AlphaGenome Variant Impact score fails to beat SpliceAI on independent benchmarks
Start here
A preprint posted on bioRxiv on October 18, 2026 benchmarks DeepMind's AlphaGenome Variant Impact (AVI) score against SpliceAI and Panglion using data excluded from the developers' own evaluations. On 87,666 post-release ClinVar rare single-nucleotide variants, AVI achieved an intronic AUROC of 0.957, statistically indistinguishable from SpliceAI (difference −0.005) and Pangolin (−0.006), and trailed SpliceAI by 0.070 on deep intronic variants. AVI also ranked splice-altering variants below raw SpliceAI on six experimental splicing assays, and flagged benign evidence for only 24% of benign variants, versus 90% for SpliceAI. The findings challenge AlphaGenome's claimed gains over other genome-wide scores and may affect how clinicians interpret its output in variant classification.
Written by a model from the reports · updated 4 hours ago
Coverage timeline
Follow the coverage to see the different angles.
- biorxiv.orgPickBenchmarking the AlphaGenome Variant Impact score against SpliceAI on post-release ClinVar classifications and experimental splicing assays
AlphaGenome's claimed gains over other genome-wide scores do not hold on ground truth excluded from its developers' analyses. On 87,666 post-release ClinVar rare single-nucleotide variants, AVI reached an intronic AUROC of 0.957, indistinguishable from SpliceAI (difference −0.005) and Pangolin (−0.006), and it trailed SpliceAI by 0.070 on deep intronic variants. It also ranked splice-altering variants below raw SpliceAI on six splicing assays and gave benign evidence to only 24% of benign variants, against 90% for SpliceAI.
Heat trend for this story
Heat now 4·Peak in the comparable set 5 (Oct 18 12:00)·Change over 24h in the comparable set –
The trend compares only the participants observed in full throughout, so its set can be smaller than the heat figure itself. Move the pointer or click the chart for an hour's heat; the left and right arrow keys work too.