MELD: Mutation Effect Learning through Differential embeddings

MELD predicts SNV functional consequences on ClinVar (1.57M variants) with AUROC 0.986, substantially exceeding a same-data ISM baseline at 0.800 using differential supervision.

• Why it matters: Interpreting SNVs in clinical genomics is limited by the inability to experimentally test millions of variants, while common sequence-to-function models score variants only after training.
• What they did: MELD trains on the WT-versus-mutant logit difference (Delta)z = zmut − zwt within the supervised loss, using a shared CNN-Transformer encoder plus mutation-mask boosting and mutation-centered pooling.
• The result: MELD (1.9M parameters) also outperforms frozen foundation models on a shared benchmark (AUROC 0.967 vs. 0.924 for AlphaGenome; 0.936 for Nucleotide Transformer) and improves BRCA1 transfer (Spearman rho 0.45).

AI summary of the abstract · may contain errors · not medical advice · read the paper.