The AlphaGenome Debate: Breakthrough or Overhyped Benchmark?

Episode 1 · Kundaje reviews DeepMind's AVI score: competitive with SOTA but not a breakthrough, and far from "AI slop" (2026-09-18, 35 posts)

On September 18, Anshul Kundaje, Professor of Computational Genomics at Stanford, posted a long thread responding to Steven Salzberg's criticism that DeepMind's AlphaGenome Atlas exemplifies "AI slop sweeping through genomics." He offered a systematic technical review of DeepMind's AVI (AlphaGenome Variant Impact) scores, concluding that the work is solid and competitive, but not a disruptive breakthrough—users need to understand its positioning and limitations.

Confirmed

• AVI is a genome-wide variant impact prioritization score that uniformly ranks all 9 billion possible variants with PHRED scaling: AVI 10 corresponds to the top 10%, 20 to the top 1%, 30 to the top 0.1%, and 40 to the top 0.01%. Kundaje specifically clarified that it is not a calibrated probability of "whether a variant is pathogenic" and should not be interpreted as a probability.
• Technically, AVI trains a classifier using common/rare variant frequencies as proxies: common variants with frequency ≥0.1% are treated as (noisy) neutral/benign proxies, while rarer variants are enriched for potentially deleterious effects. This paradigm of using negative selection signals as proxies is not new—its most direct counterpart is the widely used classic score CADD.
• Feature composition: reference vs. alternate variant effects predicted by AlphaGenome across thousands of tracks are compressed into 10 cross-cell-type features (covering chromatin accessibility, transcription factors, transcription, splicing, polyadenylation, 3D contacts, etc.), then combined with AlphaMissense protein-coding variant scores, 2 conservation scores, 3 protein loss-of-function (LoF) annotations, and an indel indicator—18 features per variant in total—fed into a linear hypernetwork ensemble to produce a single impact score.
• Evaluation conclusions: the AVI paper's authors made formal comparisons against other SOTA variant prioritization scores on multiple established benchmarks for both rare and common variants, with mixed results—slightly ahead on some metrics, notably better on others. The benchmark landscape in the original AlphaGenome paper also showed it generally outperforming existing SOTA, though not a huge leap on most specific tasks.
• Interpretability is a highlight: AVI can offer functional hypotheses for why a variant is prioritized (e.g., binding or accessibility changes), and researchers can apply interpretation methods directly on the AlphaGenome model to drill down into which sequence features (e.g., motifs) a variant disrupts and which readouts in which cell types it affects.

Kundaje's criticisms and reservations

• Compressing context-specific variant effects into a generic, disease-agnostic score is a common lossy practice: different diseases and traits are associated with different cellular contexts, so substantial information is lost. He proposes learnable phenotype scores that map molecular effects to disease-specific priorities, and argues that AI agents and model interpretation techniques create opportunities to reason directly over multi-dimensional scores.
• On hype: the paper's authors and the field at large agree that AlphaGenome and similar models remain far from "solving" the regulatory sequence code; the paper is candid about improvements and limitations. DeepMind's marketing language lacks nuance but stays within reasonable bounds compared to typical industry promotion.
• He believes the core advantage of supervised sequence-to-omics models lies in learning context-specific regulatory logic, whereas DNA language models not conditioned on cellular context struggle with non-coding variant prediction.

Why it matters

AVI is DeepMind's attempt to push the AlphaGenome family toward clinical-grade variant interpretation; its evaluation results and interpretability directly inform the selection of variant prioritization tools for rare disease genetic diagnosis. Kundaje's long post provides the community with a balanced assessment from a top third-party expert—pushing back on the excessive dismissal of "AI slop" while cautioning against being swayed by marketing.

• Computational genomics expert rebuts 'AI slop' criticism of DeepMind's AlphaGenome — anshulkundaje · 2026-09-18
• AlphaGenome Benchmarks: Better Than Prior SOTA Overall, But Not a Giant Leap in Most Cases — anshulkundaje · 2026-09-18
• AlphaGenome Paper Is Candid About Limits; the Marketing Hype Is Still Within Bounds — anshulkundaje · 2026-09-18
• Top Researcher: AlphaGenome Is Very Far From Solving the Gene Regulation Code — anshulkundaje · 2026-09-18
• What the New AVI Score Actually Is: A Genome-Wide Ranking, Not a Calibrated Pathogenicity Probability — anshulkundaje · 2026-09-18
• How AlphaGenome Compresses Thousands of Variant Effect Tracks Into 10 Features — anshulkundaje · 2026-09-18
• AlphaGenome Reduces Thousands of Variant Effect Tracks to 10 Per-Variant Features — anshulkundaje · 2026-09-18
• AVI Scores Each Variant Using 18 Features: AG, AlphaMissense, Conservation, LoF — anshulkundaje · 2026-09-18
• AVI Architecture: 18 Features Per Variant Fed Into a Linear Hypernetwork Ensemble — anshulkundaje · 2026-09-18
• DeepMind's AVI Score Uses Common-vs-Rare Variant Frequency as a Noisy Neutrality Proxy — anshulkundaje · 2026-09-18
• AVI Follows an Established Paradigm, With CADD as Its Most Direct Comparison — anshulkundaje · 2026-09-18
• AVI Scores Are PHRED-Scaled and Traceable to Contributing Features — anshulkundaje · 2026-09-18
• Interpretability Methods Run Directly on AlphaGenome to Pinpoint Disrupted Motifs — anshulkundaje · 2026-09-18
• AVI Score Offers Interpretable Functional Hypotheses for Variant Prioritization — anshulkundaje · 2026-09-18
• AVI Paper Benchmarks Against SOTA Scores With Interpretability as a Highlight — anshulkundaje · 2026-09-18
• AVI Is Competitive With SOTA on Established Variant Benchmarks — anshulkundaje · 2026-09-18
• AVI Variant Impact Score Matches SOTA on Benchmarks, Though Evals Are Biased — anshulkundaje · 2026-09-18
• Genomics researcher defends AI variant prioritization paper: validation already sufficient — anshulkundaje · 2026-09-18
• Stanford prof: Google overstated AlphaGenome claims, but calling it 'AI slop' is nonsense — anshulkundaje · 2026-09-18
• Stanford professor fires back: calling genomic variant effect scores 'AI slop' is nonsense — anshulkundaje · 2026-09-18

15 more related posts →

Episode 2 · Benchmarks Question Distilled AlphaGenome's Performance (2026-09-19, 3 posts)

Researcher Jeff Vierstra's benchmarks show that once distillation effects are removed, AlphaGenome ranks only mid-tier across cell types, with chromBPNet best at cell selectivity. He also questioned Google's approach of releasing 9 billion predictions as a distilled model.

• Genomics model comparison: chromBPNet leads on cell selectivity, AlphaGenome middling without distillation — jeffvierstra · 2026-09-19
• AlphaGenome Variant Performance Looks 'Middle of the Pack' Once Distillation Is Stripped Out, Researcher Finds — jeffvierstra · 2026-09-19
• Researcher: AlphaGenome distillation barely differs — variant correlations identical across models — jeffvierstra · 2026-09-19

Episode 3 · Researchers Defend Distilled Genomic Models, See Room for Fine-tuning (2026-09-19, 2 posts)

Responding to criticism of distilled AlphaGenome-style models, Stanford's Anshul Kundaje said pre-distilled base models remain high quality and support fast fine-tuning, while acknowledging cell-type specificity remains a challenge.

• Genomics AI researcher: distillation gains are marginal, better losses and data sampling still ahead — anshulkundaje · 2026-09-19
• Genomics AI debate: researcher argues pre-distilled base models leave big fine-tuning headroom — anshulkundaje · 2026-09-19