Cross-task benchmarking reveals variable transferability of deep learning models across molecular quantitative trait loci

Transferability of regulatory representations from six deep learning models across nine classes of molecular QTLs

Most variants associated with complex traits are non-coding, and deep learning models predict eQTLs well, but whether they generalise to other molecular phenotypes was unknown. The authors assess pretrained regulatory features from six sequence- and annotation-based models across nine classes of molecular quantitative trait loci (xQTLs). All models perform well for eQTLs, but performance varies substantially elsewhere: chromatin accessibility, transcription factor binding and histone QTLs are well predicted, whereas chromatin looping and microRNA expression QTLs remain challenging. Foundation models, particularly AlphaGenome and Borzoi, consistently outperform the others, and tree-based ensemble classifiers further improve performance. Applied to coronary artery disease GWAS loci, AlphaGenome and Borzoi identify 337 high-confidence candidate chromatin accessibility QTLs, including 11 previously reported and 326 new candidates.

Synthesis written by Geno'X. For the full original abstract, please refer to the source publication.

Research on complex traits, with no direct application to Mendelian diagnosis. The transferable message is methodological: a regulatory effect score that performs well on eQTLs does not necessarily do so for other mechanisms, and long-range models such as AlphaGenome and Borzoi are the most robust. The 326 new candidates remain predictions, with no experimental validation in the abstract.

Analysis by Dr Thibaut Benquey