AlphaGenome Enhances Personal Gene Expression Prediction but Retains Key Limitations

Recent years have seen rapid growth in sequence-to-function (S2F) models that predict diverse molecular readouts from DNA sequence alone. However, these models have faced criticism for their limited accuracy in predicting individual-level gene expression. AlphaGenome, a state-of-the-art S2F model, excels across a range of sequence-based prediction tasks, but its utility for personal expression prediction remains incompletely assessed. Here, we evaluate AlphaGenome's ability to predict tissue-specific, personal gene expression and find that it significantly outperforms its predecessor, Enformer. Across 1,000 genes and 11,471 eligible gene-tissue pairs, AlphaGenome significantly outperforms Enformer for 3,103 pairs, whereas Enformer significantly outperforms AlphaGenome for 828 pairs, yielding a win ratio of 3.75. This imbalance is supported by a global permutation test (P = 2.0e-4), with a median correlation difference of 0.046. Nevertheless, pretrained AlphaGenome substantially underperforms Elastic Net and Random Forest trained on individual-level data. These findings demonstrate that AlphaGenome better captures individual-specific gene expression variation than Enformer, while highlighting substantial room for improvement in sequence-based prediction of personal gene expression.