Developed by Polina Polunina and Wolfgang Maier of the University of Freiburg together with Alan Rubin of the Walter and Eliza Hall Institute of Medical Research, VEFill is built on LightGBM, a gradient boosting framework well suited to structured, high-dimensional tabular data. Rather than relying on a single source of information, the model integrates a rich set of biologically informed features: evolutionary conservation scores from the EVE generative model, amino acid substitution matrices such as BLOSUM62 and PAM250, physicochemical descriptors of the wild-type and variant residues, and contextual sequence embeddings from ESM-1v, a transformer-based protein language model trained on millions of natural sequences. The authors trained and optimized the model using Bayesian hyperparameter search and stored all underlying data in a structured PostgreSQL database to ensure reproducibility.

The training ground for VEFill was the Human Domainome 1 dataset, a standardized collection of stability measurements generated with an abundance-based protein fragment complementation assay across 522 human protein domains, using site-saturation mutagenesis to introduce every possible amino acid substitution. From this resource, the team assembled a feature-complete subset of 140 domains comprising 136,854 mutations for the full model, and used the broader set of 521 domains, totaling 562,208 mutations, to train reduced-feature versions that depend only on ESM-1v embeddings and positional mean scores.

The performance figures are striking. In a leave-protein-out evaluation, where the model must predict variant effects for protein domains entirely absent from training, the best configuration achieved a coefficient of determination of 0.64 and a Pearson correlation of 0.80. Feature ablation experiments revealed a clear hierarchy of information: models relying solely on evolutionary scores, substitution matrices, or biochemical descriptors performed poorly, with R-squared values below 0.15. Adding ESM-1v embeddings lifted performance substantially, and combining the embeddings with the mean experimental score per position proved even more powerful, indicating that learned sequence representations and data-driven positional priors capture complementary aspects of mutational tolerance.

Perhaps the most practically important finding concerns data scarcity. In controlled subsampling experiments on 28 high-quality datasets, VEFill consistently outperformed a battery of existing imputation methods, including Envision, FactorizeDMS, AALasso, and nearest-neighbor approaches, once at least 20 percent of variants had been experimentally measured. Below that threshold, predictions became markedly harder, and the authors are candid that true zero-shot prediction without any positional context remains challenging, particularly for functionally complex proteins. This matters because a survey of 841 public score sets from the MaveDB repository showed that fully complete datasets are vanishingly rare, with only 44 out of 841 achieving complete coverage of all possible substitutions. The vast majority of real-world experiments therefore fall squarely within the regime where VEFill delivers its greatest benefit.

The team also probed how few measurements are actually needed per position. When per-protein models were trained with a restricted number of substitutions per site, accuracy rose rapidly and began to plateau at roughly four mutations per position. Intriguingly, models trained exclusively on five carefully chosen amino acids, histidine, glutamic acid, asparagine, isoleucine, and glycine, each representing a distinct physicochemical class, matched the performance of models trained on randomly selected substitutions at equivalent coverage. This suggests that sparse, information-efficient mutational libraries could be designed deliberately, prioritizing a small representative set of substitutions per site without sacrificing predictive power.

Noise ceiling analyses added an important dose of realism. Because experimental variability imposes a fundamental upper bound on achievable accuracy, the researchers simulated replicate experiments using reported measurement uncertainties and found that estimated ceilings ranged from 0.68 to 0.99 across domains. VEFill’s per-protein correlations, between 0.52 and 0.91 under an 80/20 split, sat consistently below but generally close to these limits, indicating that much of the remaining discrepancy between predicted and observed scores reflects measurement noise rather than model failure. Error profiles were lowest for near-neutral variants, where data are densest, and predictions at the extreme deleterious and high-activity tails should be read as conservative approximations rather than precise estimates.

Generalization beyond the training data showed both promise and limits. When the cross-protein model was applied to eight full-length proteins from independent MaveDB datasets, including Parkin, PTEN, aspartoacylase, calmodulin, TPK1, and TP53, it performed substantially better on stability-based assays than on activity-based ones, a mismatch the authors attribute to differences between training and test assay modalities and the complexity of cellular phenotypes. A lightweight two-feature version using only ESM-1v embeddings and positional mean scores performed comparably to the full model in most settings, offering a practical alternative when evolutionary scores or other annotations are unavailable. Analysis of protein family composition revealed that a few Pfam families dominated the training data, but retraining on a Pfam-unique subset increased error only modestly, suggesting the findings are robust rather than inflated by family-level redundancy.

Fine-grained evaluations exposed the model’s blind spots in biologically meaningful ways. Proline substitutions consistently produced elevated errors, likely because their unusual conformational constraints disrupt secondary structure in ways that sequence-based features do not fully capture. In the TRIM44 zinc finger domain, histidine-to-cysteine mutations were poorly predicted when entire positions were held out but accurately recovered when single variants were withheld, hinting at context-dependent chemistry, such as zinc coordination, that only fine-grained positional information can resolve. The authors suggest that future versions could incorporate explicit structural features, cross-species transfer learning, or systematic assay harmonization to close these gaps.

Overall, VEFill arrives at a moment when the variant effect field is scaling rapidly, with MaveDB now listing more than 2,600 public datasets and over 1,100 for human proteins. By providing an interpretable, scalable tool that turns partially complete mutational maps into denser, more usable resources, the study lowers the experimental barrier for variant prioritization, protein engineering, and the clinical interpretation of missense mutations. The code and trained models have been released openly, inviting the community to apply sparse-library design strategies and, ultimately, to build the comprehensive Atlas of Variant Effects that the field has long envisioned.

Subject of Research: Machine learning-based imputation of missing scores in deep mutational scanning datasets across human protein domains

Article Title: VEFill: accurate and generalizable deep mutational scanning score imputation across protein domains

Article References: Polunina, P. V., Maier, W., & Rubin, A. F. (2026). VEFill: accurate and generalizable deep mutational scanning score imputation across protein domains. Molecular Systems Biology, 22(6), 979-1002. https://doi.org/10.1038/s44320-026-00203-y

Image Credits: AI Generated

DOI: 10.1038/s44320-026-00203-y

Keywords: deep mutational scanning, variant effect prediction, machine learning, gradient boosting, protein stability, ESM-1v, protein language models, MaveDB, Human Domainome, imputation, missense variants, bioinformatics

The resource is called MaveMD, short for MAVEs for Medicine, and it is a new interface built on top of MaveDB, a public repository for multiplexed assays of variant effect, or MAVEs. MAVEs are experiments that measure, in a single sweep, the functional consequences of thousands to tens of thousands of variants of a gene. Instead of testing one mutation at a time, researchers synthesize a library of variant versions of a gene, introduce them into cells en masse, and use high-throughput sequencing to count how each variant performs under a functional selection. The output is a quantitative effect score for nearly every possible single-letter change in a gene, and sometimes for many multi-letter changes as well. In principle, this is exactly the kind of evidence a clinical geneticist needs: a direct measurement of how badly a patient’s variant damages the protein it encodes.

In practice, MAVE data have struggled to cross the gap between bench and clinic, and the authors of the MaveMD paper are unusually candid about why. Raw MAVE scores are assay-specific numbers with no intrinsic clinical meaning; a score of minus two in one experiment is not comparable to minus two in another. Clinical variant interpretation, governed by frameworks such as the American College of Medical Genetics and Genomics criteria, demands calibrated evidence in which functional measurements are anchored to variants of known pathogenicity. Most MAVE datasets also live in formats designed for computational biologists, with dense matrices and command-line tooling that offer little to a genetic counselor trying to interpret a patient report. And without careful curation, it can be hard to know whether a given assay actually measured the biological function relevant to the disease in question.

MaveMD tackles these barriers one by one. The interface displays calibrated evidence strength, meaning that variant effect scores are translated into a form that maps onto the functional evidence categories used in clinical classification, so a clinician can see not just a number but how much weight that number can legitimately carry. It provides intuitive visualizations that let users place a variant in the context of the full distribution of measured effects across the protein, a picture that is often far more informative than a single score. It integrates with established clinical and genomic resources so that MAVE evidence can be viewed alongside the annotations clinicians already use. And it exports ready-to-use clinical evidence, packaging the functional data in a form that can be dropped directly into variant classification workflows rather than requiring bespoke analysis for every case.

The scale of what has been assembled is substantial. MaveMD currently contains 476,076 variant effect measurements spanning 39 genes with established associations to human disease. According to the authors, this body of functional data is sufficient to enable classification of 75 percent of variants of uncertain significance within the genes covered. That figure deserves emphasis. Three-quarters of the VUS in these disease genes, variants that would otherwise sit in the ambiguous middle of a clinical report, can now be assigned functional evidence that moves them toward a definitive interpretation, whether pathogenic or benign. For the families waiting on those interpretations, the difference between an uncertain result and a resolved one can determine whether relatives get tested, whether surveillance begins, and whether a diagnosis is finally named.

The technical architecture behind the resource reflects a deliberate design philosophy. MaveMD is built as an interface layer over MaveDB rather than a separate silo, which means the underlying variant effect scores remain in the community repository where they were deposited, versioned and credited to the original experimentalists. The clinical layer adds calibration, curation, and presentation on top of that foundation. Supplementary material accompanying the paper details the per-assay and per-publication curation decisions that went into the curated datasets, an important transparency measure because the clinical value of a MAVE depends heavily on the quality and relevance of the underlying experiment. A schematic of the system’s architecture and additional examples of the web interface are also provided, showing how a user moves from a gene to a specific variant to its calibrated functional evidence.

The collaboration itself spans two continents and multiple disciplines. The work was led by researchers at the University of Washington, including the Brotman Baty Institute for Precision Medicine and the departments of Laboratory Medicine and Pathology, Genome Sciences, and Bioengineering, together with colleagues at the Collaborative Centre for Genomic Cancer Medicine, the Walter and Eliza Hall Institute of Medical Research, and the University of Melbourne. Corresponding authors Douglas M. Fowler and Alan F. Rubin anchor teams that include clinical genetics expertise alongside deep experience in multiplexed functional assays, a combination that is essential for a resource whose entire purpose is to translate between laboratory measurement and clinical decision-making. The project was supported by the National Human Genome Research Institute, including an Advancing Medical Genomics award, along with funding from the Australian Government and several foundation grants.

What makes MaveMD notable in the broader landscape of genomic medicine is its timing. The cost of genome sequencing has collapsed to the point where sequencing is no longer the bottleneck; interpretation is. Clinical laboratories report VUS rates that can exceed a third of findings in some genes, and each VUS represents not only diagnostic uncertainty but also a recurring burden: the same variant gets re-evaluated independently by different laboratories, families receive periodic letters saying the classification has not changed, and cascade testing stalls. Functional data at scale offers a way to resolve large blocks of that uncertainty systematically rather than case by case. MAVEs have been maturing rapidly over the past decade, with assays now routine for well-studied genes, but the clinical community has lacked the connective tissue needed to use them routinely. MaveMD is positioned as exactly that connective tissue.

The resource also has implications for how future functional data get generated. By providing a clear clinical destination for MAVE experiments, MaveMD creates an incentive structure for researchers to design assays that measure disease-relevant functions, calibrate their scores against known pathogenic and benign variants, and deposit results in shareable, standardized form. The authors frame the platform as supporting future data generation efforts, and the feedback loop is straightforward: the more clinically usable the interface, the more valuable new datasets become, and the more the community is encouraged to fill gaps in the 39 genes currently covered. Genes beyond the current set, particularly those where the relevant function is hard to assay in a high-throughput format, remain a challenge, and functional evidence will always be one strand of evidence among many, to be weighed alongside segregation data, computational predictions, and case reports.

Still, the trajectory is clear. Genomic medicine has spent two decades building the ability to read genomes cheaply and completely, and it is now in the harder business of understanding what the readings mean. Resources like MaveMD represent the infrastructure of that second phase: databases that do not merely store experimental results but actively translate them into the evidentiary currency of clinical practice. If the 75 percent classification figure can be extended across more of the disease-associated genome, the population of patients left holding an uncertain result will shrink correspondingly. For a field whose promise has always been precision, that is the kind of progress that matters most, measured not in sequenced bases but in diagnoses finally delivered.

Subject of Research: A clinical database interface for translating multiplexed assays of variant effect into genomic medicine evidence

Article Title: MaveMD: a functional data resource for genomic medicine

Article References: McEwen, A. E., Stone, J., Tejura, M., Gupta, P., Capodanno, B. J., Da, E. Y., Grindstaff, S. B., Moore, N., Reinhart, D., Snyder, A. E., Stergachis, A. B., Starita, L. M., Fowler, D. M., & Rubin, A. F. (2026). MaveMD: a functional data resource for genomic medicine. Genome Medicine. https://doi.org/10.1186/s13073-026-01782-z

Image Credits: AI Generated

DOI: 10.1186/s13073-026-01782-z

Keywords: MaveMD, MaveDB, variants of uncertain significance, multiplexed assays of variant effect, genomic medicine, variant interpretation, functional genomics, clinical genetics, precision medicine, genetic databases, disease-associated genes, variant classification