Response representation shapes AlphaGenome enhancer-gene prediction in human astrocytes

Abstract

Background: Sequence-to-function models predict regulatory perturbation responses, but their added value beyond established enhancer-gene predictors and across prediction tasks remains unclear. We evaluated whole-enhancer deletion and sparse variant responses in human astrocytes. Results: We analyzed 2,307 CRISPR-interference-tested relations, including 133 expression-decreasing hits. Adding deletion responses increased fully nested EGrf average precision from 0.5475 to 0.6137. The paired increment was 0.0662 (enhancer-cluster 95% interval 0.0144-0.1181). In a separate exploratory analysis, sparse Atlas responses were combined with seven-input logistic and random-forest comparators by equal-weight rank fusion. Atlas magnitude fusion exceeded absolute-deletion fusion by 0.0519 and 0.0637 in average precision. Enhancer-cluster 95% intervals were 0.0098-0.0957 and 0.0220-0.1094, respectively. Standalone Atlas magnitude reached average precision 0.5570, but its advantage over absolute deletion remained uncertain. Signed Atlas fusion performed worse than either comparator alone. Incremental value for within-enhancer target choice and in the external WTC11 MHC evaluation remained uncertain. The reported performance intervals use one-way clustering of fixed predictions; Atlas analyses were exploratory within the same cohort. Conclusions: In AstroREG, whole-enhancer deletion responses increased observed retrieval performance beyond a fully nested chromatin-based predictor. Sparse variant responses were more informative as magnitudes than as signed averages. Improved within-enhancer target choice and external transfer were not established.