The abstract is read at the publisher; the summary is JClub's.
Fine-tuning Enformer with LAMP—genome-wide artificial SNVs perturbing transcription factor motifs—still underperforms on held-out genes and people (HOGP), indicating persistent limits in learning variant effects.
- Why it matters: Post-training sequence-to-function models often fail to improve personal-genome prediction accuracy on HOGP, leaving causal variant effects underestimated and causing fine-mapping failures.
- What they did: LAMP substitutes easy-to-learn motif-perturbing aSNVs into personal genome datasets with large, consistent expression effects, then fine-tunes Enformer and analyzes attention rewiring and motif-variability effects.
- The result: Underperformance is linked to data-inefficient learning of enhancer–gene links due to strong distal priors and to sequence variability that obscures motif perturbation patterns; targeted augmentation under LAMP improves learning.
The findingWhy it mattersWhat they didThe result
AI summary of the abstract · may contain errors · not medical advice · read the paper.