VEFill

Deep mutational scanning measures the functional consequence of thousands of amino acid substitutions in one experiment, and MaveDB now lists over 2,600 public datasets, more than 1,100 of them for human proteins. Almost none of those maps are complete. Library construction, sequencing depth, and assay dropout leave holes scattered across the substitution grid, and a variant with no measured score is a variant that cannot be used in clinical interpretation, protein engineering, or any integrative analysis that assumes a full matrix. Existing imputation methods largely fill those holes one protein at a time, or lean on structural annotations that are unavailable for most targets.

VEFill was built by Polina Polunina, Wolfgang Maier, and Alan Rubin across the Walter and Eliza Hall Institute of Medical Research, the University of Freiburg, and the University of Melbourne. It is a gradient-boosting regressor trained once on Human Domainome 1 — a uniformly assayed collection of site-saturation mutagenesis experiments across 522 protein domains — and then applied without retraining to domains and proteins it has never seen. Its input features combine ESM-1v sequence embeddings, EVE evolutionary conservation scores, amino acid substitution matrices, and physicochemical descriptors, making it a thin, interpretable predictive head over a frozen protein language model rather than a new representation learner.

The model documented here is the general cross-protein version, the one that transfers between targets; the paper also reports narrower per-protein models fitted to individual datasets for limited-data settings.

#Key Features

• Trained once, applied across proteins: A leave-protein-out design holds out all variants of 10% of proteins, so reported accuracy reflects transfer to unseen domains rather than interpolation within a familiar one.
• Language model embeddings carry the signal: Substitution matrices, physicochemical descriptors, or conservation scores alone each land below R² = 0.15; adding ESM-1v embeddings lifts performance to R² = 0.42, and combining them with positional priors reaches the best result.
• A cheap two-feature alternative: A reduced model using only ESM-1v embeddings and per-position mean DMS scores performs comparably to the full feature set, and scales to all 521 usable domains.
• No structural annotations required: Dropping structure as an input is what lets the model run on any domain with a sequence, at the cost of weaker predictions at structurally sensitive positions such as proline substitutions.
• Honest assay boundary: Generalization was tested separately on stability-based and activity-based external assays, and the gap between them is reported rather than averaged away.

#Technical Details

VEFill is implemented with LightGBM 4.5.0, tuned by Bayesian hyperparameter optimization through Optuna, with all training features held in a PostgreSQL schema for reproducibility. Human Domainome 1 measures mutational stability with an abundance-based protein fragment complementation assay under standardized cellular conditions; after excluding one artificial sequence absent from MaveDB, 521 domains remained. A feature-complete subset of 140 domains covering 136,854 mutations was used for the full-feature models, while the reduced two-feature models were trained on all 521 domains and 562,208 mutations.

The best cross-protein configuration — every feature except the ESM-1v difference vector — achieved R² = 0.64 and Pearson r = 0.80. Tested on eight full-length proteins from MaveDB absent from training, performance held up on stability assays (PRKN R² = 0.799, ASPA R² = 0.702, PTEN R² = 0.550) and fell sharply on activity assays (PTEN R² = 0.465, TP53 R² = 0.473 and −0.125, TPK1 R² = 0.158, CALM1 R² = 0.152). Error analysis shows a U-shaped RMSE profile across the score range: predictions are most accurate near neutral, where training data are densest, and compress toward the mean at both the deleterious and high-activity extremes.

#Applications

The immediate use is completing published DMS maps so that downstream variant interpretation is not silently truncated by missing measurements. Beyond gap-filling, the model supports mutation prioritization — ranking which substitutions merit a targeted follow-up assay — and the design of smaller mutational libraries, which matters most for groups whose experimental budget cannot cover a full site-saturation panel.

#Impact

Published in Molecular Systems Biology after preprinting on bioRxiv, VEFill argues that DMS imputation is best treated as one cross-protein learning problem on a uniformly assayed corpus rather than as a per-dataset curve fit. The code is MIT licensed on GitHub with the trained LightGBM model committed to the repository, a frozen release archived on Zenodo, and processed datasets, a database backup, and pretrained models deposited separately. Two limits define its scope: the training corpus is a stability assay, and transfer to activity-based readouts is materially weaker; and the strongest configurations use per-position mean DMS scores, which exist only where some measurements are already available, so true zero-shot prediction without positional context remains an open problem for functionally complex proteins.