Human Proteome Exhaustive Missense Variant Effect Prediction Resource

Disclaimer. This resource is intended exclusively for research purposes. It is not intended for clinical diagnosis, and the data provided do not constitute medical or any other kind of advice. The authors are not responsible for its use or misuse. The authors have worked with care in the development of this dataset, but assume no liability or responsibility for any error, weakness, incompleteness or temporariness of the resource and of the data provided. This dataset provides a proteome-wide exhaustive missense variant effect prediction resource for the human reference proteome (UniProt Proteome ID: UP000005640). It comprises predictions for all 19 possible amino acid substitutions at every protein position, covering 20,521 human proteins and 217,408,803 amino-acid substitutions. Predictions in this version are based on QAFI ensembles of 30 protein-specific predictors (PSPs) and on QAFImt 1.0, as described in the original publication (Ye et al., 2026). Each protein is distributed as an individual TSV file containing the maximum available set of variant effect predictors for that protein, including the QAFI family and external predictors (AlphaMissense, popEVE, EVE, CPT-1, REVEL, BayesDel). Changes with respect to version 1.x: (i) QAFI predictors based on 30 PSPs; (ii) QAFIsplit3 updated and extended to 2,558 additional proteins; (iii) QAFIMeta v1–v3 rescaled (z-score) to the QAFIsplit3 distribution and QAFIMeta v4 recomputed; (iv) one row per amino-acid substitution. Version 1.x listed one row per underlying genomic variant, so several rows could correspond to the same substitution. Where several genomic variants produce the same substitution, genomic coordinates and genomic-level scores (REVEL, BayesDel) correspond to the first mapped genomic variant; (v) the column QAFImt was renamed QAFImt1.0. Technical validation. The objective of this validation is to ensure that the scores accessed by users correspond to those of the methodology presented in the original publication (Ye et al., 2026). Three tests were performed, and their outputs are provided as separate files:(i) Deep mutational scanning data from Beltran et al. (exhaustive_mutagenesis_validation_beltran_v2.0): per-protein Pearson correlation between observed fitness and predicted QAFIsplit3 for 337 proteins (median r = 0.476), together with the agreement between the distributed scores and the predictions computed independently by the authors of the original publication (median r = 0.995).(ii) CAGI7 (ARSA, ATP7B, LPL, TSC2) (exhaustive_mutagenesis_validation_cagi7_v2.0): comparison of the distributed QAFIMeta scores with the predictions submitted to CAGI7 (Pearson r = 0.96–0.98).(iii) ClinVar/Humsavar (exhaustive_mutagenesis_validation_clinvar_v2.0): AUC-ROC per predictor (QAFIClass AUC = 0.928). These labels were used to train QAFIClass; QAFIClass scores are leave-one-protein-out predictions. Which version to use. Version 2.0 corresponds to the published methodology and is the most accurate; version 3.0 provides the most complete coverage of the proteome. The choice of predictor depends on the problem: To predict the functional impact of missense variants as a continuous value: QAFIsplit3 (Ye et al., 2026). For the binary classification of variants (pathogenic vs. benign): QAFIMeta v4. Note that QAFIMeta v4 provides a continuous score (higher values = more benign), not a class label; a threshold must be chosen by the user. Usage advice: use version 2.0 when a score is available in both versions, and version 3.0 when it is available only there. Note: in this version, QAFImt 1.0 and the QAFI models derived from it are not available for all proteins. Where QAFImt 1.0 is missing, QAFIClass uses the training-set median for this feature. In proteins without QAFIsplit3, QAFIMeta v1–v3 are not rescaled. Where QAFIsplit3 is missing for a variant, QAFIMeta v4 is the median of the available external predictors., The computational pipeline is available at the accompanying GitHub repository. This dataset accompanies the QAFI methodology described in: Ye S, Özkan S, Díaz-Vásquez A, Padilla N, de la Cruz X (2026). Toward Advanced Quantitative Prediction of Missense Variant Effects. Int. J. Mol. Sci. 27(19), 8777. The authors acknowledge support from the Ministerio de Ciencia, Innovación y Universidades de España through grant PID2022-142753OB-I00, and from the European Regional Development Fund (FEDER).