MELD Model Predicts Variant Effects Efficiently
A 1.9-million-parameter model trained on wild-type versus mutant differences outperforms larger frozen foundation models on a shared benchmark.
Updated on Nov. 9, 2026 in Life Sciences

Researchers have introduced MELD, a model that predicts the functional consequences of single nucleotide variants by training on differences between wild-type and mutant sequences. MELD reached an AUROC of 0.986 on ClinVar, which contains 1.57 million variants, and outperformed larger frozen foundation models on a shared variant-effect benchmark.
Why it matters
Experimental characterization cannot scale to millions of observed variants, so computational methods that predict variant effects are needed to interpret genetic data. MELD's results suggest that placing the mutant-versus-wild-type comparison directly inside the training objective can yield strong performance with a comparatively small model.
The benchmark results compare MELD's AUROC with a same-data baseline and with larger foundation models on ClinVar and a shared benchmark, while frozen MELD reached a Spearman ρ of 0.42 on BRCA1 saturation genome editing data. The extraction does not report performance on variant classes or genomic regions outside these evaluations.
The details
MELD combines a shared CNN-Transformer encoder, a mutant-specific mutation-mask boost, multiscale mutation-centered pooling, and a learned adaptive amplification factor. In low-data settings, MELD trained on 50,000 variants reached an AUROC of 0.903, compared with 0.796 for the in-silico mutagenesis baseline trained on 1.48 million variants.
The Big Picture
MELD builds on the in-silico mutagenesis approach of scoring variants by comparing predicted outputs for mutant and wild-type sequences, and the extraction reports that it outperforms that baseline on ClinVar with the same data and backbone. The model's gains over larger frozen foundation models such as AlphaGenome and Nucleotide Transformer on a shared benchmark position it as an alternative to scaling model size alone.
Tools that predict variant effects from sequence could help researchers prioritize which genetic changes to study experimentally. The extraction does not describe any clinical or commercial deployment of MELD.
The takeaway
Models that directly learn the difference between mutant and wild-type sequences may offer a more efficient route to variant interpretation than simply scaling up general-purpose genomic models. Readers following genomic data tools can watch how these benchmark results hold up in independent evaluations.
Further reading
For related coverage, see the Life Sciences section.
Source note: This article includes information reported by Biorxiv.







