Alignment-free missense variant effect predictor that combines ProtT5-predicted conservation, BLOSUM62 and language model log-odds.
No providers recorded yet. Browse all providers
The classic way to judge a missense variant starts with a multiple sequence alignment: gather the protein's relatives, see which positions evolution has kept fixed, and flag substitutions at those positions as likely to hurt. That works, but it costs an alignment search per protein and fails outright for residues with too few homologs to align. VESPA (Variant Effect Score Prediction without Alignments) asks whether a protein language model has already absorbed enough of that signal to skip the alignment entirely.
It answers in two steps. First, a small convolutional network reads frozen per-residue embeddings from ProtT5 and predicts the nine-class conservation score that ConSurf-DB would assign from a full alignment; this component, ProtT5cons, came within a whisker of the alignment-based ConSeq (two-state MCC 0.596 versus 0.608). Second, a logistic regression ensemble combines that predicted conservation with the BLOSUM62 substitution score and ProtT5's masked-reconstruction log-odds for the specific substitution, and outputs a probability that the variant has an effect. A lighter variant, VESPAl, drops the log-odds input, which removes the slowest step because log-odds need one masked forward pass per residue while embeddings need one pass per protein.
VESPA was developed by Céline Marquet, Michael Heinzinger and colleagues in Burkhard Rost's group (Rostlab) at the Technical University of Munich, and published in Human Genetics in December 2021 after an earlier Research Square preprint. It is the predecessor of VespaG, a separate 2024 model from an overlapping author set that replaces this feature-engineered ensemble with a head over ESM-2 embeddings trained to imitate GEMME.
ProtT5cons is a two-layer CNN (window size 7, about 231,000 parameters) trained on ConSurf10k, 10,507 structurally resolved proteins redundancy-reduced to 25% pairwise identity, with the ProtT5-XL-U50 encoder frozen throughout. The effect predictor is an ensemble of ten balanced logistic regressions, one per cross-validation split of Eff10k, the SNAP2 development set of 100,737 binary variant annotations from 9,594 proteins; the ten outputs are averaged into a score in [0, 1], with 0.5 as the effect threshold. On DMS39, 39 deep mutational scans assembled for DeepSequence (135,665 variants), VESPA reached a mean absolute Spearman correlation of 0.51, against 0.53 for GEMME, 0.50 for DeepSequence and 0.49 for ESM-1v, with VESPAl at 0.47. The authors stress that no method outperformed all others consistently or with statistical significance. On four binarized human DMS sets (BRCA1, PTEN, TPMT, PPARG), VESPAl reached an MCC of 0.405 against 0.204 for SNAP2. VESPAl scored all substitutions in the roughly 20,000-protein human proteome in 40 minutes on one Nvidia Quadro RTX 8000.
VESPA suits anyone who needs variant effect estimates across many proteins at once: clinical
and population geneticists triaging missense variants, groups annotating newly sequenced
proteomes, and protein engineers pre-screening positions before mutagenesis. It installs as the
vespa-effect Python package, and precomputed VESPAl predictions for the human and
Drosophila melanogaster proteomes are archived on Zenodo, so many users can look up scores
without running anything. Scope is limited to single amino-acid substitutions; insertions,
deletions and combinations of mutations are not modeled.
VESPA was an early demonstration that frozen language model embeddings could stand in for alignments in variant effect prediction and land in the same range as alignment-based methods, at a fraction of the compute. It has become a routine baseline in later variant effect work and led directly to VespaG. Its limits are candid in the paper: accuracy drops for residues outside alignable regions, the training labels are sparse binary annotations rather than quantitative measurements, and the full VESPA model still needs a GPU with substantial memory for the log-odds step.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.