Stanford's Kundaje Reviews DeepMind's AVI Score: Competitive, Not Revolutionary, and Not "AI Slop"

On September 18, Stanford computational genomicist Anshul Kundaje posted a long thread responding to Steven Salzberg's criticism that DeepMind's AlphaGenome Atlas exemplifies "AI slop sweeping genomics," and systematically reviewed the AVI (AlphaGenome Variant Impact) score. He concluded the work is solid and competitive but not a disruptive breakthrough, and that users must understand its positioning and limitations. The next day he responded to jeffvierstra's doubts about the value of AlphaGenome distillation, reiterating that it is far from "AI slop" and essentially at the frontier, with its real value to be proven by community use in variant prioritization.

Confirmed

• AVI is a genome-wide variant-impact prioritization score that ranks all 9 billion possible variants on a PHRED scale: AVI 10 corresponds to the top 10%, 20 to 1%, 30 to 0.1%, 40 to 0.01%. Kundaje stressed it is not a calibrated probability of pathogenicity and should not be read as one.
• AVI trains a classifier using variant-frequency proxies: common variants with frequency ≥0.1% serve as (noisy) neutral/benign proxies, while rarer variants are enriched for deleterious effects. This negative-selection-proxy paradigm is not new; the most direct comparator is the widely used classic score CADD.
• Feature composition: AlphaGenome's predicted reference-vs-alternate effects across thousands of tracks are compressed into 10 cross-cell-type features (covering chromatin accessibility, transcription factors, transcription, splicing, polyadenylation, 3D contacts), then combined with AlphaMissense protein-coding variant scores, 2 conservation scores, 3 loss-of-function annotations and an indel indicator — 18 features per variant fed into a linear hypernetwork ensemble producing a single impact score.
• Evaluation: AVI's authors benchmarked against other SOTA variant prioritization scores on established rare- and common-variant benchmarks, with mixed results — slightly ahead on some, notably better on others. The original AlphaGenome paper's benchmark panorama shows broad improvements over prior SOTA, but rarely dramatic leaps on specific tasks; benchmarks typically involve perturbation experiments, synthetic libraries, and silver-standard causal variant sets.
• Interpretability is a highlight: AVI provides functional hypotheses (e.g., binding or accessibility changes) for why a variant is prioritized, and researchers can apply interpretability methods directly on AlphaGenome to drill down into which sequence features (e.g., motifs) a variant disrupts and in which cell types and readouts.

Unconfirmed

• ctokheim questioned in a reply that although noncoding interpretation is AlphaGenome's main selling point, several of its protein-coding variant interpretation results raise suspicion of benchmark optimization for hype; this remains an individual researcher's concern without systematic evaluation.

Kundaje's critiques and reservations

• Compressing context-specific variant effects into a generic, disease-agnostic score is a lossy common practice: different diseases map to different cellular contexts, losing substantial information. He proposed learnable phenotype scores mapping molecular effects to disease-specific priorities, noting AI agents and model interpretation create opportunities to reason directly over multi-dimensional scores.
• On hype: there is strong consensus among the authors and the field that AlphaGenome and similar models remain very far from "solving" the regulatory code; the paper is candid about limitations, and while DeepMind's marketing lacks nuance, it is within the range common in the industry.
• He argued supervised sequence-to-omics models' core strength is learning context-specific regulatory logic, whereas DNA language models not conditioned on cellular context struggle with noncoding variant prediction.

Why it matters

AVI is DeepMind's push toward clinical-grade variant interpretation, and its benchmark performance and interpretability bear directly on tool selection for variant prioritization in rare-disease genetic diagnosis. Kundaje's thread offers the community a balanced assessment from a top independent expert — rebutting the "AI slop" dismissal while cautioning against marketing hype — and leaves ctokheim's benchmark-optimization dispute for the community to test.

2026-09-18 ~ 2026-09-19 · 35 related posts

Full story(3 episodes)→

• Episode 1: Stanford's Kundaje Reviews DeepMind's AVI Score: Competitive, Not Revolutionary, and Not "AI Slop"(2026-09-18, 35 posts)
• Episode 2: Benchmarks Question Distilled AlphaGenome's Performance(2026-09-19, 3 posts)
• Episode 3: Researchers Defend Distilled Genomic Models, See Room for Fine-tuning(2026-09-19, 2 posts)

• [source] 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 Is Competitive With SOTA on Established Variant Benchmarks — anshulkundaje · 2026-09-18
• [source] 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
• Aggregate variant scores persist as a convenience, not a principled choice — anshulkundaje · 2026-09-18
• Clinical variant prioritization guidelines focus almost only on coding variants — anshulkundaje · 2026-09-18
• Context-specific non-coding effects vs coding-only clinical guidelines — anshulkundaje · 2026-09-18
• Cell-context specificity is why DNA LMs struggle with non-coding variants — anshulkundaje · 2026-09-18
• Supervised sequence-to-omics models excel at context-specific regulatory predictions — anshulkundaje · 2026-09-18
• Collapsing context-specific variant scores into one universal score loses key information — anshulkundaje · 2026-09-18
• Learning phenotypic scores to map molecular effects into disease-relevant prioritization — anshulkundaje · 2026-09-18
• Aggregate variant scores like CADD still useful when paired with context-specific effects — anshulkundaje · 2026-09-18
• Genomics researcher: nobody has 'solved' variant prioritization, but old-score-only views are outdated — anshulkundaje · 2026-09-18
• Researchers argue aggregate variant effect models already enable ALS therapeutic discovery — anshulkundaje · 2026-09-18
• [source] Defending AlphaGenome: 'not AI slop' — predictions are cutting edge, researcher argues — anshulkundaje · 2026-09-19
• Genomics AI researcher: bigger models held back by data quality, small models still best for local signals — anshulkundaje · 2026-09-19
• Stanford's Anshul Kundaje: gene expression AI models capped by data, not scale — anshulkundaje · 2026-09-19
• Researchers clash over whether AlphaGenome's coding-variant results are benchmark-gamed — anshulkundaje · 2026-09-19
• Commenter questions how generalizable AlphaGenome's precomputed AVI scores really are — anshulkundaje · 2026-09-19

1 near-duplicate retellings: anshulkundaje