The underlying insight of the work is deceptively simple. Over the past two decades, researchers have released a remarkable number of computational predictors that estimate how likely a given missense variant is to damage a protein. Some of these tools rely on evolutionary conservation, comparing the affected amino acid position across related species to see whether that position tolerates change. Others build statistical models from observed mutation frequencies in large human populations, asking whether the amino acid substitution appears in healthy individuals more often than chance would allow. A newer generation of predictors is powered by protein language models, neural networks trained on massive databases of protein sequences that learn the statistical grammar of amino acid chains and can flag substitutions that violate that grammar. Each of these tools produces a score, and each has blind spots, biases and strengths. Clinicians and researchers, however, rarely know which tool to trust for which kind of variant, and the scores from different predictors can disagree dramatically. MissenseHMM treats this landscape of disagreement as information in its own right.

Technically, the method applies a hidden Markov model, a statistical framework long used in genomics to segment sequences into latent states, to the problem of variant annotation. Li and Ernst collected scores from 43 different missense pathogenicity predictors and discretized them into categories for each variant. The hidden Markov model then learns a set of hidden states, each corresponding to a characteristic combinatorial pattern of predictor outputs. In other words, rather than asking a single question such as whether a variant is pathogenic, the model asks which signature of agreement and disagreement among 43 tools the variant displays. Variants in which nearly all predictors agree on harm end up in one state; variants in which tools disagree in characteristic ways end up in others. The model learns these states and their emission patterns from the data itself, and each variant is assigned to the state that best explains its full profile of scores. In the published application, the authors learned 20 such states, and every missense variant in the reference protein annotations of the human genome, more than 77 million of them, received a state assignment. The result is a genome-wide annotation resource that compresses a 43-dimensional vector of predictor scores into a single, interpretable state label for every possible amino acid substitution at every position.

The biological content of these states turned out to be striking. The authors show that the 20 states differ not only in their predictor score patterns but in their enrichment for amino acid substitution types and for other functional annotations of the affected protein positions. Some states capture variants at highly conserved positions where even conservative amino acid swaps are flagged as damaging by most tools. Others capture variants in regions where predictors disagree, perhaps because the underlying tools weight evolutionary and structural evidence differently. The states also show distinct relationships to real clinical data. When the researchers examined how strongly each state is associated with known pathogenic variants from clinical databases, the states formed a graded spectrum from highly enriched for disease variants to strongly enriched for benign ones. Importantly, the state annotations enhanced the association with clinical pathogenic variants beyond what any individual predictor achieved on its own, demonstrating that the joint model extracts information that no single tool captures.

A second major validation came from deep mutational scanning, an experimental technique in which thousands of amino acid substitutions in a single protein are systematically generated and their effects on protein function measured in the laboratory. Deep mutational scanning datasets provide ground truth about variant effects that computational predictors try to approximate, and they have become a standard benchmark for variant effect prediction. Li and Ernst found that MissenseHMM state annotations improved the correspondence between predictor outputs and deep mutational scanning measurements across the proteins they examined. Variants assigned to the states associated with pathogenicity were more likely to have measured functional effects in the experiments, and the state framework helped reconcile cases where individual predictors had performed poorly. This convergence between computational state assignments and laboratory-measured variant effects strengthens the case that the hidden states represent genuine biological categories rather than statistical artifacts of the predictor landscape.

The method also delivers an unexpected dividend for the growing field of protein language models. Because the 43 predictors integrated by MissenseHMM include a range of language-model-based tools, the learned states provide a window into how these neural models behave across the variant space. The authors report that the state annotations give insight into the relative performances of various protein language models, revealing which models contribute complementary information and which are largely redundant with one another. As protein language models such as those based on transformer architectures increasingly dominate variant effect prediction, understanding their collective behavior becomes important both for method developers and for end users deciding which predictions to trust. MissenseHMM offers a framework for that comparison that is systematic and genome-wide rather than anecdotal.

The practical implications for medicine are considerable. Variant interpretation is a bottleneck in clinical genetics: when a patient’s genome is sequenced, the laboratory must decide which variants to report, and missense variants of uncertain significance are a persistent source of ambiguity. Guidelines from the American College of Medical Genetics allow computational evidence to support pathogenicity assessments, but the choice of predictor and the handling of discordant predictions remain contentious. An annotation resource that synthesizes 43 predictors into interpretable states, and that has been validated against both clinical variant databases and experimental mutational scanning, could give diagnostic laboratories a more robust and transparent evidence layer. Because MissenseHMM annotates over 77 million variants in advance, the state assignments can simply be looked up for any missense variant encountered in practice, avoiding the need to rerun dozens of predictors for each new case.

The work also reflects a broader trend in computational genomics toward integrative annotation. Rather than pitting predictors against one another in leaderboards, the integrative philosophy asks whether the ensemble contains structure that individual tools miss. Hidden Markov models have a distinguished history in this regard: they underpin gene finders, chromatin state maps such as the ChromHMM method developed in the Ernst laboratory, and numerous sequence segmentation tools. MissenseHMM adapts this proven machinery to the variant interpretation domain, treating the pattern of predictor scores at a variant as the observable emission and the underlying functional class as the hidden state. The analogy with chromatin state discovery is explicit in the design: just as combinations of histone marks define reproducible chromatin states with distinct biological functions, combinations of predictor score patterns define variant states with distinct relationships to disease.

The study was conducted in the Bioinformatics Interdepartmental Program and the Department of Biological Chemistry at UCLA, where Ernst holds appointments across the Broad Center of Regenerative Medicine and Stem Cell Research, the Computer Science Department, the Jonsson Comprehensive Cancer Center, the Molecular Biology Institute and the Department of Computational Medicine. The research was supported by US National Institutes of Health grants including DP1DA044371, U01MH105578, U01HG012079 and U01MH130995, and by a UCLA Dissertation Year Award to Li. The article, published open access in Genome Biology on September 5, 2026, was edited by Wenjing She, and the authors acknowledge contributions from members of the Ernst laboratory, including preliminary work by Siddharth Naidu and discussions with Tristan Bepler, Seon-Kyeong Jang, Zhengtong Liu, Sriram Sankararaman and Noah Zaitlen.

As sequencing continues to scale and the catalog of observed human genetic variation expands faster than clinical interpretation can keep pace, resources like MissenseHMM address a widening gap. The method does not replace existing pathogenicity predictors; instead, it complements them by making their collective judgment legible. With 20 learned states spanning the space from near-unanimous pathogenicity calls to clear benign consensus, and with validation against clinical databases and deep mutational scanning experiments, the framework provides both an immediate practical annotation layer for variant interpretation and a scientific lens on the predictor landscape itself. For a field drowning in scores, a method that turns discordance into meaning may prove as valuable as any single new predictor.