformation, transcription-factor binding, transcript abundance, splicing and 3′ end processing. Each route requires a different assay to observe and, historically, a different predictor to model. Sequence-to-function models unify these by learning a shared sequence representation and attaching per-assay output heads to it.[1]

The key property is that they are not annotation lookups. They do not require the variant to overlap a catalogued element, to be evolutionarily conserved, or to have been observed at sufficient frequency for a QTL study to have power — the three failure modes that most conventional variant-to-gene pipelines share.

The two governing trade-offs¶

Model design in this family has been shaped by two trade-offs, both explicitly diagnosed in the AlphaGenome paper.[1]

1. Input context length versus output resolution¶

Computational cost scales with (sequence length × output resolution), so models have had to sacrifice one.

Regime	Examples	Context	Output resolution	What is lost
Short context, base resolution	SpliceAI, BPNet, ProCapNet	≤10 kb	1 bp	distal regulatory elements
Long context, binned resolution	Enformer, Borzoi	200 kb / 524 kb	128 bp / 32 bp	splice sites, TF footprints, polyadenylation signals blurred
Both	AlphaGenome	1 Mb	1 bp (some heads 128 bp / 2,048 bp)	—

The trade-off is not cosmetic. Coarse bins blur exactly the fine-scale features — splice sites, transcription-factor footprints, polyadenylation sites — that determine a large fraction of variant effects.[1] Conversely, short context misses the distal enhancers that carry much of the regulatory signal: 99% (465 of 471) of CRISPRi-validated enhancer–gene pairs lie within 1 Mb, so a 10 kb window sees only a small fraction of them.[1]

2. Modality breadth versus specialist accuracy¶

The second axis runs between models specialised for one modality — SpliceAI and Pangolin for splicing, ChromBPNet for local accessibility, Orca for 3D architecture — and generalists (DeepSEA, Basenji, Enformer, Sei, Borzoi) that predict many modalities from one shared representation.[1]

Specialists have historically won on their home task but cannot describe a variant that acts through several routes at once, and give no basis for comparing effects across routes.
Generalists offer one model and a reusable representation for fine-tuning, but have lagged specialists on some tasks (notably splicing) or omitted modalities entirely (notably contact maps).[1]
Even within a modality, specialisation can be partial: splice-site predictors such as SpliceAI and Pangolin model site probability but not splice junction identity or competition between sites.[1]

Evidence that the generalist penalty is not inevitable: AlphaGenome's ablations show the fully multimodal model generally outperforming models trained on single modality groups, with the benefit strongly task-dependent — accessibility variant prediction was fine on accessibility data alone, while eQTL prediction benefited from the full multimodal signal. Excluding any single modality group degraded performance only modestly, indicating substantial redundancy between modalities.[1]

Recurring design lessons¶

Ablations reported for AlphaGenome give the clearest published evidence on what actually drives performance in this class:[1]

Base-pair training targets help selectively. The gain concentrates in tasks needing fine detail (splicing PSI, ATAC profile shape). Contact maps, histone ChIP correlation, and variant scores that aggregate over gene bodies were largely resolution-insensitive.
Long context must be present during training, not just inference. Models trained on ≤32 kb underperformed 1-Mb-trained models even when given 1-Mb context at inference. The converse is more forgiving: a 1-Mb-trained model evaluated on shorter contexts often matched models trained at those shorter lengths, so context can be traded for inference speed after the fact.
Distillation can replace ensembling. Distilling many teachers into one student matched or beat mean ensembles of independently pretrained models, at a fraction of the inference cost.
Perturbed inputs during distillation are essential for variant tasks. Distilling on unmutated sequences alone cost 0.06 auROC on eQTL sign prediction — the student must see altered sequences to learn variant effects rather than memorising the reference genome.[1]

Evaluating these models¶

Two benchmark families are used, and conflating them is a common error:[1]

Genome track prediction — how well does the model reproduce a held-out assay track on unseen reference-genome intervals? This tests generalisation across sequence but says nothing directly about variant sensitivity.
Variant effect prediction — how well does the REF-versus-ALT difference recover a measured variant effect? Ground truth comes from fine-mapped QTLs (eQTL, sQTL, caQTL, dsQTL, bQTL, paQTL), CRISPRi enhancer perturbation, MPRA saturation mutagenesis, and clinical variant classification (ClinVar).

Independent, protocol-controlled benchmarking of this model class is scarce; GENEB is the closest available, evaluating 40 genomic foundation models under a unified linear-probing protocol — though it explicitly under-tests the >10 kb long-range regime these models occupy.

Track accuracy is a prerequisite for, but not a guarantee of, variant accuracy: a model can predict the reference track well while being insensitive to the single-base perturbations that matter. This is why the AlphaGenome authors use fold-specific pretrained models for track evaluation but the distilled student — trained on mutated inputs — for all variant evaluations.[1]

Synthesis: The benchmark separation above and the vault's independent record of train/test leakage in enzyme kinetics prediction (see DLKcat Generalization Failure and Train/Test Leakage) describe the same underlying hazard from different fields: a model can score well on a held-out split that shares structure with training data while failing the transfer the metric is taken to certify. Neither source discusses the other; the parallel is drawn here.

How the leading models complement each other¶

The three models this vault covers from primary sources are not three attempts at one task. They occupy different output spaces, and the published benchmarks show each winning where the evaluation phenotype most resembles what it natively predicts.

	SpliceAI	Borzoi	AlphaGenome
Context	10 kb [4]	524 kb [5]	1 Mb [6]
Output resolution	1 bp [4]	32 bp [5]	1 bp (some heads 128 bp / 2,048 bp) [6]
Predicts	donor/acceptor probability per base [4]	RNA-seq coverage; other statistics derived from it [5]	11 output types incl. splice sites, usage, junctions, chromatin, contact maps [6]
Trained on	GENCODE annotations, reference sequence only [4]	ENCODE/GTEx RNA-seq + chromatin assays [5]	multimodal experimental tracks, then distilled on mutated sequence [6]
Native strength	de novo splice-site creation anywhere in an intron [5]	usage/ratio changes near junctions and 3′ ends [5]	multimodal mechanism reconstruction at one locus [6]

Three published complementarity results, not opinions:

Distance-dependent splicing performance. Pangolin (a SpliceAI-lineage splice-site model) held a slight overall advantage over Borzoi on fine-mapped sQTLs within 10 kb of a splice site (ΔAUPRC 0.01), because most distant variants are de novo splice-gain mutations that a site-probability model classifies easily from the local effect. Borzoi won within 200 bp of a junction (ΔAUPRC 0.02), where the question is competitive usage rather than site existence. The average rank of both models beat either alone (ΔAUPRC > 0.02).[5]
AlphaGenome does not dominate splicing either. On the MFASS minigene MPRA it lost to Pangolin (auPRC 0.51 vs 0.54), while beating SpliceAI and DeltaSplice (0.49 each); it is state of the art on 6 of 7 splicing benchmarks, not 7 of 7.[6]
Sequence models and conservation are not redundant. Borzoi and CADD each reached AUROC 0.55 discriminating gnomAD singletons from common variants in cCREs; combined, 0.57.[5] Conservation scores are agnostic to mechanism; sequence models are informative at non-conserved bases. Neither subsumes the other.

Synthesis: the pattern across all three results is that these models fail differently, not uniformly worse or better. A splice-site model asks "is there a site here?"; a coverage model asks "how much of each isoform?"; a multimodal model asks "what changed across every layer?". Where a benchmark phenotype matches one question, that model wins. Combining them helps because their errors are structurally distinct — which is why ensembling repeatedly beat any single model in the cited evaluations [5], and why the practical recommendation is to run more than one rather than to pick the highest headline number. See RNA-seq-Derived Molecular Phenotypes for why the benchmark phenotype is doing so much of the work here.

A generational caveat. SpliceAI, Borzoi and AlphaGenome are also successive generations (2019, 2025, 2026), and the later models were designed partly to fix the earlier ones' limits — Borzoi's stated next goals were megabase receptive fields and finer resolution [5], which is what AlphaGenome delivered [6]. Complementarity and supersession are both real here, and which one applies depends on the task: for genome-wide exome annotation with a calibrated threshold and precomputed tables, SpliceAI remains in production [4]; for a specific locus needing mechanism, AlphaGenome; for coverage-shape questions such as intronic polyadenylation competition, Borzoi [5].

Known limitations of the class¶

These are properties of the approach, not of any one implementation, and all are acknowledged by the AlphaGenome authors as shared challenges:[1]

Distal effects degrade. Accuracy falls with variant-to-TSS distance across every expression task; elements beyond ~100 kb remain poorly captured, and very distal enhancer effects are systematically underestimated.
Cell-type specificity is the weak axis. Average expression levels are predicted well; deviations between cell types — the part that actually carries tissue-specific biology — are much harder.
Reference-genome centric. Personal-genome prediction is a known weakness of this model class and was not benchmarked for AlphaGenome.
Protein-coding bias. Training data and evaluations concentrate on protein-coding genes; non-coding genes such as microRNAs are under-represented.
No calibrated uncertainty. Predictions come without confidence estimates, which the authors name as a target for future work.

The molecular-to-phenotypic gap: these models predict molecular consequences of a variant, not phenotypic ones. Complex-trait phenotypes additionally involve gene function, development, environment and gene-to-disease effects that lie outside the sequence-to-function scope.[1] Empirically this shows as an enrichment-versus-recall trade-off: thresholds strict enough to enrich causal trait-altering variants 5–11 fold over matched controls retain only a small minority of them, and recall is worst for GWAS (complex trait) variants.[1] Treating a high molecular score as evidence of phenotypic causality is the central misuse of this model class.

Relationship to other approaches in this vault¶

Versus conservation-based deleteriousness scores (phyloP, and by analogy AlphaMissense for coding variants): conservation scores are agnostic to mechanism and blind to fast-evolving or lineage-specific regulatory elements. Sequence models predict a direction and a mechanism, and remain informative at non-conserved bases. They are complements, not substitutes.[1]
Versus QTL-based variant-to-gene mapping: colocalization and TWAS require a well-powered molecular QTL study in a matching tissue, so power depends on allele frequency, sample size and LD. Sequence models have no such dependence, which is why they resolve loci that colocalization cannot — see Direction-of-Effect Assignment at GWAS Loci.
Versus supervised enhancer–gene models: ENCODE-rE2G and the Activity-by-Contact (ABC) Model consume measured assay data (DNase, H3K27ac, Hi-C) for a specific biosample; sequence models predict those same signals from sequence alone. The two compose: AlphaGenome features added to ENCODE–rE2G (extended) improved it beyond either alone.[1]
As a feature source for expression imputation: scPrediXcan fine-tunes Enformer on pseudo-bulk single-cell expression to build cell-type-specific TWAS models, an example of using a pretrained sequence representation rather than its zero-shot predictions.

Citations¶

[1] Avsec, Latysheva, Cheng, Novati, Taylor et al. (2026), "Advancing regulatory variant effect prediction with AlphaGenome", Nature 649:1206–1217. Supports: the definition of the model class; both trade-offs and the model taxonomy; the 98%-non-coding and 465/471-enhancer-pair figures; the ablation lessons; the track-versus-variant benchmark distinction; all stated limitations; the enrichment-versus-recall figures for trait-altering variants. Location: Introduction; "Unifying DNA sequence-to-function model"; "Performance overview"; "Multimodal view of variant effects"; "Model and data ablations"; "Discussion"; Figs. 1, 6f, 7. Source paper: s41586-025-10014-0.pdf

[4] Jaganathan, K. et al. (2019). Predicting Splicing from Primary Sequence with Deep Learning. Cell 176(3):535–548.e24. Supports: SpliceAI's 10 kb context, 1 bp resolution, donor/acceptor-probability output space, GENCODE-annotation training, and its continued production use via calibrated thresholds and precomputed genome-wide tables. Location: Results ("Accurate Prediction of Splicing from a Primary Sequence Using Deep Learning"); Discussion. Source copy: j.cell.2018.12.015.md

[5] Linder, J., Srivastava, D., Yuan, H., Agarwal, V. & Kelley, D. R. (2025). Predicting RNA-seq coverage from DNA sequence as a unifying model of gene regulation. Nature Genetics 57:949–961. Supports: Borzoi's 524 kb / 32 bp design and coverage-first output space; the Pangolin-versus-Borzoi distance-stratified sQTL result and the ensemble gain; the Borzoi+CADD gnomAD result; the stated next goals of megabase context and finer resolution. Location: Results ("RNA-seq model design"; "Functional splicing variant interpretation"; "Borzoi prioritizes genetic variants that influence expression"); Discussion; Figs. 5, 7. Source paper: s41588-024-02053-6.pdf

[6] Avsec, Ž. et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature 649:1206–1217. Supports: AlphaGenome's 1 Mb / base-pair design and 11 output types; the MFASS result versus Pangolin and the 6-of-7 splicing benchmark record. Location: "Unifying DNA sequence-to-function model"; "Improved splicing variant predictions"; Fig. 3i. Source paper: s41586-025-10014-0.pdf

Further models characterised here from secondary description¶

These models are described on this page only as the AlphaGenome and Borzoi papers characterise them [1,5]; their originals were not independently inspected, and specific claims should be verified against the primary source before reuse. (Enformer and Pangolin were on this list until their primary papers were read; they now have their own sourced pages and citations [7,8].)

Zhou, J. (2022). Sequence-based modeling of three-dimensional genome architecture from kilobase to chromosome scale (Orca). Nat. Genet. 54:725–734.
Avsec, Ž. et al. (2021). Base-resolution models of transcription-factor binding reveal soft motif syntax (BPNet). Nat. Genet. 53:354–366.
Cochran, K. et al. (2024). Dissecting the cis-regulatory syntax of transcription initiation with deep learning (ProCapNet). bioRxiv.
Trevino, A. E. et al. (2021) / ChromBPNet — the local chromatin-accessibility specialist benchmarked against by AlphaGenome.
Kelley, D. R., Snoek, J. & Rinn, J. L. (2016). Basset: learning the regulatory code of the accessible genome with deep convolutional neural networks. Genome Res. 26:990–999.

[7] Avsec, Ž. et al. (2021). Effective gene expression prediction from sequence by integrating long-range interactions. Nature Methods 18:1196–1203. Supports: Enformer's 200 kb input, 128 bp resolution and 100 kb receptive field; attention-versus-dilated-convolution ablations; the cell-type generalisation limitation. Location: Results; Discussion. Source paper: s41592-021-01252-x.pdf

[8] Zeng, T. & Li, Y. I. (2022). Predicting RNA splicing from DNA sequence using Pangolin. Genome Biology 23:103. Supports: Pangolin's ±5 kb receptive field, tissue-specific usage output, and the finding that multi-tissue/multi-species/quantitative training targets each improved accuracy independently of architecture. Location: Main text; Methods. Source paper: s13059-022-02664-4.pdf