MELD learns the wild-type-to-mutant difference directly, lifting ClinVar AUROC from 0.800 to 0.986 with 1.9 million parameters
Synopsis
MELD puts the score difference between wild-type and mutant sequence directly into the supervised objective, instead of training on absolute signal and comparing two sequences after the fact. On 1.57 million ClinVar variants it raised AUROC from 0.800 to 0.986 against a same-data, same-backbone post-hoc baseline, and on a shared benchmark its 1.9 million parameters outperformed frozen foundation models with 3 to 1,300 times more parameters.
Interpretation
The training target for variant-effect prediction changes from absolute sequence signal to the difference between wild-type and mutant, which the model learns directly. Sequence-to-function models such as Enformer, Borzoi and AlphaGenome are trained on absolute genomic tracks and score variants post hoc by contrasting reference and alternate alleles, so that difference never enters the training loss. On the held-out test set of the ClinVar standard 1 kb benchmark, MELD reached AUROC 0.986 and AUPRC 0.934, against 0.800 and 0.700 for the same-data, same-backbone post-hoc baseline.
In practice, wild-type and mutant windows share one compact CNN-Transformer encoder, and the variant site is marked by a mask that applies a learned boost to the mutant branch. Encoded representations are pooled globally and in mutation-centered neighborhoods at several radii; at each scale the mutant-minus-wild-type scalar score difference is weighted by an adaptive amplification factor, averaged across scales, and passed through a sigmoid to give a pathogenicity probability. This reference model has 1,917,059 trainable parameters, uses 1,025 bp variant-centered windows, and reached a five-seed mean AUROC of 0.9859±0.0007 on the held-out test set.
Perspective
The demonstrated scope is binary pathogenicity classification on the ClinVar standard 1 kb benchmark and continuous effect regression on 466 held-out BRCA1 saturation genome editing variants; external foundation-model comparisons use a fixed shared subset of 200k training, 20k validation and 20k test variants, with frozen backbones and only prediction heads trained.
Quantitative evaluation covers one gene and 466 test variants, and absolute correlations are moderate; on scoreable single-nucleotide variants, the integrative CADD v1.7 score reached AUROC 0.989 on ClinVar against 0.982 for MELD; ClinVar labels carry ascertainment bias and label noise, and a few duplicated sequence windows were retained.