henhaobio
Sign in

bioRxivGenomicsPreprint

Interpretable variant effect prediction from genomic foundation model embeddings

Michael T. Pearce, Thomas DoomsiD, Ryo YamamotoiD, Shant AyanianiD, Alexander J. RyuiD, Joshua MeehliD, Carl MolnariD, Tuomo KiiskineniD, Mark BisselliD, Dron HazraiD, Ching Fang, Nam Nguyen, Michael Anderson, Collin OsborneiD, Patrick Emmet DuffyiD, Bridget ToomeyiD, Elena MyasoedovaiD, Eric W. KleeiD, Panagiotis D. KorfiatisiD, Matt RedloniD, Archa JainiD, Daniel Balsam, Nicholas K. WangiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Scientific foundation models learn high-dimensional representations from diverse data modalities, yet what they encode and how to extract that knowledge remain open questions. Here we show that probing the internal representations of Evo 2, a 7-billion-parameter genomic foundation model, enables accurate and interpretable genetic variant effect prediction. We introduce a covariance-based probe that captures second-order structure from Evo 2 sequence embeddings to predict variant pathogenicity across variant types and functional consequences, matching or exceeding specialized predictors within their domains. To ground these predictions in known biological mechanisms, we train a complementary panel of probes on existing annotations to detect which genomic properties are disrupted by a variant. This categorized evidence is then integrated with each variants genomic context through a language model to generate variant-specific mechanistic hypotheses. Our pathogenicity predictions correlate with experimental measures of variant function, clinical penetrance, and biobank disease associations while the mechanistic hypotheses are consistent with expert reviews, known mechanism classes, and downstream molecular readouts. We release pathogenicity scores, disruption profiles, and contextualized interpretations for 4.2 million variants from the ClinVar database as an open resource through the Evo Variant Effect Explorer (EVEE). More broadly, this structured probing approach offers a general framework for interrogating foundation models across scientific disciplines and grounding their outputs in existing domain concepts.

In-depth summary

Claude Opus 5.5 · xhigh effort · 7 Nov 2026 · from the full text of version 4 (11 Jul 2026) · CC BY 4.0

A low-rank covariance probe on Evo 2 embeddings predicts ClinVar pathogenicity across SNVs and indels, while annotation probes plus an LLM produce mechanistic hypotheses; scores released for 4.2 million variants.

Variant-effect predictors are split by regime (protein models mostly cover missense, sequence-to-function models mostly regulatory effects) and typically output a score without a mechanism, whereas clinical classification under ACMG/AMP guidelines relies on categorized evidence. Instead of scoring with Evo 2 output likelihoods, this work probes the model's embeddings to predict pathogenicity across variant classes and to name the genomic properties a variant disrupts.

Key results

  1. ShownTrained only on ClinVar single-nucleotide variants (SNVs), one covariance probe separates pathogenic from benign variants within every consequence type, stays competitive on a consequence-balanced benchmark, and transfers zero-shot to indels.

    Gene-held-out AUROC 0.900 (nonsense) to 0.984 (intronic) on 833,970 SNVs, above Evo 2 likelihood scoring, CADD, GPN-MSA, NTv3, AlphaGenome and a mean-pooled Evo 2 probe, level with AlphaMissense on missense. Matched or beat the best comparator on 158,616 deconfounded variants. Indels: AUROC 0.991 vs 0.980 (CADD v1.7) on 73,961, but 0.772 for frameshifts. · A single covariance probe matches or exceeds…; SNV benchmark; Indel benchmark; Deconfounded benchmark

  2. ShownProbe scores correlate with deep mutational scan (DMS) readouts in four genes excluded from probe training, beating Evo 2 likelihood scoring and mean-pooled probes and roughly matching AlphaMissense and CADD.

    Spearman |ρ| for LDLR 0.41 vs 0.34 (AlphaMissense) and 0.32 (CADD); BRCA1 0.55 and TP53 0.70, equal to AlphaMissense; BRCA2 0.44 vs 0.46 for CADD. The covariance probe consistently exceeded the mean-pooled probe. · Predictions generalize beyond conservation to…; LDLR

  3. ShownPer-variant disruption profiles from annotation probes separate literature-curated disease-mechanism classes within genes, whereas pathogenicity scores alone do no better than chance.

    236 binary annotation probes, median AUROC 0.919. Multinomial logistic regression, stratified 5-fold CV, in LDLR (N = 1,581), LMNA (241), MYH7 (269), TP53 (488): mean balanced accuracy 0.74 vs 0.51 for position only, 0.33 for AlphaMissense, 0.26 for CADD. EVEE alone did not lead on MYH7, whose classes follow primary structure. · Disruption profiles recover disease mechanisms…; Annotation panel; Within-gene supervised mechanism-class recovery…

  4. ShownRNA-seq of Cancer Cell Line Encyclopedia (CCLE) lines confirmed a predicted BRCA1 cryptic splice acceptor at nucleotide resolution; across predicted splice and nonsense-mediated decay (NMD) variants, 57% of variant-carrier pairs showed the expected RNA disruption.

    BRCA1 c.5278-1G>A in JHOS-4 (variant allele fraction 0.96): no canonical exon-20 reads, 24 exon-skip and 25 reads at the predicted +8 acceptor, vs 185–190 and 248–294 canonical reads in wild-type JHOS-2 and A2780. Overall 47/82 validated: 19/23 (83%) homozygous or loss-of-heterozygosity vs 28/59 (47%) heterozygous carriers; canonical splice 30/47 (64%), NMD 14/32 (44%). · Disruption profiles recover disease mechanisms…; RNA-seq validation of EVEE splice and NMD…; Illustrative case

  5. SuggestedEVEE scores track clinical penetrance among LDLR carriers and graded disease-association enrichment in FinnGen R12, including ClinVar-unclassified variants, and nominated six rare variants with enough ACMG/AMP evidence to support ClinVar submission.

    Mayo Clinic Tapestry, familial hypercholesterolemia-manifesting (n = 131) vs presymptomatic (n = 16) carriers: AUROC 0.91 (95% CI 0.84–0.96) vs 0.70 AlphaMissense, 0.65 CADD. FinnGen genomic inflation λGC 1.33 for EVEE ≥ 0.95 (n = 958), 1.24 for ≥ 0.80, 1.00 for allele-frequency-matched EVEE < 0.80 controls, 1.17 for unclassified ≥ 0.95 variants (n = 495). · Predictions track clinical penetrance and…; Mayo Clinic Tapestry cohort — LDLR…; FinnGen R12 biobank — variant-level p-value…; ACMG/AMP structured-evidence pipeline (FinnGen…

Methods, data and reproducibility

Reference and alternate 65k bp sequences (both strands) were run through Evo 2 7B; layer-27 embedding differences at the 256 most divergent positions feed a 64 × 64 low-rank covariance probe trained on ClinVar labels with a gene-level split. Covariance probes trained on genomic annotations give disruption profiles (alternate minus reference) that Claude Sonnet 4.6 turns into mechanism hypotheses.

  • ClinVar release 2026-03-21, filtered and checked against GENCODE v49 (1,539,596 labelled, 2,716,001 unlabelled) · human (ClinVar; GRCh38 coordinates) · n = 4,252,870 variants
  • ClinVar benchmarks (≥1 star, genes ≤ 100 kb): SNVs, zero-shot indels, and a consequence-balanced deconfounded subset · human (ClinVar) · n = 833,970 SNVs; 73,961 indels; 158,616 deconfounded
  • BRCA1 and BRCA2 saturation genome editing function scores (Findlay et al.; Sahu et al.) via MaveDB · human (genes held out of ClinVar training) · n = 2,803 and 6,270 SNVs · urn:mavedb:00000097; urn:mavedb:00001225
  • TP53 deep CRISPR mutagenesis scores (Funk et al.) and LDLR cell-surface abundance / LDL uptake scores (Tabet et al.) · human (genes held out of ClinVar training) · n = 1,920 SNVs; 24,999 codon-level variants
  • CCLE cell lines with whole-exome genotypes and RNA-seq, including ovarian lines JHOS-4, JHOS-2 and A2780 · human (CCLE cancer cell lines) · n = 82 variant-carrier pairs
  • Mayo Clinic Tapestry cohort: LDLR variant carriers with familial hypercholesterolemia severity tiers · human (clinical cohort) · n = 147 unique variants (111 / 20 / 16 per tier)
  • FinnGen R12 GWAS summary statistics (REGENIE) across 2,489 disease endpoints · human (Finnish biobank) · n = ~500,000 participants
  • Mayo Clinic clinical exome cohort of patients with rheumatoid arthritis · human (patients) · n = 588 patients; 299 rare variants
Methods & tools
Evo 2 (7B), Claude Sonnet 4.6, Haiku 4.5, Opus 4.6, AlphaMissense, CADD v1.7, GPN-MSA, Nucleotide Transformer v3 (650M), AlphaGenome, Muon optimizer, MaveDB, GENCODE v49, gnomAD, ClinGen gene-disease validity
Software & environment
Versions stated for models and references (Evo 2 7B layer 27, CADD v1.7, GENCODE v49, ClinVar 2026-03-21); covariance-probe weights released; no container or requirements file described in the text.
Assay conditions & compute
Embedding extraction took ~20,000 H100-hours; the stored embeddings (4,252,870 × 2 × 512 × 4,096, bf16) total ~34 TB.
Reproducibility gaps
Mayo Clinic Tapestry and rheumatoid arthritis cohort data have no stated access route; release of the ~34 TB embedding set and of annotation-probe weights is not stated; LLM interpretations are generated on demand rather than deposited.
Limitations (6)
  • Evo 2's broad evolutionary prior may capture regulatory, epistatic or polygenic effects poorly where these lack clear evolutionary signatures, and the annotation probes only read known, labelled features. · stated by the authors
  • Because the probe is trained on ClinVar labels, within-ClinVar benchmarks give it a home advantage over comparators not trained on the same labels; the DMS, penetrance, RNA-seq and biobank analyses carry the generalization claim. · stated by the authors
  • Clinical and population validations face cohort and ascertainment limits: only 16 presymptomatic LDLR carriers in Tapestry, and a single biobank (FinnGen). · stated by the authors
  • The EVEE-based PP3 criterion is a Supporting-strength heuristic; formal calibration against observed pathogenicity odds has not been done. · stated by the authors
  • LLM-as-judge scores are only a relative measure: the interpreting model may recall well-known variants, the judge can verify claims only against ClinVar submission text, and generated explanations are hypotheses needing expert or experimental review. · stated by the authors
  • RNA-seq validation tested only variants EVEE called pathogenic (score ≥ 0.99), with no lower-scoring or competing-predictor set, so the 57% concordance has no baseline for comparison. · noted in this summary

Read this if you build or apply variant-effect predictors and want to see how probing a DNA foundation model's embeddings, rather than its likelihoods, yields both calibrated pathogenicity scores and named mechanistic evidence.

Citations
2

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISMathur '26Wu '25Xu '26Zeng '26Brixi '25Kato '26
68% semantic matchGenomics
Vihaan Mathur, Ravi Sachidanandam·bioRxiv·Mar 11, 2026· 1 citations