Genomic Foundation Models in 2026: What Survives a Held-Out Test Set (rewire.it)

A recent analysis of genomic foundation models reveals a mixed landscape of performance and capabilities, suggesting that while some models excel in specific tasks, others fall short under rigorous independent evaluation. Notably, Evo 2 and AlphaGenome have made significant advances in predicting variant effects, surpassing traditional task-specific tools in noncoding variants. However, the findings also underscore that many models underperform compared to simple linear baselines in tasks like perturbation prediction. The key takeaway is the importance of a thorough validity ledger, which evaluates model performance against held-out test sets, contrasting the marketing-driven capability ledger that often inflates model capabilities without robust benchmarks. Central to understanding the state of genomic models is the introduction of GENEB, a comprehensive diagnostic benchmark that assessed 40 genomic models across 100 tasks. It reveals variability in model performance across functional categories and highlights that scale and architectural design often outweigh sheer parameter counts in determining effectiveness. As the field evolves, researchers are urged to prioritize transparency in model evaluations, recognize the complexities of genomic data representation, and navigate the challenges of benchmarking to make genuinely clinical advancements in genomics.