AnnotateMissense: Genome-Wide Variant Scoring
- AnnotateMissense is a comprehensive framework that integrates genomic annotations, engineered biological features, and multiple pathogenicity predictors to score missense variants.
- The system benchmarks various machine-learning and deep-learning models, achieving high performance metrics (ROC-AUC >0.99, MCC >0.94) on a ClinVar-labeled dataset.
- It serves as an integrative meta-predictor for research prioritization and variant triage, while acknowledging limitations like potential data circularity and reliance on prior predictor inputs.
Searching arXiv for the primary paper and closely related missense-prediction work cited in the provided material. AnnotateMissense is a genome-wide missense variant annotation, feature-integration, benchmarking, and prediction framework developed to score the pathogenicity of missense single-nucleotide variants across the human hg38/GRCh38 genome (Muneeb et al., 23 May 2026). It begins with a genome-wide missense catalog, enriches each variant with a large integrated annotation set, trains supervised classifiers on ClinVar-labeled missense variants, benchmarks multiple machine-learning and deep-learning models under controlled feature configurations, performs circularity-controlled ablations and temporal validation, and then produces pathogenicity scores and binary prediction labels for 90,643,830 hg38 missense variants (Muneeb et al., 23 May 2026). The framework is best understood as a large-scale integrative meta-framework rather than as a single de novo biological model, because it combines transcript and gene annotation, conservation, population frequency, prior pathogenicity predictors, AlphaMissense, ESM-derived features, and engineered amino-acid, codon, and nucleotide-context features within supervised predictive models (Muneeb et al., 23 May 2026).
1. Scope, problem formulation, and released resource
AnnotateMissense addresses the problem of predicting whether a missense variant is likely pathogenic or benign (Muneeb et al., 23 May 2026). The framework is framed around the observation that missense pathogenicity depends on heterogeneous evidence from population frequency, evolutionary conservation, transcript context, amino acid substitution severity, prior pathogenicity predictors, and protein-language-model-derived features (Muneeb et al., 23 May 2026). Its operational target is binary pathogenic-versus-benign classification for missense single-nucleotide variants, with training labels derived from ClinVar after harmonization of pathogenic or likely pathogenic as the positive class and benign or likely benign as the negative class, while uncertain, conflicting, ambiguous, and unsupported categories are excluded (Muneeb et al., 23 May 2026).
The framework has two distinct but connected roles. First, it is an annotation and benchmarking system built on a curated ClinVar-labeled benchmark of 132,714 missense variants, comprising 76,804 pathogenic or likely pathogenic variants and 55,910 benign or likely benign variants (Muneeb et al., 23 May 2026). Second, it is a genome-wide prediction resource that scores 90,643,830 hg38 missense single-nucleotide variants derived from dbNSFP v5.1 through the CAGI7 Annotate-All-Missense challenge input (Muneeb et al., 23 May 2026). The released output consists of pathogenicity scores and binary predictions for this full missense universe (Muneeb et al., 23 May 2026).
This design places AnnotateMissense within a broader class of missense interpretation systems that integrate complementary evidence sources. Related work has described integration over strong pretrained predictors such as GPN-MSA, ESM1b, and AlphaMissense for difficult missense classification (Boulaimen et al., 2024), report-centric platforms combining missense prediction with structural and literature evidence (Vroling et al., 2021), protein-feature delta models derived from fine-tuned ESM2 (Saadat et al., 2024), and structure-centered graph models such as ALPHAGMUT (Wang et al., 2024). AnnotateMissense differs from those systems in being explicitly genome-wide, benchmark-oriented, and centered on a very large integrated annotation stack (Muneeb et al., 23 May 2026).
2. Data sources and annotation architecture
The starting point for AnnotateMissense was the CAGI7 Annotate-All-Missense challenge input, derived from dbNSFP v5.1, provided as chromosome-divided files aligned to hg38/GRCh38, and containing 90,643,830 missense single nucleotide variants (Muneeb et al., 23 May 2026). The annotation workflow converted each chromosome-level file into ANNOVAR-compatible input, ran ANNOVAR gene-based protocols, merged outputs chromosome-wise, augmented the merged tables with amino-acid substitution information and transcript-level descriptors from dbNSFP, and then added AlphaMissense scores, ESM-derived scores, and engineered biological variables (Muneeb et al., 23 May 2026).
Gene and transcript annotation was performed with ANNOVAR using RefSeq, Ensembl, UCSC Known Gene, CCDS, and GENCODE, together with exon function annotations and transcript consequence annotations (Muneeb et al., 23 May 2026). Population-frequency variables included gnomAD, ExAC, 1000 Genomes, population-specific allele frequencies, and allele rarity metrics (Muneeb et al., 23 May 2026). Conservation variables included PhyloP, PhastCons, GERP++, SiPhy, and related vertebrate and mammalian conservation features; examples explicitly named in the paper include phyloP100way_vertebrate_dbnsfp42c and GERP++_RS_rankscore (Muneeb et al., 23 May 2026).
A central property of the framework is the breadth of prior pathogenicity predictors included in the full benchmark. These predictors include SIFT, SIFT4G, PolyPhen-2, CADD, DANN, REVEL, MetaSVM, MetaLR, MetaRNN, BayesDel, ClinPred, FATHMM, LRT, PROVEAN, DEOGEN2, MVP, VEST4, MPC, PrimateAI, MutationTaster, and MutationAssessor (Muneeb et al., 23 May 2026). The AI-derived components explicitly added were AlphaMissense and ESM-derived missense effect predictions (Muneeb et al., 23 May 2026). The engineered biological feature layer included BLOSUM62 scores, amino-acid physicochemical changes, codon composition, CpG context, transition/transversion status, consensus voting features, interaction features, codon bias, nucleotide context, amino-acid severity classes, and protein-context engineered features; representative engineered variables reported in feature-importance analyses include blosum62_score, refcodon_merged, blosum_non_conservative, alt_charged, aa_structural_impact, and codon_gc_content (Muneeb et al., 23 May 2026).
Preprocessing and filtering were fold-specific during supervised benchmarking, with shared training and test features retained within each fold, missing values filled with zero, and inputs standardized using StandardScaler fitted on the training split and applied to the test split (Muneeb et al., 23 May 2026). Automatic filtering removed features with excessive missingness, constant features, genomic coordinate fields, and ClinVar-derived annotations likely to leak target labels (Muneeb et al., 23 May 2026). For genome-wide prediction, the supplement reports 334 candidate features evaluated across 25 chromosome-level files, 207 features retained for modelling, 127 excluded, 151 retained numeric features, and 56 retained categorical features, with average missingness among retained features of 14.6% and median missingness of 16.4% (Muneeb et al., 23 May 2026).
3. Feature configurations and model families
The benchmark is organized around several explicitly defined feature configurations (Muneeb et al., 23 May 2026). The central configuration is Dataset 2, the full 303-feature benchmark matrix, which includes computational pathogenicity predictors, conservation scores, functional annotations, population-frequency variables, categorical variables, raw biological sequence variables, unified prediction mapping, amino-acid property features, codon features, nucleotide substitution features, consensus-voting features, interaction features, AlphaMissense, and ESM-derived features (Muneeb et al., 23 May 2026). The remaining configurations probe restricted or circularity-controlled versions of this integrated representation.
Dataset 3 is a naive 41-feature matrix retaining conservation scores, functional annotations, and categorical variables (Muneeb et al., 23 May 2026). Dataset 4 is a location-oriented 56-feature matrix retaining functional annotations, categorical variables, raw biological sequence variables, and derived biological variables such as BLOSUM62, amino-acid property changes, codon composition features, and nucleotide substitution features (Muneeb et al., 23 May 2026). Dataset 5, NoPriorPredictors, removes computational pathogenicity predictor scores and population-frequency metrics while retaining conservation, functional and transcript annotations, categorical variables, and engineered biological features (Muneeb et al., 23 May 2026). Dataset 6, NoClinicalFeatures, more conservatively excludes prior pathogenicity predictors, BayesDel allele-frequency scores, AlphaMissense, ESM-derived scores, MetaRNN, and ClinPred (Muneeb et al., 23 May 2026). Dataset 7, NoAlphaMissenseESM, excludes only AlphaMissense and ESM-derived features (Muneeb et al., 23 May 2026). Dataset 8, EngineeredBiologicalOnly, retains only engineered sequence-derived biological features including BLOSUM62, amino-acid physicochemical changes, codon composition, CpG context, codon bias, transition/transversion status, and nucleotide context (Muneeb et al., 23 May 2026).
The study evaluates both classical machine-learning and deep-learning models (Muneeb et al., 23 May 2026). XGBoost was used with n_estimators = 200, max_depth = 8, learning_rate = 0.1, subsample = 0.8, colsample = 0.8, and fold-specific scale_pos_weight in the primary benchmark, while the final genome-wide model used max_depth = 6, subsample = 0.9, colsample_bytree = 0.9, reg_alpha = 0.1, reg_lambda = 1.0, random_state = 42, and log loss as evaluation metric (Muneeb et al., 23 May 2026). Random Forest used 200 trees, max depth 15, min split 5, min leaf 2, max_features = sqrt, bootstrap aggregation, out-of-bag scoring, and balanced class weights (Muneeb et al., 23 May 2026). FLAML AutoML optimized ROC-AUC within a 600-second time budget per fold over XGBoost, Random Forest, LightGBM, and Extra Trees (Muneeb et al., 23 May 2026). TabNet used n_d = 64, n_a = 64, n_steps = 7, gamma = 1.3, lambda_sparse = 1e-3, Adam, entmax masking, and step-wise learning-rate scheduling (Muneeb et al., 23 May 2026). The PyTorch DNN used four hidden layers of 512, 256, 128, and 64 units with batch normalization, dropout 0.4, 0.3, 0.2, and 0.1, Adam, and early stopping, while the TensorFlow DNN used dense 512–256–128–64 layers with batch normalization, dropout, Adam, binary cross-entropy loss, and internal validation-based early stopping (Muneeb et al., 23 May 2026).
The main benchmark employed five-fold stratified cross-validation, with benign encoded as 0 and pathogenic encoded as 1, Matthews correlation coefficient as the primary metric, and ROC-AUC as the secondary metric (Muneeb et al., 23 May 2026). The choice of MCC is consistent with the paper’s focus on balanced binary discrimination across feature ablations and model families (Muneeb et al., 23 May 2026).
4. Benchmark performance and temporal validation
On the full 303-feature Dataset 2, XGBoost was the best-performing model overall, reaching ROC-AUC and MCC , with a 95% confidence interval for MCC of (Muneeb et al., 23 May 2026). FLAML AutoML was close behind with ROC-AUC and MCC , followed by Random Forest, TabNet, PyTorch DNN, and TensorFlow DNN, all with ROC-AUC above 0.992 and MCC above 0.931 on Dataset 2 (Muneeb et al., 23 May 2026). These results identify the full integrated feature space, rather than any single model architecture, as the dominant source of benchmark performance (Muneeb et al., 23 May 2026).
Performance dropped sharply on reduced feature sets. On Dataset 3, the naive 41-feature configuration, the best MCC was with FLAML AutoML and ROC-AUC was (Muneeb et al., 23 May 2026). On Dataset 4, the location-oriented 56-feature configuration, the best MCC was with Random Forest and ROC-AUC was (Muneeb et al., 23 May 2026). Dataset 8, EngineeredBiologicalOnly, produced very similar performance, with the best MCC at using Random Forest and ROC-AUC 0 (Muneeb et al., 23 May 2026). The paper explicitly interprets these results as showing that engineered sequence and protein-context features contain useful signal but are insufficient on their own (Muneeb et al., 23 May 2026).
The framework was also compared with established pathogenicity predictors across 135,512 missense variants (Muneeb et al., 23 May 2026). The Dataset 2 XGBoost model reached Pearson correlation 0.983, Spearman correlation 0.853, MCC 0.981, and accuracy 0.991 against ClinVar binary labels (Muneeb et al., 23 May 2026). Examples of comparator values include MetaRNN with MCC 0.883 and accuracy 0.942, BayesDel_addAF with MCC 0.818 and accuracy 0.908, PROVEAN with MCC 0.615 and accuracy 0.806, PolyPhen-2 HVAR with MCC 0.582 and accuracy 0.785, SIFT4G with MCC 0.556 and accuracy 0.774, AlphaMissense with MCC 0.420 and accuracy 0.658, REVEL with MCC 0.426 and accuracy 0.670, and CADD with MCC 0.205 and accuracy 0.476 (Muneeb et al., 23 May 2026). The paper cautions that such high concordance with ClinVar should be interpreted in light of possible evidence overlap and circularity (Muneeb et al., 23 May 2026).
Temporal validation was performed by parsing a newer ClinVar variant_summary.txt, restricting to GRCh38 SNVs, retaining only pathogenic and benign labels, removing variants already present in the older annotation-derived set, and evaluating the final categorical predictions joined by chromosome, VCF position, reference allele, and alternate allele (Muneeb et al., 23 May 2026). From 4,129,645 newer ClinVar GRCh38 SNVs, the authors derived 1,459,305 newly observed strict pathogenic or benign variants, of which 298,353 had AnnotateMissense categorical predictions (Muneeb et al., 23 May 2026). On this temporal validation, AnnotateMissense achieved MCC 0.7613, accuracy 0.8798, sensitivity 0.9070, specificity 0.8563, precision 0.8453, and F1-score 0.8750 (Muneeb et al., 23 May 2026). The paper explicitly states that this is temporal concordance with ClinVar rather than independent clinical validation (Muneeb et al., 23 May 2026).
5. Circularity-controlled ablations and methodological interpretation
The ablation analyses are among the most consequential aspects of AnnotateMissense because they quantify how much benchmark performance depends on potentially overlapping evidence sources (Muneeb et al., 23 May 2026). When prior pathogenicity predictors and population-frequency metrics were removed in Dataset 5, XGBoost dropped from MCC 0.94108 on Dataset 2 to MCC 0.71977, with ROC-AUC decreasing from 0.99497 to 0.93979 (Muneeb et al., 23 May 2026). Dataset 6, which also excluded broader clinically overlapping features such as BayesDel allele-frequency scores, AlphaMissense, ESM-derived scores, MetaRNN, and ClinPred, produced nearly identical performance to Dataset 5; for XGBoost, MCC was 0.71906 and ROC-AUC was 0.93977 (Muneeb et al., 23 May 2026). The paper interprets this as showing that once prior predictors and population-frequency information are removed, excluding the additional clinically overlapping tools does not cause a much larger drop (Muneeb et al., 23 May 2026).
By contrast, removing only AlphaMissense and ESM-derived features in Dataset 7 had minimal effect (Muneeb et al., 23 May 2026). For XGBoost, MCC changed only from 0.94108 on Dataset 2 to 0.94029 on Dataset 7, while ROC-AUC changed from 0.99497 to 0.99470 (Muneeb et al., 23 May 2026). The paper therefore concludes that AlphaMissense and ESM contribute little independent gain once the broader feature space is already present (Muneeb et al., 23 May 2026). This sharply distinguishes AnnotateMissense from integration studies in which AlphaMissense-like or ESM-like signals are more central. For example, an integration study over GPN-MSA, ESM1b, and AlphaMissense reported that adding AlphaMissense materially boosted downstream performance on hard ProteinGym and ClinVar-linked missense subsets (Boulaimen et al., 2024). A plausible implication is that the marginal value of AlphaMissense- and ESM-derived signals depends strongly on the rest of the feature stack rather than being invariant across frameworks.
These ablations also clarify the meaning of AnnotateMissense as a predictor. The paper states that the full model functions substantially as an integrative meta-predictor rather than purely as a de novo biological predictor (Muneeb et al., 23 May 2026). This characterization is important when comparing it with platforms such as Helix, which explicitly avoid using predictions from other classifiers as features and do not use minor allele frequency as a feature (Vroling et al., 2021), or with sequence-centered models such as MutFormer, whose sequence-only form is intended to provide a relatively independent learned sequence prior (Jiang et al., 2021). AnnotateMissense instead demonstrates the performance obtainable when many established signals are aggregated at scale (Muneeb et al., 23 May 2026).
The framework’s biological use-case analysis on ClinVar missense VUS adds a secondary interpretive layer. Applied to 49,990 ClinVar missense VUS and compared against gnomAD v2.1.1 missense constraint Z-score (mis_z), AnnotateMissense assigned higher scores to variants in more missense-intolerant genes: mean 0.631 and median 0.748 for mis_z > 3, mean 0.600 and median 0.705 for 1, and mean 0.548 and median 0.614 for mis_z < 0 (Muneeb et al., 23 May 2026). The proportion of variants with score 2 was 37.4% in intolerant genes, 34.0% in intermediate genes, and 31.5% in tolerant genes, with Mann–Whitney 3 for intolerant versus tolerant and 4 for intolerant versus intermediate (Muneeb et al., 23 May 2026). The paper interprets this as biologically consistent with stronger predicted pathogenicity in genes under greater missense constraint (Muneeb et al., 23 May 2026).
6. Outputs, intended use, and limitations
For genome-wide inference, the final model was XGBoost trained on merged chromosome feature files after retaining only variants with benign or pathogenic CLNSIG, excluding VUS and unknown labels, applying a stratified 80:20 train/test split, and using SMOTE only on the training data (Muneeb et al., 23 May 2026). The resulting classifier generated a continuous pathogenicity score and a binary label for each missense variant in the 90,643,830-variant hg38 universe (Muneeb et al., 23 May 2026). Binary predictions are typically thresholded at 0.5, so score 5 corresponds to pathogenic or damaging and score 6 corresponds to benign or tolerated (Muneeb et al., 23 May 2026). The outputs were released through a GitHub repository and a Zenodo deposition that includes genome-wide prediction output and a DuckDB database with selected annotations and predictions (Muneeb et al., 23 May 2026).
The intended use is research prioritization, benchmarking, and variant triage, with the paper explicitly stating that AnnotateMissense should not be interpreted as a standalone clinical classification system (Muneeb et al., 23 May 2026). This positioning aligns with a wider pattern in missense interpretation tools. Helix is likewise presented as decision support rather than autonomous clinical interpretation (Vroling et al., 2021). Protein-feature delta models based on fine-tuned ESM2 are described as best used as explanatory annotation layers or PM1-support modules rather than as standalone clinical pathogenicity predictors (Saadat et al., 2024). Structure-centered systems such as ALPHAGMUT are described as particularly suitable as structural annotation modules or ensemble components when good 3D structures are available (Wang et al., 2024). AnnotateMissense belongs to this decision-support landscape but is distinguished by its genome-wide scope and explicit emphasis on benchmarking (Muneeb et al., 23 May 2026).
The paper is unusually explicit about its limitations (Muneeb et al., 23 May 2026). First, ClinVar-based benchmarking may inflate performance because many included predictors were trained on ClinVar, HGMD, or overlapping databases, so shared prior knowledge may contribute to apparent predictive strength (Muneeb et al., 23 May 2026). Second, the ablations show that much of the high benchmark performance depends on prior pathogenicity predictors and population-frequency information, confirming that the full model behaves substantially as an integrative meta-predictor (Muneeb et al., 23 May 2026). Third, temporal validation is described as temporal concordance with ClinVar rather than true external clinical validation (Muneeb et al., 23 May 2026). Fourth, although direct ClinVar fields such as CLNDN, CLNREVSTAT, CLNDISDB, and CLNALLELEID were removed, residual indirect leakage remains possible through external features trained on overlapping sources (Muneeb et al., 23 May 2026). A further implementation caveat noted in the supplement is that TensorFlow DNN results are omitted for Datasets 5–8 because of a file caching issue identified post hoc (Muneeb et al., 23 May 2026).
In methodological terms, AnnotateMissense shows that very high ClinVar-concordant performance can be obtained from a large integrated annotation stack, but it also shows that such performance is strongly contingent on previously learned predictors and population-frequency evidence (Muneeb et al., 23 May 2026). This suggests that the framework is most informative when interpreted as a scalable, queryable, and benchmarked aggregation layer over heterogeneous missense evidence rather than as a self-sufficient mechanistic model.