• We downloaded the entire Ensembl database (6.18GB) from https://web.expasy.org/swissvar.html, which includes variants from various sources (e.g., ClinVar [1], dbSNP [2], gnomAD [3], 1000 Genomes [4], UniProt [5], ExAC [6], COSMIC [7]) (Figure 1A). To select the relevant MMs, we applied two filters: variant type (setting as MM or missense) and clinical significance (setting as benign/likely benign or pathogenic/likely pathogenic), getting 622,270 variants (Figure 1B). We then applied five additional filters to remove unwanted mutations: (1) those with conflicting interpretations of pathogenicity, (2) uncertain significance, (3) repetitive and inconsistent MMs from different sources, and (4) variants classified as benign and pathogenic/likely pathogenic, (5) variants being likely benign and pathogenic/likely pathogenic (obtaining 91,072 variants, Figure 1C). Using the Ensembl variant effect predictor (https://asia.ensembl.org/Homo_sapiens/Tools/VEP), we mapped the mutation VCF information to GRch38, excluding inconsistent chromosome/position and unmatched mutations, getting 77,700 variants (Figure 1D). Following the data processing illustrated in Figure 1, the benchmark dataset consists of 37,317 benign variants in 2,595 proteins and 40,383 pathogenic variants in 3,294 proteins. However, a subset of variants lacked annotations when processed with Ensembl VEP v104 [14]. Consequently, we excluded these variants along with the corresponding proteins. After aforementioned procedures, we finally obtained a large-scale non-redundant MM benchmark dataset. The final number of variants and proteins is presented in Table 1.

• We downloaded pathogenic GOF/LOF variants [8] (HGMD_based) from https://itanlab.shinyapps.io/goflof/, which contains a total of 9,619 variants. From this set, we selected single nucleotide variants occurring in protein coding regions excluded those that overlapped with variants in the benchmark dataset (resulting in 3,039 variants, Figure 1E). Subsequently, for the gene set within the GOF/LOF data, we searched each gene in gnomAD and applied the following filters to get benign variants: (1) removal of variants stored in ClinVar; (2) retained MMs with an allele frequency > 0.1%; (3) kept mutations with clinical significance set as benign or likely benign; (4) removed MMs that overlapped with the benchmark dataset. This process yielded 2,919 variants (Figure 1F). Finally, we merged the pathogenic and benign variants, removing redundant ones to construct blind test set. The statistical summary of variants and proteins in the benchmark dataset and blind test set is presented in the following table.

Benchmark Missense Mutation Dataset-Download

Blind Test Set-Download

Python codes--GitHub link