Loading…
Missense variant effect predictor that distills GEMME evolutionary scores into a small head over frozen protein language model embeddings.
No providers recorded yet. Browse all providers
Scoring every possible missense variant in a proteome is arithmetic that defeats the most accurate methods. A human proteome of roughly 20,000 proteins holds tens of millions of single amino-acid substitutions, and the predictors near the top of the ProteinGym leaderboard earn their accuracy through per-protein work: the alignment-based expert GEMME builds a multiple sequence alignment and an evolutionary tree for each query, while protein language models compute masked-token log-odds position by position. On one consumer laptop, GEMME finished 25 proteins in the under 30 minutes VespaG needed for all 20,000.
VespaG moves that expense from inference into training. ESM-2 embeds a sequence once; a feed-forward head then reads the 2,560-dimensional vector for a residue and emits scores for all 20 amino acids at that position in a single forward pass. The head is trained to reproduce GEMME, using 39 million substitutions drawn from about 5,000 non-redundant human proteins that GEMME had already scored. Those predicted scores act as a pseudo standard-of-truth in place of the far scarcer experimental measurements, which is what lets the training set be orders of magnitude larger than a deep-mutational-scanning-supervised one. No alignment is built and no log-odds ratio is computed at prediction time, so inference needs neither homology search nor a GPU.
The model comes from Céline Marquet and Julius Schlensok at the Technical University of Munich, with Marina Abakarova and Elodie Laine at the Laboratory of Computational and Quantitative Biology (Sorbonne University / CNRS) and Burkhard Rost. It was posted to bioRxiv in April 2024 and published in Bioinformatics that November. It is a separate artifact from VESPA, an earlier variant-effect predictor from a partly overlapping author set, with its own weights, training signal and benchmark results.
VespaG is a feed-forward network with a single 256-unit hidden layer and roughly 660,000 free parameters, sitting on frozen ESM2-T36-3B-UR50 embeddings; the backbone is never fine-tuned. Alternative heads over ProtT5-XL-U50 embeddings and deeper feed-forward, convolutional and ensemble variants were evaluated, and the one-hidden-layer network on ESM-2 embeddings won on a random 80/20 validation split. Training proteomes were built from UniProt reference proteomes for human, Drosophila melanogaster, Escherichia coli and all viruses, restricted to sequences of 25–1,024 residues and redundancy-reduced. On the ProteinGym substitution benchmark of 217 assays and 2.5 million variants, VespaG reached a mean Spearman correlation of 0.480 ± 0.021, statistically indistinguishable from its teacher GEMME and from SaProt, and ahead of zero-shot ESM-2 log-odds, VESPA and TranceptEVE L on the 189-assay organismal subset. On a 66-assay subset it was on par with PoET and comparable to AlphaMissense.
The practical draw is exhaustive precomputation. A lab can score the complete single-site mutational landscape of a whole proteome on a laptop, then use the resulting table for clinical variant triage, prioritizing candidate positions for site-directed mutagenesis, or filtering large designed libraries before assay. Because the only input is a FASTA file and the only dependency at inference is the language model encoder, it fits into pipelines that cannot assume GPU access or a maintained sequence database for alignment building.
VespaG's contribution is a recipe as much as a model: distilling an alignment-based expert into a small regression head over language model embeddings buys near-expert accuracy at a fraction of the compute, and the same recipe transfers to newer backbones. Its limits are equally concrete. It handles single amino-acid substitutions only, not insertions or deletions; its accuracy is bounded by the teacher it imitates; and it gives no mechanistic account of why a variant is damaging, where GEMME at least exposes the conservation pattern behind a score.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.