Learning sequence to predict gain- or loss-of-function variants

Learning sequence to predict gain- or loss-of-function variants | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Learning sequence to predict gain- or loss-of-function variants Doyeon Ha, Sungnam Kim, Kisang Kwon, Wonseok Chung, Joohyun Han This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6705195/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract A clear understanding of mutational effects can advance genetics and biomedical research by providing valuable insights into gene functions, disease mechanisms, and therapeutic approaches. However, methods to determine the pathogenicity of genetic variants are limited by the absence of information on the direction of mutational effects. Here, we present ClearVariant, a deep learning system to classify pathogenic variants into gain- or loss-of-function, achieving state-of-the-art performance validated with data from ClinVar and Human Gene Mutation Database (HGMD). The model contains protein language models (PLMs) for training mutated sequences alongside their reference counterparts, showing similar predicted outcomes when a residue changed to another amino acid belonging to the same property group. We evaluated its ability to learn the protein language by observing high attention scores on coevolutionary relationships. To support advancements in biomedicine, we provide a database of pathogenic human missense variants labelled with their predicted mutational effects. Biological sciences/Computational biology and bioinformatics/Genome informatics Biological sciences/Genetics/Functional genomics/Mutagenesis Biological sciences/Computational biology and bioinformatics/Sequence annotation Biological sciences/Genetics/Clinical genetics Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Elucidating the phenotypic consequences of genetic mutations is a fundamental goal in genetics and biomedical research 1 . Since the effects of all mutational alterations on the human genome cannot be observed or confirmed, variant effect predictors (VEPs) are used to predict the functional impacts of genetic mutations. The American College of Medical Genetics and Genomics (ACMG) Standards and Guidelines, aimed at improving genetic diagnosis and clinical decision-making, have provided recommendations for the use of computational methods to assess the pathogenicity of genetic variants 2 . Recently, several studies have reported the effectiveness of performing in silico predictions using multiple VEPs and proposed a higher level of evidence strength 3 , 4 . Recent progress in artificial intelligence has led to the development of deep learning models. Cheng et al . reported that the performance scores of their deep learning model assessing missense variants within multiple sequence alignment (MSA) coverage were the highest across numerous clinical and experimental pathogenic benchmarks 5 . The model was introduced as an adaptation of AlphaFold, which is a neural network-based model previously built for predicting 3D protein structures with approximately 93 million trainable parameters. Moreover, Brandes et al . used evolutionary scale modelling (ESM1b), a protein language model with 650 million parameters, to predict the effects of not only missense variants but also coding variants affecting multiple residues, including in-frame indel and nonsense mutations 6 . However, the mechanisms of pathogenic variants are heterogeneous. Most VEP tools are limited in their ability to predict whether a mutation is pathogenic or benign 7 . As a recent review article reported, none of the widely used computational tools used to predict variant effects, except for MutPred2, can predict mutational effects on gene function 8 , 9 . This limitation leads to the next step in computational mutation analysis: identifying whether the mutations are either gain-of-function (GOF) mutations or loss-of-function (LOF) mutations. Tools that are used to predict the mutational effects on a variant of interest could be called directional variant effect predictors (D-VEPs). The classification of mutational effects on gene function is important for obtaining a clear understanding of disease mechanisms. GOF mutations typically confer proteins with enhanced or novel functions that can affect cellular processes whereas LOF mutations result in complete or partial loss of normal protein function 10 . For example, the PCSK9 D374Y mutation, labelled as a GOF mutation, increases the binding affinity of PCSK9 to low-density lipoprotein receptors (LDL-Rs) and is thus considered a causal determinant of hypercholesterolemia, a disease characterized by elevated levels of LDL cholesterol 11 , 12 . In contrast, PCSK9 LOF mutations, such as R46L and G106R, are associated with hypocholesterolemia, which is a disease characterized by low levels of LDL cholesterol 11 , 12 . Importantly, elucidating mutational effects on the functions of genes such as PCSK9, LDL-R, APOB, and ANGPTL3 has facilitated sophisticated diagnostic strategies and the identification of drug targets for cholesterol-related diseases 13 . The biological mechanisms of LOF mutations are well established; however, the mechanisms underlying functional enhancement or the acquisition of a new function are more complicated 14 . Existing D-VEP methods, including MutPred2 and LoGoFunc, exhibit poor GOF mutation prediction performance due to an insufficient amount of biological information for models to generate a specific inference 9 , 15 . The low amount of data on GOF also makes it difficult to identify new GOF mutations 15 . The ability to precisely predict GOF mutations is connected to understanding protein language, as it relies on recognizing the nuanced relationships between a protein's sequence and its function 16 , 17 . Like how written words form coherent meanings in natural languages, functional sequence regions, including specific motifs and domains within the gene sequence, dictate the functions of the encoded proteins in biological systems 18 . Protein language models (PLMs) depend heavily on the broad sequence context, which might distinguish between GOF and LOF mutations. The increasing collection of annotations on variant functionality has supported deep learning training with mutated protein sequences. ClinVar and the Human Gene Mutation Database (HGMD) harbour information on mutational effects and their phenotypic consequences 19 . Multiplexed assays of variant effects (MAVEs) systematically measure protein variant effects, and the data have been organized as large-scale experimental benchmarks 20 . Furthermore, approximately 12 million reference sequences across over a million species have been collected over time, providing the opportunity to train models with wild-type and variant protein languages 21 . With increasingly large genetic datasets, PLM training operates at the amino acid level by embedding vectors for each residue 22 . By employing attention mechanisms, these models enable the interpretation of residue importance, specifically for identifying which residues are key to determining the effects of mutations. Here, we present ClearVariant, a model for classifying GOF and LOF mutations. The model explicitly integrates both reference and variant sequences to facilitate direct comparisons of mutational effects and leverages a self-attention mechanism to enhance interpretability. Evaluations using datasets on pathogenic mutations across diverse genes revealed that ClearVariant outperforms existing models, achieving a GOF prediction precision of 0.814 and an F1 score of 0.778, which are much higher than the GOF prediction precision of 0.578 and F1 score of 0.670 achieved by the previous state-of-the-art (SOTA) model. Furthermore, experimentally measured single-gene mutational effects demonstrate that its inference outcomes and decision-making process align well with established biological insights, underscoring its reliability for functional variant prediction. Results Learning protein sequences to predict GOF and LOF variants We developed ClearVariant, an attention-based deep learning model, to predict GOF and LOF variants on the basis of ESM-2 (Fig. 1 a; Methods) 15 . ESM-2 was trained using 250 million protein sequences for protein structure prediction. We restructured the model architecture of ESM-2 for the following two principles. First, a protein’s reference sequence and variant sequence were jointly considered since the direction of the mutational effects is assigned by comparing consequences of mutated genes to those of wild-type genes. Therefore, it is necessary to explicitly introduce the reference sequence to provide context for direct comparison. Proteins with GOF mutations have enhanced functions or new functions, whereas proteins with LOF mutations present either partial or complete loss of function relative to the WT protein 10 . Second, the model architecture was structured to allow interpretability by incorporating a self-attention mechanism within the classification layer. Attention refers to the weighting of input features on the basis of their relevance to the model’s decision-making process 23 . The CLS (classification) vector is subsequently used as the input of the classification head. To validate the predictive performance of ClearVariant regarding mutational effects, we conducted training and evaluation experiments using two distinct datasets. The first dataset comprised pathogenic GOF and LOF mutations spanning a diverse array of genes from the ClinVar database and the HGMD (Fig. 1 b) 19 , 24 . The other dataset consisted of experimental measurements quantifying the effects of mutations in a single gene obtained from ProteinGym (Fig. 1 c) 20 . To ensure the reliability of the model's predictions, we performed ‘explanation’ and ‘interpretation’ analyses. In the ‘explanation’ analysis, we verified that the model's predictions aligned with established biological patterns such as amino acid chemical properties (Fig. 1 d) 25 . In the ‘interpretation’ analysis, we evaluated whether the underlying rationale of the model's decisions was consistent with biological information that was not explicitly provided during training (Fig. 1 e). By analysing the attention placed between residue pairs within the model, we can interpret the extent to which each residue interaction provides crucial information for GOF and LOF prediction. For example, we assessed whether residues associated through coevolution were recognized as key determinants in predicting mutational effects 26 . Precise prediction of mutational effects on gene function To assess the performance of ClearVariant in predicting mutational effects, we compared basic controls and previous D-VEP models in all available pathogenic missense variants with GOF and LOF labels from the ClinVar database and the HGMD 19 , 24 . ClearVariant achieved a GOF precision of 0.814 and a GOF F1 score of 0.778, outperforming all the other controls, including the previous SOTA model, LoGoFunc, with a GOF precision of 0.578 and a GOF F1 score of 0.670 (Fig. 2 a and b) 15 . In LOF prediction, which is a relatively easy task where performance has converged, ClearVariant performs similarly to the other basic models (Fig. 2 c and d). To more accurately evaluate the overall performance of the model, we examined the precision‒recall curve with positive labels of GOF and LOF (Fig. 2 e and f), and also assessed the receiver operating characteristic (ROC) curve for comprehensive evaluation (Supplementary Figs. 1 and 2). Thus, the model revealed a greater proportion of correct answers in the GOF prediction (Fig. 2 g; odds ratio = 99.24 and P = 2.43 × 10 − 87 (ClearVariant), odds ratio = 34.31 and P = 2.55 × 10 − 58 (LoGoFunc)). ClearVariant exhibited better performance than a single ESM-2 encoder, achieving higher MCC values in the GOF and LOF predictions (Supplementary Fig. 3). Basic controls included i) random , ii) all one label , and iii) gene-biased prediction (gray bars in Fig. 2 a-d; Methods). The control D-VEP models included LoGoFunc, MutPred2, and Gerasimavicius, etc. (salmon bars in Fig. 2 a-d; Methods) 9 , 14 , 15 . Notably, the gene-biased prediction had high prediction performance comparable with those of the previous models, although it is a simple rule-based model where inference is decided by other labels on variants of the gene. It is not surprising that no significant difference between GOF and LOF variants was observed in the pathogenic scores calculated by AlphaMissense or 3Cnet (Fig. 2 h and Supplementary Fig. 4; P = N.S., Mann–Whitney U test) 5 , 27 . Similarly, the pathogenicity scores of the predicted GOF and LOF variants were similar (Supplementary Fig. 5; P = N.S., Mann–Whitney U test). Previous VEPs have not been developed to predict the direction of the mutation effect; thus, most VEPs are not suitable for predicting GOF and LOF variants. Small biological datasets often present challenges in AI due to inherent characteristics such as high measurement noise and complex data structures, which can lead to greater variability in model performance during cross-validation 28 . Driven by this concern, we proceeded k-fold cross validation (k = 5) (blue lines in Fig. 2 i; Methods). An ESM-2 model with randomized parameters showed far lower performance between the initial and final epochs, highlighting the importance of pretraining the model with protein language (gray line in Fig. 2 i). Furthermore, we evaluated the GOF and LOF prediction performance of ClearVariant for variants classified as likely pathogenic and likely benign by AlphaMissense in the test set of the 5-fold cross-validation (Supplementary Fig. 6). The results revealed that ClearVariant exhibited greater predictive performance across all evaluation metrics for variants classified as likely pathogenic by AlphaMissense. This might be due to relatively larger data numbers in AM-likely pathogenic variants (Supplementary Fig. 7), and correct prediction by ClearVariant is achieved with a larger dataset (Supplementary Fig. 8). Biological explanation of inference outcomes Human-understandable explanations are required for the results of biological analyses with in silico tools, leading to their application in subsequent research and medical fields 8 , 29 . In studies involving DNA or protein sequences, the results can be explained in terms of amino acid residues, which biological researchers intuitively understand 30 , 31 . First, in the collection of benchmarks for mutational effects, changes to amino acid groups by residues with similar characteristics are more likely to have similar mutational effects (Fig. 3 a-d, Supplementary Fig. 9) 25 . A mino acids can be categorized into four groups on the basis of their side chain properties: charged positive (CP), charged negative (CN), polar uncharged (PU), and hydrophobic (HP) . Specifically, one residual change to CP groups has a similar mutational effect with other residual changes to CP groups compared with changes to CN , PU , or HP groups (Fig. 3 a). The similarity in mutational effects driven by the chemical properties of amino acids is consistently maintained for wild-type and variant amino acid pairs (Supplementary Figs. 10 and 11). For example, in interleukin enhancer binding factor 3 ( ILF3 ), mutational changes to CN amino acids (Asp and Glu) at residues 3 and 4 increased the stability of the protein, whereas changes to HP amino acids (Ala, Val, Ile, Leu, Met, Phe, Tyr, and Trp) decreased the stability of both the observed values in ProteinGym and the values predicted by ClearVariant (Fig. 3 e and f). The predicted and observed values of ILF3 mutations were highly significantly correlated (Fig. 3 g; ρ = 0.936 and P < 5.00 × 10 − 324 ). Amino acid residues form interresidue interactions within a protein to construct wild-type structures of a protein; thus, those interaction residues essentially contribute to the protein’s functions and mechanisms 32 . Compared with variants at non-interaction residues, variants at interaction residues exhibit a greater reduction in predicted values relative to that of the WT (Fig. 3 h and Supplementary Fig. 12). Two standards were utilized for the detection of residual interactions: Ca5 and Min3 (Methods) 33 . We analysed the within-protein interactions using 32 proteins that had AlphaFold protein structures and WT experimental values. Under Ca5 standard conditions, 4 out of 32 proteins (12.5%) presented higher values for their interaction residues than did the WT, whereas 11 out of 32 proteins (34.4%) presented elevated values for their non-interaction residues than did the WT. For example, the mutational effects predicted by ClearVariant for peroxisome proliferator-activated receptor γ (PPARG) were significantly different between interaction and non-interaction residues (Fig. 3 i). Surprisingly, variants with non-interaction residues presented higher activity values than the WT did. Residues located in the protein interior are more likely to engage in residue‒residue interactions 34 , 35 . When the average score of the predicted activity index values for all 19 possible missense variants at each residue were mapped onto the protein structure, we observed that residues in the protein interior tended to have lower predicted activity index values than the wild type did, whereas residues with higher predicted values were predominantly located on the protein surface. The PPARG mutation M280I, located on the protein surface, is related to overexpression, which is a gain-of-function mechanism (green colour in Fig. 3 i) 36 . These findings are consistent with those of previous studies, suggesting functional differences between the core and surface of proteins 14 . ClearVariant had a Spearman median of 0.73 and a Pearson median of 0.78 across all 92 human proteins (Supplementary Fig. 13). However, the standard deviation was large, which means that the prediction performance varied across proteins. To investigate this, we measured performance across different experimental categories and assessed data coverage. As a result, we found that only the data coverage of the variant ratio of stability, highest prediction performance, was greater than that of the other experimental categories (Supplementary Figs. 14 and 15). Within-model interpretation at amino acid residue resolution Understanding the specific information an AI model uses as the basis for its decisions is essential for improving the model's reliability 23 , 37 . In ClearVariant, the self-attention layer of the interpretation module plays a critical role in comparing the mutated protein sequence with the reference protein sequence at the residue level, determining which residue's information should be prioritized for prediction (Fig. 4 a). Residues receiving higher attention weights are deemed more significant and contribute more substantially to the formation of downstream representations (Fig. 4 b). Since the final prediction is made using the CLS vector, analysing the attention weights that contribute to forming this vector allows us to identify the residues that ClearVariant considers crucial for predicting the direction of mutational effects (Fig. 4 c). Attention analysis revealed that coevolution residues provide key information for predicting the direction of mutational effects. 44 mutational effect datasets with information on protein coevolution were utilized for attention analyses 20 , 26 . We obtained the attention distribution for four residue groups: 1) the residue where the variant occurs, referred to as the residue of interest (ROI), 2) the coevolved residues with the ROI, and 3) all other residues. In the ornithine transcarbamylase (OTC) protein enzymatic activity dataset, coevolution residues presented a greater distribution of attention scores than did all other residues, both when all missense variants across the entire protein were considered and when only those within a specific domain were analysed (Fig. 4 d and e). This finding suggests that, for each missense variant, coevolving residues consistently receive high attention scores (Fig. 4 f and g). The median attention weights indicated that, in most cases, the ROIs and coevolved residues had higher attention weights than the other residues did (Fig. 4 h and i; coevolution of in-protein, P = 7.33 × 10 − 6 , coevolution of in-domain; P = 4.51 × 10 − 5 , ROI of in-protein; P = 3.09 × 10 − 6 , ROI of in-domain; and P = 1.05 × 10 − 5 Wilcoxon signed rank test; Supplementary Fig. 16). Notably, when ROIs or coevolution residues had higher attention weights, the differences were often substantial, exceeding 10 4 -fold in some instances. Conversely, in cases where other residues had higher attention weights, the differences were marginal, typically close to onefold. The median attention weight for coevolution residues in 33 models was greater than that for other residues, with the highest cases showing a 10 3.5 -fold difference. This suggests that the model effectively learns the biological significance of interactions involving coevolution residues, which are well known to play a critical role in protein functions. Additionally, to understand whether high performance in the gene-biased model (Fig. 2 a-d) is derived from research bias or the evolution of protein language, we scrutinized the degree of expression, stability and activity changes caused by a missense mutation. If the proportions are distributed at both extremes, most genes might be considered affected by research bias. We observed a wide distribution in the proportion of variants with experimental scores higher than those of the WT, meaning that each protein could obtain divergent protein languages by evolving in separate paths to different proportions of GOFs and LOFs (Supplementary Fig. 17, Methods). GOF/LOF inference dataset for the science community We provide the integrated resource of GOF/LOF inference on pathogenic variants by ClearVariant, and inference is conducted on pathogenic variants (Fig. 5 a). Four additional annotations are tagged: i) ClinVar pathogenicity annotation, ii) AlphaMissense prediction, iii) ClearVariant prediction, and iv) whether the variant belongs to a gene encountered during training. Notably, ClearVariant predicted the KCNQ1 p.Arg231His variant—reported in multiple families with Familial Atrial Fibrillation (FAF)—as a GOF variant 38 . This prediction aligns with previous studies suggesting that GOF mutations in KCNQ1 are associated with the development of FAF 39 , 40 . The number of GOF variants was 723 (14.5%) for the ClinVar/HGMD variants mapped with GOF/LOF (Fig. 5 b and Supplementary Table 1), 20,831 (14.2%) for the ClearVariant’s prediction of ClinVar pathogenic variants (Fig. 5 c), and 7.7 M (9.4%) for the ClearVariant’s prediction of AM-likely pathogenic variants (Fig. 5 d). The GOF percentages are between 8.7% and 15.4% across diverse intervals of the minor allele frequency (Fig. 5 e). Discussion The development of D-VEPs, defined as VEPs with a determined direction of mutational effects, could lead to new insights into various biological fields from the mechanistic study of diseases to genetic testing for diagnosis and drug development. The diagnostic yield of Mendelian disorders has steadily increased over time. For example, an after-five-year reanalysis of paediatric neurological diseases increased the yield from 24.8–46.8% 41 . One of the core issues for undiagnosed cases is the inability to identify causal variant candidates. The progressive development of VEPs in recent years has contributed to molecular diagnosis and reanalysis by facilitating the selection of causal variants 42 . By enabling molecular diagnostic specialists to identify the alignment between the patient's symptoms and mutation effects, D-VEP could increase the diagnostic rates of rare diseases and decreases the diagnostic turnaround time, enhancing both patient quality of life and survival outcomes. Trained on human genomes and clinical annotations, D-VEPs could yield a significant contribution to drug development by accelerating the entire pipeline. Human genetic evidence significantly increases the probability of FDA approval, making drugs with genetic evidence 2.6 times more likely to receive approval than those without 43 , 44 . Genetics-based approaches aid in the selection of appropriate drug targets. Identifying diseases driven by GOF mutations and subsequently targeting the affected proteins with inhibitory compounds is a representative strategy for drug development. Furthermore, D-VEPs can find a variant with a reverse effect on the causal variant using data on protein functional annotations. These mutations, provided that they result in LOF mechanisms, offer a basis for developing inhibitory therapies that replicate the protective effects. For example, the first genetic evidence that ANGPTL3 is associated with hypobetalipoproteinaemia was reported in 2010 44 . After 11 years, by mimicking the beneficial effects of LOF variants in target genes, EVINACUMAB, an angiopoietin-related protein 3 inhibitor, was approved for this disease. This strategy, with the assistance of D-VEPs, will revolutionize the identification of targets for drug development. The prediction performance of an in silico tool alone does not guarantee its adoption in medical applications 8 . The reliability of such tools, especially deep learning models, in predicting pathogenicity or mutational effects can be determined by two factors: i) the outputs of the tool are aligned with established biological knowledge, and ii) the prediction mechanism of the tool is human-understandable. To evaluate the former, which we call ‘explanation’, we analysed the amino acid characteristics of the inference outputs from ClearVariant (Fig. 3 ). We observed that alterations to the same chemical property tended to have similar mutational effects on a single residue. Interestingly, alterations in intraprotein interaction residues are more likely to be LOF mutations. For the latter, which we call ‘interpretation’, we scrutinized the internal representations of ClearVariant and observed that coevolution residues place considerable attention on the residue of interest (Fig. 4 ). This observation means that the inference process of ClearVariant is largely guided by the coevolutionary relationship with the residue of interest. Beyond prediction performance, ‘explanation’ and ‘interpretation’ analyses could make the model more usable in real-world applications, providing biological insights for further discussion. Rather than starting from scratch, ClearVariant was built on the basis of models reported in prior works in both artificial intelligence and life sciences. Vaswani et al ., whose work introduced the Transformer enabled parallel processing and improved efficiency by training a large amount of text sequence data with self-attention mechanisms 45 . By inheriting the structure and methodology of the Transformer, advanced PLMs like evolutionary scale modelling foundation models are refined to learn and interpret protein language within its broader biological context, through training a large volume of reference sequences across over 138 million proteins 6 . Pathogenic variant data mapped to GOFs and LOFs have been collected and imported into various databases, including the ClinVar database and the HGMD, over time 19 , 24 . The determination of LOFs has been achieved by many tools based on simple standards, and even their prediction precision reached approximately 0.95 in a previous study (Fig. 2 ) 15 . Unlike LOFs, GOFs are not easily identifiable, and the amount of available data on GOFs is significantly smaller, making GOF prediction highly challenging. Leveraging these sophisticated PLMs and the increasing availability of extensive biological data, ClearVariant—a model fine-tuned with a large amount of labelled genomic data—is poised to predict GOFs precisely. To unlock the full potential of D-VEP and broaden its applications, effectively addressing its current limitations is paramount. Indeed, the current scope of D-VEP often focuses on GOF and LOF, while the broader spectrum of mutational effects, including dominant-negative effects and haploinsufficiency, presents further complexities 14 . Encouragingly, the ongoing efforts to collect and annotate data on such diverse mutational effects are paving the way for more sophisticated and comprehensive predictive models 19 . Another recognized challenge lies in maintaining high prediction performance for variants emerging after a model's training period or for mutations in novel genes (Supplementary Figs. 2 and 18). This performance discrepancy often highlights the limitations of employed datasets; for instance, the dataset used in this study, while substantial, covered only 975 proteins with pathogenic GOF or LOF labels (Supplementary Table 1), representing a small fraction of the human proteome and thus a potentially biased subset (Supplementary Fig. 19) 46 . Nevertheless, these hurdles, although their resolution requires considerable time, are surmountable. Significant improvements are anticipated through the strategic integration of diverse data types beyond gene sequences. For example, incorporating allele frequency data, phenotypic information from large population cohorts such as the UK Biobank, and detailed knowledge of signalling pathways and functional processes holds considerable promise for more accurately determining the direction and impact of mutational effects 47 – 49 . Ultimately, adaptive evolution of D-VEP, building upon methodologies such as those demonstrated herein, is anticipated to be pivotal for the future of genetics and biomedical research. Methods Data Curation Protein reference sequence We collected protein reference sequences, used them as inputs for deep learning, and generated variant sequences. A total of 20,359 reviewed human protein reference sequences were downloaded on March 7, 2024, from the UniProtKB website ( https://www.uniprot.org/uniprotkb?query=* ). Protein sequences and their corresponding gene names and UniProt and RefSeq IDs (NM_, NP_) were obtained. Gene names and IDs were used to match reference sequences with variants across multiple databases. ID versions were not considered. Training and test sets for LOF/GOF prediction We obtained pathogenic GOF and LOF variants from Bayrak et al . The datasets were generated on the basis of known pathogenic variants from the HGMD and the ClinVar database. For each pathogenic variant, they searched for GOF- and LOF-related nomenclature in all relevant publications using natural language processing libraries. They curated each dataset using the 2019.1 professional version of the HGMD and the 2021.06 version of ClinVar. The GOF and LOF variant datasets were downloaded from their website ( https://itanlab.shinyapps.io/goflof/ ). A total of 9,619 variants were obtained from the HGMD-derived dataset (goflof_HGMD2019_v032021_allfeat.csv), and 5,085 variants were obtained from the ClinVar-derived dataset (goflof_ClinVar_v062021.csv). The numbers of missense variants were 5,436 and 3,103, respectively. Information for each variant, including the GOF/LOF label, gene name, mRNA RefSeq ID (NM_), and HGVSp, was extracted from the two datasets. Using the gene names and RefSeq IDs, we retrieved each variant's corresponding protein reference sequence from the UniProt database. These protein reference sequences served as direct inputs to ClearVariant and, when combined with HGVSp, were used to derive variant sequences—another input to ClearVariant. Finally, duplicates were removed on the basis of the variant sequence. A total of 8,270 variants remained, of which 4,984 were missense variants. A total of 7,316 (4,656 missense) variants were from the HGMD-based dataset, and 3,795 (2,483 missense) variants were from the ClinVar-based dataset. There were 2,841 (2,155 missense) duplicate variants between the two datasets. No duplicate variants had both LOF and GOF labels. Additional LOF/GOF dataset We obtained 1,665 labelled pathogenic GOF and LOF missense variants from Stein et al . 15 . We constructed an LOF/GOF variant dataset using the same method used by Bayrak et al . along with chronological updates of known pathogenic variants. The dataset was generated on the basis of variants from the 2021.3 professional version HGMD and the 2023.08.13 version of ClinVar. We downloaded the training and test sets of LoGoFunc from https://gitlab.com/itan-lab/logofunc . The training and test sets contained LOF/GOF labels and LoGoFunc input features for each variant but did not contain variant IDs such as HGVSp. To match the variant IDs with LOF/GOF labels for each variant, we downloaded LoGoFunc predictions from their website ( https://itanlab.shinyapps.io/goflof/ ). It contains LoGoFunc predictions for canonical missense variants in the human genome, providing both the variant ID and LoGoFunc features for each variant. Since the variant IDs corresponded to DNA-level variants in VCF format, we used VEP (v104) ( https://asia.ensembl.org/info/docs/tools/vep/ ) with the –hgvs option to annotate the protein-level variants, HGVSp. ProteinGym We downloaded deep mutational scanning (DMS) substitution datasets from the ProteinGym website on 2024-08-23 ( https://proteingym.org/ ) 20 . We used this dataset to test whether our model could predict the direction of protein mutation within a gene. Each DMS substitution dataset in ProteinGym contains substitution mutations for one gene and corresponding DMS assay scores. The DMS assays conducted for each dataset can be broadly categorized into one of the following five types: activity, binding, expression, organismal fitness, or stability. We used 92 datasets for human proteins among 217 datasets for various species. Detailed information about the datasets used is available in Supplementary Table 2. The DMS assay scores were min–max scaled to a value between 0 and 1 for efficient and stable training of the deep learning model. For the test set, the scaling parameters were determined exclusively from the training set and then applied to the test set. Coevolution Coevolution refers to the accumulation of interdependencies between specific residues during the evolutionary of protein sequences. These interdependencies, shaped by natural selection, reflect the structural constraints and functional properties of proteins, providing crucial insights for predicting protein 3D structures and the effects of mutations. We downloaded the evolutionary coupling dataset from the Debbie Marks laboratory website ( https://marks.hms.harvard.edu/evmutation/ ) 26 . This dataset contains evolutionary coupling scores for each pair of residues within the domain of each protein. When the domain length was L, the top L residue pairs with the highest scores were defined as coevolution pairs. A higher score indicates greater interdependency between two residues, and a score of 0 might indicate that no coupling was detected in multiple sequence alignments. Coevolution data were obtained for 44 out of the 92 proteins in the ProteinGym dataset. AlphaFold2 protein structure AlphaFold2 is a SOTA model for predicting protein structures from amino acid sequences. A total of 90 human protein structures included in AlphaFold2 protein structure v4 (UP000005640_9606_HUMAN_v4) were downloaded from the AlphaFold protein structure database ( https://alphafold.ebi.ac.uk/ ). AlphaMissense AlphaMissense is a model for predicting the pathogenicity of missense variants and is fine-tuned from the protein structure prediction model AlphaFold2. By utilizing the structural information and evolutionary conservation data learned by AlphaFold2, it has achieved state-of-the-art performance across a wide range of genes. We downloaded the predicted pathogenic scores calculated with AlphaMissense from https://console.cloud.google.com/storage/browser/dm_alphamissense;tab=objects?pli=1&inv=1&invt=AbjufA&prefix=&forceOnObjectsSortingFiltering=false . A prediction score is allocated for all possible human missense variants by amino acid substitutions. 3Cnet 3Cnet is a model for predicting the pathogenicity of all types of genetic variants, including missense variants. By utilizing the amino acid context of human variants, it has achieved state-of-the-art performance across various types of variants. We downloaded 3Cnet from its github ( https://github.com/KyoungYeulLee/3Cnet.git ). All possible missense variants were generated for 67,584 human RefSeq sequences using 21 amino acids, including selenocysteine. 3Cnet prediction scores were measured for a total of 779,456,981 variants. ClearVariant model architecture The key principle of ClearVariant’s architecture consists of two key components: i) the use of ESM2, which has learned protein evolutionary information, and ii) leveraging ESM2 to compare reference protein sequences with their mutated counterparts. ESM2 has been trained on nearly all protein sequences found in nature (UniRef50) with masked language modelling (MLM). As a result, ESM2 can embed protein sequences into vectors that encapsulate key aspects of protein structure and function. The direction of mutational effects can be measured by comparing the function of the reference protein with the function of the mutated protein. We used two ESM2 modules to embed the reference protein sequence and the mutated protein sequence. The information contained in these two vectors effectively serves as information for predicting the direction of mutational effects. We designed a classification module to extract key information for distinguishing GOF/LOF mutations from ESM2-embedded vectors. To predict the direction of mutational effects, we hypothesized that the functionality of the reference protein, the functionality of the mutated protein, and the changes between these two states are crucial. To analyse these three components collectively, we concatenated the embedded vectors of the reference protein and the mutated protein at the same protein positions. This approach allows for a comprehensive analysis of the values and differences between the reference and mutated proteins for each residue. The concatenated vectors are subsequently passed through an interpretation layer, which incorporates a self-attention mechanism, enabling the model to identify residue-specific patterns relevant to GOF and LOF prediction. We anticipate that certain changes in the embedded vectors of individual residues caused by mutations will provide more critical information for GOF/LOF prediction, and we expect the interpretation layer to effectively learn and prioritize this information. Finally, the CLS vector, which represents the overall protein sequence and variant information, was passed through a dense layer to predict the GOF/LOF label. This architecture aims to capture and utilize the most informative signals for accurate mutational effect direction classification. The parameters of the ESM2 pretrained model were downloaded from Hugging Face using a Python module. The classification module was designed using the PyTorch Python module. Learning protein sequences to predict GOF and LOF variants Model training The missense variants with pathogenic GOF/LOF labels downloaded from Bayrak et al . were used for training. The input features included the protein reference sequence and the mutated protein sequence, whereas the output labels were binary (GOF/LOF). The input features were tokenized using the pretrained ESM tokenizer, downloaded via the Hugging Face Python module, and used as inputs to the model. The labels were one-hot encoded. To address the class imbalance in the dataset (with LOF mutations being approximately nine times more frequent than GOF mutations), we utilized the imbalanced data sampler from Python’s torchsampler module. This sampler oversamples the smaller class (i.e., GOF) and undersamples the larger class (i.e., LOF) to ensure that the model learns equally from both labels during training (Supplementary Fig. 20). In the test set, the 9:1 ratio was maintained because it reflects the real-world scenarios encountered in medical practice. The model was trained using BCEWithLogitsLoss as the loss function, and AdamW as the optimizer. Performance was evaluated through 5-fold cross-validation, employing stratified random splitting on the basis of the GOF/LOF labels. The performances on both the test set and the chronological test set were evaluated, with the mean and 95% confidence interval for each metric calculated from the results of 5-fold cross-validation. We evaluated whether learning the directional effect of mutations by comparing variant sequences with reference sequences improved predictive performance compared with learning from variant sequences alone. We built a model using a single ESM-2 encoder trained on variant sequences and compared its performance with that of ClearVariant. Controls (other tools and basic controls) We evaluated the performance of random baselines or existing tools. All controls were assessed using the same 5-split dataset previously employed for performance evaluation in ClearVariant. Basic controls We evaluated the performance of our model using four basic control methods: random, all-GOF, all-LOF, and gene bias. The random control method predicts each variant in the test set as either a GOF or an LOF with an equal probability of 50:50. All the GOF controls assume that all the variants in the test set are GOF mutations, whereas all the LOF controls predict all the variants as LOF mutations. The gene-biased control predicts test set labels on the basis of the label distribution of each gene in the training set. For genes where only GOF or only LOF variants are observed in the training set, the test set variants for those genes are assigned the same label. In cases where both GOF and LOF variants are present for a gene in the training set, test set predictions are made randomly. These controls provide baseline metrics for evaluating the performance of ClearVariant. LoGoFunc LoGoFunc is a state-of-the-art model designed to predict LOF and GOF variants. To compare its performance with that of ClearVariant, we downloaded the training and evaluation code provided in LoGoFunc and assessed its performance on a comparable dataset. The comparable dataset was defined by two criteria: first, the use of the same 5-split dataset; second, training solely on genetic features. To ensure consistency with the 5-split dataset, we identified the features of variants in the ClearVariant dataset from the LoGoFunc dataset and used them for training. Among the 4,984 variants in the ClearVariant dataset, 4,301 variants were available in the LoGoFunc dataset and used for training and evaluation. LoGoFunc provides a wide range of training features, including population data and other annotations. For this comparison, only genetic features were used as training inputs. Out of the 501 features offered by LoGoFunc, 399 genetic features were selected for the analysis (detailed in Supplementary Table 3). Gerasimavicius et al . Gerasimavicius et al . reported that the absolute change in protein stability caused by variants is significantly greater for LOF variants than for GOF variants. Since this finding is based on analysis rather than a predictive model, we directly used the dataset provided in their study for performance evaluation. This approach may represent a more lenient method of performance evaluation than models that divide the data into distinct training and test sets. We utilized data from the “HGMD_variant_level” sheet in the “DiseaseMech_Stability_VEPS.xlsx” file, which was downloaded from https://osf.io/h62fq/ . This dataset includes information on the functional classification of each variant (GOF or LOF, indicated by the “Disease_mechanism” field) and the corresponding absolute change in protein structural stability (“abs_FoldX_Monomer”). Using the absolute change in protein stability as the predictive feature and the functional classification as the label, we determined the threshold that achieved the highest F1 score to differentiate between the GOF and LOF variants with the training set. Variants with “abs_FoldX_Monomer” values below the threshold were predicted as GOF variants, whereas those with values above the threshold were predicted as LOF variants in the test set. MutPred2 MutPred2 is a widely used tool for predicting the pathogenicity of missense variants and classifying molecular mechanisms, such as GOF or LOF. We downloaded the standalone tool executable from the MutPred2 website ( http://mutpred.mutdb.org/index.html ) and evaluated its performance on our test sets. The datasets used in ClearVariant were converted into input formats compatible with MutPred2. The default command provided on the MutPred2 help page ( http://mutpred.mutdb.org/help.html ) was used for analysis without modifications. To assess the GOF and LOF prediction performance, we analysed the "molecular mechanism" terms in the MutPred2 output. Variants in the test set were classified as GOF variants if the highest probability term contained "gain" and as LOF variants if it contained "loss". However, since variants in the test set might have already been included in MutPred2’s training data, these performance results could overestimate the tool’s true predictive ability compared with experiments with a strict separation of training and test sets. Learning protein DMS assay Model training Among the ProteinGym DMS substitution datasets, 92 human protein datasets were used for training and evaluation. The input features included the protein reference sequence and the mutated protein sequence, whereas the output labels had continuous values between 0 and 1 (min–max scaled DMS score). The input features were tokenized using the pretrained ESM tokenizer, downloaded via the Hugging Face Python module, and used as inputs to the model. The labels are scaled between 0 and 1. The model was trained using MSELoss as the loss function, and AdamW as the optimizer. Performance was measured by randomly dividing the training and test sets at a 4:1 ratio. The model's performance was evaluated at the epoch with the highest r2 score on the test set. Biological explanation of inference outcomes To evaluate whether the model reflects the biological characteristics, we analysed the relationships between the predicted scores and amino acid properties. The impact of a substitution is often proportional to the chemical disparity between the original and substituted amino acids. We examined whether this biological knowledge is reflected in ClearVariant's predictions. For each of the models trained on one of the 92 datasets, we generated predictions for all possible single substitution variants. Each variant was created by substituting one position in the protein reference sequence with one of the 20 standard amino acids. For each protein position, the similarity between the predicted DMSs for different substitutions was calculated. $$\:Similarity\left(pos,\:aa1,\:aa2\right)=1-|Prediction\left(pos,\:aa1\right)-Prediciton\left(pos,\:aa2\right)|$$ where \(\:Prediction\left(10,\:aa1\right)\) represents the predicted DMS when the amino acid at the 10th position of the protein is substituted with aa1. As a result, substitutions with the same amino acid were assigned a similarity score of 1, and similarity scores were always ≤ 1. Each data point in the boxplot represents the median of the similarity values calculated for each dataset. Amino acids are grouped according to the characteristics of their side chains, including positive charge, negative charge, polar uncharged, and hydrophobicity. $$\:DataPoint\left(Dataset,aa1,\:aa2\right)=\:{Median}_{\:pos}\left(Similarity\left(pos,aa1,aa2\right)\right)$$ To evaluate whether the model reflects biological characteristics, we analysed the relationship between whether each residue has residue‒residue interactions and the predicted scores. The presence of interactions for each residue was determined using the protein structure predicted by AlphaFold. To achieve this goal, we conducted separate experiments on the basis of two independent criteria. The first criterion considers a residue to have an interaction if the shortest atomic distance between the residue and any other residue is less than 3 Å (Min3), which is usually used for potential strong interactions such as hydrogen bonds, ionic bonds, or π‒π stacking. The second criterion considered a residue to interact if the C-alpha distance between the residue and any other residue was less than 5 Angstroms (Ca5). The interaction status of each residue was determined separately for each criterion and then compared with the prediction scores. For 32 proteins from the 92 proteins ProteinGym 20 expression, stability, and activity collections, whose wild-type experimental values could be identified from the original publications, those wild-type values were used in the analysis. Within-model interpretation at amino acid residue resolution We evaluated whether the features the model focuses on during prediction align with biologically well-established knowledge. In the classification module of ClearVariant, the concatenated ESM-embedded vector is first processed by the interpretation layer. Subsequently, only the embedded vector corresponding to the CLS token (CLS vector) is passed through a dense layer to produce the final prediction. The CLS token is a placeholder positioned at the beginning of the input sequence, containing no intrinsic information. During fine-tuning, it is specifically utilized to aggregate information relevant to prediction. In our case, the interpretation layer learns to encode information that is critical for determining the direction of mutational effects into the CLS vector. More specifically, it is trained to assign higher attention weights to residues critical for prediction, thereby ensuring that their information is effectively captured in the CLS vector 23 . We identified residues that are critical for determining the direction of mutational effects. These include the residue of interest (ROI) where the substitution mutation occurs and residues coevolving with the ROI. For each mutated sequence, we examined whether these critical residues received higher attention weights than did all other residues that are not ROIs or that coevolve with ROIs. For each dataset, we generated all possible single substitution variants and extracted the attention values to generate the CLS vector in the interpretation layer for each variant. We subsequently analysed the distribution of attention scores between the ROI, coevolving residues, and all other residues, both at the domain level and across the entire protein. GOF/LOF inference dataset for the science community For the characterisation of predictions, we utilised proteome-wide predictions of pathogenic or likely pathogenic missense variants. We applied ClearVariant inference to two sets of missense variants: (i) those annotated as pathogenic or likely pathogenic with a review status of at least one star in ClinVar (release date 2024-09-04), and (ii) those predicted to be pathogenic by AlphaMissense. Each variant corresponds to a specific nucleotide substitution resulting in an altered amino acid sequence from the RefSeq reference. Allele frequency (AF) data were obtained from gnomAD version 4.1.0 (non-UK Biobank subset). For the minor allele frequency (MAF), we used the global allele frequency. Variants in gnomAD were linked to our altered amino acid sequences using the corresponding protein RefSeq ID (NP_). Declarations ACKNOWLEDGMENTS We thank all the members of the artificial intelligence and bioinformatics team at 3billion for their helpful discussions. S.K. and D.H. conceived and designed the experiments. S.K. and D.H. performed the experiments. S.K., K.K., and D.H. prepared the data. S.K., W.C., J. H., and D.H. analysed the results. S.K. and D.H. wrote the paper. AUTHOR INFORMATION Authors and Affiliations Department of Artificial Intelligence, Research and Development Center, 3billion, Inc., Seoul, Republic of Korea Sungnam Kim, Wonseok Chung & Doyeon Ha Department of Bioinformatics, Research and Development Center, 3billion, Inc., Seoul, Republic of Korea Kisang Kwon & Joohyun Han Interdisciplinary Program of Medical Informatics, College of Medicine, Seoul National University, Seoul, Republic of Korea Joohyun Han SUPPLEMENTARY DATA Supplementary data are available at Nature Genetics online. DATA AVAILABILITY The source code of ClearVariant is available at public GitHub (https://github.com/Doyeon-Ha/ClearVariant). Predictions for all pathogenic human missense variants and amino acid substitutions are available at Zenodo (https://doi.org/10.5281/zenodo.15429133) 50 . ETHICS DECLARATIONS Competing interests Financial Support: none declared. Conflict of Interest: none declared. References Lek, M. et al. Analysis of protein-coding genetic variation in 60,706 humans. Nature 536 , 285–291 (2016). Richards, S. et al. Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. Genetics in Medicine 17 , 405–424 (2015). Rastogi, R. et al. Critical assessment of missense variant effect predictors on disease-relevant variant data. Hum Genet 144 , 281–293 (2025). Livesey, B. J. et al. Guidelines for releasing a variant effect predictor. Genome Biol 26 , 97 (2025). Cheng, J. et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science (1979) 381 , eadg7492 (2023). Brandes, N., Goldman, G., Wang, C. H., Ye, C. J. & Ntranos, V. Genome-wide prediction of disease variant effects with a deep protein language model. Nat Genet 55 , 1512–1522 (2023). Lin, Y. J., Menon, A. S., Hu, Z. & Brenner, S. E. Variant Impact Predictor database (VIPdb), version 2: trends from three decades of genetic variant impact predictors. Hum Genomics 18 , 90 (2024). Karchin, R. et al. Improving transparency of computational tools for variant effect prediction. Nat Genet 56 , 1324–1326 (2024). Pejaver, V. et al. Inferring the molecular and phenotypic impact of amino acid variants with MutPred2. Nat Commun 11 , 5918 (2020). Boisson-Dupuis, S. et al. Inborn errors of human STAT1: allelic heterogeneity governs the diversity of immunological and infectious phenotypes. Curr Opin Immunol 24 , 364–378 (2012). Guo, Q., Feng, X. & Zhou, Y. PCSK9 Variants in Familial Hypercholesterolemia: A Comprehensive Synopsis. Front Genet 11 , 1020 (2020). Benjannet, S., Hamelin, J., Chrétien, M. & Seidah, N. G. Loss- and gain-of-function PCSK9 variants: Cleavage specificity, dominant negative effects, and low density lipoprotein receptor (LDLR) degradation. Journal of Biological Chemistry 287 , 33745–33755 (2012). Srivastava, R. A. K. A., Ajit, R. & Srivastava, K. A Review of Progress on Targeting LDL Receptor-Dependent and -Independent Pathways for the Treatment of Hypercholesterolemia, a Major Risk Factor of ASCVD. Cells 12 , 1648 (2023). Gerasimavicius, L., Livesey, B. J. & Marsh, J. A. Loss-of-function, gain-of-function and dominant-negative mutations have profoundly different effects on protein structure. Nature 13 , 3895 (2022). Stein, D. et al. Genome-wide prediction of pathogenic gain- and loss-of-function variants from ensemble learning of a diverse feature set. Genome Med 15 , 1 (2023). Vu, M. H. et al. Linguistically inspired roadmap for building biologically reliable protein language models. Nat Mach Intell 5 , 485–496 (2023). Unsal, S. et al. Learning functional properties of proteins with language models. Nat Mach Intell 4 , 227–245 (2022). Paysan-Lafosse, T. et al. InterPro in 2022. Nucleic Acids Res 51 , D418–D427 (2023). Landrum, M. J. et al. ClinVar: improvements to accessing data. Nucleic Acids Res 48 , D835–D844 (2020). Notin, P. et al. ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design. Adv Neural Inf Process Syst 36 , 64331–64379 (2023). O’Leary, N. A. et al. Reference sequence (RefSeq) database at NCBI: current status, taxonomic expansion, and functional annotation. Nucleic Acids Res 44 , D733–D745 (2016). Xiao, Y. et al. Protein Large Language Models: A Comprehensive Survey. ArXiv (2025). Wiegreffe, S. & Pinter, Y. Attention is not not Explanation. ArXiv (2019). Stenson, P. D. et al. The Human Gene Mutation Database (HGMD®): optimizing its use in a clinical diagnostic or research setting. Hum Genet 139 , 1197–1207 (2020). Grantham, R. Amino Acid Difference Formula to Help Explain Protein Evolution. Science (1979) 185 , 862–864 (1974). Hopf, T. A. et al. Mutation effects predicted from sequence co-variation. Nat Biotechnol 35 , 128–135 (2017). Won, D. G., Kim, D. W., Woo, J. & Lee, K. 3Cnet: pathogenicity prediction of human variants using multitask learning with evolutionary constraints. Bioinformatics 37 , 4626–4634 (2021). Maleki, F., Ovens, K., McQuillan, I. & Kusalik, A. J. Size matters: how sample size affects the reproducibility and specificity of gene set analysis. Hum Genomics 13 , 42 (2019). Chen, V. et al. Applying interpretable machine learning in computational biology—pitfalls, recommendations and opportunities for new developments. Nat Methods 21 , 1454–1461 (2024). Liu, Y. et al. Rotamer-free protein sequence design based on deep learning and self-consistency. Nat Comput Sci 2 , 451–462 (2022). Henikoff, S. & Henikoff, J. G. Amino acid substitution matrices from protein blocks. Proc Natl Acad Sci U S A 89 , 10915 (1992). Gromiha, M. M. & Selvaraj, S. Inter-residue interactions in protein folding and stability. Prog Biophys Mol Biol 86 , 235–277 (2004). McDonald, I. K. & Thornton, J. M. Satisfying Hydrogen Bonding Potential in Proteins. J Mol Biol 238 , 777–793 (1994). Richards, F. M. Areas, volumes, packing and protein structure. Annu Rev Biophys Bioeng 6 , 151–176 (1977). Majithia, A. R. et al. Prospective functional classification of all possible missense variants in PPARG. Nat Genet 48 , 1570–1575 (2016). Rochel, N. et al. Recurrent activating mutations of PPARγ associated with luminal bladder tumors. Nat Commun 10 , 253 (2019). Gilpin, L. H. et al. Explaining explanations: An overview of interpretability of machine learning. 2018 IEEE 5th International Conference on data science and advanced analytics (DSAA) IEEE , 80–89 (2018). Bartos, D. C. et al. A KCNQ1 Mutation Causes a High Penetrance for Familial Atrial Fibrillation. J Cardiovasc Electrophysiol 24 , 562 (2013). Chen, Y. H. et al. KCNQ1 gain-of-function mutation in familial atrial fibrillation. Science (1979) 299 , 251–254 (2003). Olesen, M. S. et al. Mutations in the potassium channel subunit KCNE1 are associated with early-onset familial atrial fibrillation. BMC Med Genet 13 , 24 (2012). Liu, P. et al. Reanalysis of Clinical Exome Sequencing Data. New England Journal of Medicine 380 , 2478–2480 (2019). Salfati, E. L. et al. Re-analysis of whole-exome sequencing data uncovers novel diagnostic variants and improves molecular diagnostic yields for sudden death and idiopathic diseases. Genome Med 11 , 1–8 (2019). Minikel, E., Painter, J., Dong, C. & Nelson, M. Refining the impact of genetic evidence on clinical success. Nature 629 , 624–629 (2024). Trajanoska, K. et al. From target discovery to clinical drug development with human genetics. Nature 620 , 737–745 (2023). Vaswani, A. et al. Attention is All you Need. Adv Neural Inf Process Syst 30 , (2017). Ponomarenko, E. A. et al. The size of the human proteome: the width and depth. Int J Anal Chem 2016 , 7436849 (2016). Rajeevan, H., Soundararajan, U., Kidd, J. R., Pakstis, A. J. & Kidd, K. K. ALFRED: an allele frequency resource for research and teaching. Nucleic Acids Res 40 , D1010–D1015 (2012). Gene Ontology Consortium. The gene ontology resource: 20 years and still GOing strong. Nucleic Acids Res 47 , D330–D338 (2019). Halldorsson, B. V et al. The sequences of 150,119 genomes in the UK Biobank. Nature 607 , 732–740 (2022). Learning sequence to predict gain- or loss-of-function variants. https://zenodo.org/records/15429133. Additional Declarations There is NO Competing Interest. Supplementary Files Hasupplementarytable1GofLofDetermination.xlsx Supplementary Table 1 Hasupplementarytable2GofLofDetermination.xlsx Supplementary Table 2 Hasupplementarytable3GofLofDetermination.xlsx Supplementary Table 3 HasupplementaryinformationGofLofDetermination.docx Supplementary Figures HasupplementaryinformationGofLofDeterminationforreportingsummarychecklist.docx Supplementary Figures Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6705195","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":467083000,"identity":"433e7941-b5dc-47ce-bf26-729475bda965","order_by":0,"name":"Doyeon Ha","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA7ElEQVRIiWNgGAWjYBACA2YGxgNw3gcgZmMnrIUBroVxBkgLMyEtDEhamHnAJAEt5uzMBw58+HXYnn9G8uHPNr+2yfMBnfrhYw5uLZbNbAkHZ/YdTpxxIy1NOrfvtmEbMwOz5MxteBx2mMfgMG/P4QSGM2fMmHN7bjMCtbAx8xKhxV7+zPnPny17btsTp4Xnx2HGDcd7GKQZftxOJKgF4peG9MSNx9vMJHsbbie3MTM24/WLOf/hgw8+/LG2lzvM/PjDjz+3bee3Nx/88BGPFjBgbGuGMcBkAwH1IPCnDsYgQvEoGAWjYBSMOAAAk9RWeCq3n4kAAAAASUVORK5CYII=","orcid":"https://orcid.org/0000-0002-6574-7311","institution":"3billion, Inc.","correspondingAuthor":true,"prefix":"","firstName":"Doyeon","middleName":"","lastName":"Ha","suffix":""},{"id":467083001,"identity":"936fd55e-4bab-4638-851b-073c1a4ce4e9","order_by":1,"name":"Sungnam Kim","email":"","orcid":"https://orcid.org/0000-0002-1427-3515","institution":"Pohang University of Science and Technology","correspondingAuthor":false,"prefix":"","firstName":"Sungnam","middleName":"","lastName":"Kim","suffix":""},{"id":467083002,"identity":"15ba29b5-3dec-4e40-b31f-a2c6c62c255d","order_by":2,"name":"Kisang Kwon","email":"","orcid":"","institution":"3billion, Inc.","correspondingAuthor":false,"prefix":"","firstName":"Kisang","middleName":"","lastName":"Kwon","suffix":""},{"id":467083003,"identity":"5afeb624-7760-4809-adee-1561dbacb2b8","order_by":3,"name":"Wonseok Chung","email":"","orcid":"","institution":"3billion, Inc.","correspondingAuthor":false,"prefix":"","firstName":"Wonseok","middleName":"","lastName":"Chung","suffix":""},{"id":467083004,"identity":"f4bd659e-71eb-4221-b7e4-b889f19e8464","order_by":4,"name":"Joohyun Han","email":"","orcid":"","institution":"3billion, Inc.","correspondingAuthor":false,"prefix":"","firstName":"Joohyun","middleName":"","lastName":"Han","suffix":""}],"badges":[],"createdAt":"2025-05-20 08:01:03","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-6705195/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6705195/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84059343,"identity":"4bf2e4ab-596b-40c2-b77a-be2cab2fa40a","added_by":"auto","created_at":"2025-06-06 09:56:40","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":248237,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the ClearVariant framework.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea,\u003c/strong\u003e Model architecture of ClearVariant, consisting of two ESM-2 encoders, an interpretation layer, and a classification head. Each ESM-2 encoder embeds the variant and reference protein sequences separately, followed by residue-level concatenation. No layer in the model is frozen during training. \u003cstrong\u003eb,\u003c/strong\u003e Binary classification: predicting gain-of-function and loss-of-function variants in clinical assessment databases. \u003cstrong\u003ec,\u003c/strong\u003e Real-value regression: predicting deep mutational scanning assayresults. \u003cstrong\u003ed,\u003c/strong\u003e Biological explanation demonstrating whether ClearVariant predictions follow known biological rules on the basis of the similar chemical properties of amino acid residues. \u003cstrong\u003ee,\u003c/strong\u003e Within-model interpretation to verify whether key features with high attention scores in ClearVariant align with residue evolution relationships.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/2260312ce8cde5205c9114c4.png"},{"id":84059769,"identity":"78e06d9a-cebf-41ae-a488-e6ea28303f52","added_by":"auto","created_at":"2025-06-06 10:04:40","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":575029,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePrecise prediction of variant functional effects.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea-d,\u003c/strong\u003e Performance scores of the ClearVariant and baseline models for predicting gain-of-function (GOF) and loss-of-function (LOF) variants. (\u003cstrong\u003ea\u003c/strong\u003e) GOF precision, (\u003cstrong\u003eb\u003c/strong\u003e) GOF F1 score, (\u003cstrong\u003ec\u003c/strong\u003e) LOF precision, and (\u003cstrong\u003ed\u003c/strong\u003e) LOF F1 score. The average and standard error were measured using 5-fold cross-validation. ClearVariant, the basic control model with simple logic, and the models reported in previous studies are shown with blue, grey, and pink bars, respectively. The top score is shown in bold. \u003cstrong\u003ee,f,\u003c/strong\u003e Precision‒recall curves for the GOF and LOF predictions made by the ClearVariant and baseline models. The average precision over recall is represented by a solid line, whereas the standard deviation is indicated by a shaded region. \u003cstrong\u003eg,\u003c/strong\u003e Proportion of actual GOF and LOF variants among those predicted as GOF and LOF variants by ClearVariant and LoGoFunc. The proportions are calculated as the average over 5-fold validation. The two-sided Fisher’s exact test was used to assess statistical significance. Notably, LoGoFunc requires a variety of data types, resulting in lower inference coverage. \u003cstrong\u003eh,\u003c/strong\u003e Kernel density estimate (KDE) plot of AlphaMissense scores for the GOF and LOF variants (\u003cem\u003eP\u003c/em\u003e= N.S., Mann–Whitney U test). \u003cstrong\u003ei,\u003c/strong\u003e Changes in the GOF precision over 1,000 epochs for ClearVariant (5-fold) in blue and model with the randomized parameters of the pretrained ESM-2 encoder in grey.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/0a253da4cfefd152f0ff93e5.png"},{"id":84059345,"identity":"52a22e85-088f-40db-8d09-9a84c1dffc6d","added_by":"auto","created_at":"2025-06-06 09:56:40","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":588496,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBiological explanation of inference outcomes.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea-d,\u003c/strong\u003e Similarity in ClearVariant mutation effect predictions on the basis of the chemical properties of the variant residue. The similarity was calculated using the mutation effect prediction values of two variants and defined as 1 minus the absolute difference between the predictions for two variants. Boxplots showing the similarity in the two mutation effect predictions when one variant residue of the two variants is (\u003cstrong\u003ea\u003c/strong\u003e) charged positive, (\u003cstrong\u003eb\u003c/strong\u003e) charged negative, (\u003cstrong\u003ec\u003c/strong\u003e) polar uncharged, or (\u003cstrong\u003ed\u003c/strong\u003e) hydrophobic. The Mann–Whitney U test was used to assess statistical significance. Statistical significance was denoted * for \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.01, ** for \u003cem\u003eP\u003c/em\u003e \u0026lt; 1.0 × 10⁻⁵, and *** for \u003cem\u003eP\u003c/em\u003e \u0026lt; 1.0 × 10⁻¹⁰. \u003cstrong\u003ee\u003c/strong\u003e,\u003cstrong\u003ef,\u003c/strong\u003e Observed and predicted stability indices of each variant in the \u003cem\u003eILF3\u003c/em\u003e protein. The x-axis represents the actual protein sequence position, and the y-axis represents the variant residue. \u003cstrong\u003eg,\u003c/strong\u003e Scatter plot showing the predicted and observed stability indices for the \u003cem\u003eILF3\u003c/em\u003e variants. \u003cstrong\u003eh,\u003c/strong\u003e Median DMS scores of residues with and without inter-residue interactions in 32 Proteingym datasets with WT scores. The DMS scores were min–max scaled such that the wild-type score was fixed at 0.5, with the minimum and maximum scores set to 0 and 1, respectively. The red dots represent mutated residues involved in inter-residue interactions, whereas the blue dots represent mutated residues without such interactions. Each pair of dots connected by a solid line corresponds to the same dataset. If the interacting residue has a higher DMS score, the line is coloured red; if the non-interacting residue has a higher score, the line is coloured blue. Dark-coloured dots represent residue pairs with a C-alpha distance ≤ 5 Å, whereas light-coloured dots represent cases where the shortest atomic distance is ≤ 3 Å, defining an interaction. \u003cstrong\u003ei,\u003c/strong\u003e Distribution of the activity indices of the \u003cem\u003ePPARG\u003c/em\u003e variants, categorized by the presence or absence of residue‒residue interactions, where interactions are defined as atomic distances ≤ 3 Å (P \u0026lt; 5.00 × 10\u003csup\u003e-324\u003c/sup\u003e, Mann–Whitney U test). The protein structure was predicted using AlphaFold2, and the predicted mean activity index for all 19 possible substitutions at each residue was visualized using colour mapping.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/4d9efa8f9f032e04c8ff90fd.png"},{"id":84059346,"identity":"3fe72667-fed9-4ba9-bfbf-5511f5ed6174","added_by":"auto","created_at":"2025-06-06 09:56:41","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":543636,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eWithin-model interpretation at amino acid residue resolution.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea-c,\u003c/strong\u003e Within-model interpretation using attention scores from the interpretation layer. (\u003cstrong\u003ea\u003c/strong\u003e) The concatenated embedding vector passes through the interpretation layer followed by the classification head. (\u003cstrong\u003eb\u003c/strong\u003e) In the interpretation layer, the CLS vector formed by the weighted sum is passed to the classification head, resulting in model inference. (\u003cstrong\u003ec\u003c/strong\u003e) The attention scores for each residue were compared with coevolution residue pairs. \u003cstrong\u003ed,e,\u003c/strong\u003e The density distribution of attention scores for coevolution residues was subtracted from the density distribution for non-coevolution residues. A negative y-axis value indicates that coevolution residues have a higher density at the corresponding attention value on the x-axis. (\u003cstrong\u003ed\u003c/strong\u003e) Density difference across the entire OTC protein. (\u003cstrong\u003ee\u003c/strong\u003e) Density difference in the \u003cem\u003eL-ornithine-binding domain\u003c/em\u003e of the \u003cem\u003eOTC\u003c/em\u003e protein. The density distributions of coevolution and non-co-evolution were significantly different for both the OTC protein and the OTC domain (both P \u0026lt; 5.00 × 10\u003csup\u003e-324\u003c/sup\u003e, Mann–Whitney U test). \u003cstrong\u003ef,g,\u003c/strong\u003e Structure of the \u003cem\u003eL-ornithine-binding domain\u003c/em\u003e of the \u003cem\u003eOTC\u003c/em\u003e protein. Among the 174 residues, (\u003cstrong\u003ef\u003c/strong\u003e) the 50 residues with the highest attention values, (\u003cstrong\u003eg\u003c/strong\u003e) coevolution residues and the residue of interest (ROI) are highlighted in different colours. The protein structure was predicted using AlphaFold2. \u003cstrong\u003eh,i,\u003c/strong\u003e For each protein in the 92 ProteinGym dataset, (\u003cstrong\u003eh\u003c/strong\u003e) the median attention score of coevolved residues divided by the median attention score of other residues and (\u003cstrong\u003ei\u003c/strong\u003e) the median attention score of the ROI divided by the median attention score of other residues are shown. A ratio greater than 0.0 indicates that coevolution residues (or ROIs) have higher attention scores than do non-coevolution residues (or other residues). The x-axis is log10-transformed. Dots are coloured blue if they are greater than 0.0 and pink if they are less than 0.0. The Wilcoxon signed rank test was used to assess how positive values have statistically larger absolute values than do negative values.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/f6fe33dcc556fbf11cde0947.png"},{"id":84059770,"identity":"e9bf468f-414e-4ce4-bd04-4cc191b75eaf","added_by":"auto","created_at":"2025-06-06 10:04:41","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":315670,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eClearVariant predictions as a community resource.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ea,\u003c/strong\u003e Example row of the dataset containing GOF/LOF predictions made by ClearVariant. This dataset includes missense variants that are either annotated as pathogenic in ClinVar or predicted to be likely pathogenic by AlphaMissense. Each altered amino acid sequence was generated from the corresponding RefSeq sequence. Each row contains the gene name, RefSeq transcript ID, UniProt protein ID, amino acid-level variant notation (HGVSp), ClinVar annotation, AlphaMissense prediction, ClearVariant's GOF/LOF prediction, and an indicator of whether the protein containing the variant was included in the ClearVariant training set. \u003cstrong\u003eb-d\u003c/strong\u003e, Pie charts showing the distributions of the GOF and LOF predictions across different datasets. (\u003cstrong\u003eb\u003c/strong\u003e) Distribution in the ClearVariant training set. (\u003cstrong\u003ec\u003c/strong\u003e) Predictions for variants annotated as pathogenic or likely pathogenic in ClinVar. (\u003cstrong\u003ed\u003c/strong\u003e) Predictions for variants predicted to be pathogenic by AlphaMissense. \u003cstrong\u003ee\u003c/strong\u003e, GOF and LOF distributions stratified by the minor allele frequency (MAF).\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/b46c54b035e55ed9a39ce06e.png"},{"id":108491017,"identity":"3b2c5abe-83f7-40ae-9b43-0107f87b9d08","added_by":"auto","created_at":"2026-05-05 09:51:12","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2747928,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/f8791ab5-2c0b-4fcd-822d-e9baa00797f4.pdf"},{"id":84059339,"identity":"656e56bb-edb1-47f6-80e9-e2ccf165ba92","added_by":"auto","created_at":"2025-06-06 09:56:40","extension":"xlsx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":33446,"visible":true,"origin":"","legend":"Supplementary Table 1","description":"","filename":"Hasupplementarytable1GofLofDetermination.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/0bd9451b1fc9cb945d944b05.xlsx"},{"id":84059768,"identity":"c7b428f5-d0c8-4340-9331-1d2fbd75c494","added_by":"auto","created_at":"2025-06-06 10:04:40","extension":"xlsx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":18389,"visible":true,"origin":"","legend":"Supplementary Table 2","description":"","filename":"Hasupplementarytable2GofLofDetermination.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/cc845f005aa8432ea327077c.xlsx"},{"id":84059342,"identity":"63926721-7fb0-458c-a116-a9d9438c8526","added_by":"auto","created_at":"2025-06-06 09:56:40","extension":"xlsx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":26583,"visible":true,"origin":"","legend":"Supplementary Table 3","description":"","filename":"Hasupplementarytable3GofLofDetermination.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/b6eb795ebe34a9568cea7334.xlsx"},{"id":84059772,"identity":"132f8806-71f1-47c5-b7dc-16f7c45f82bc","added_by":"auto","created_at":"2025-06-06 10:04:41","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":1993763,"visible":true,"origin":"","legend":"Supplementary Figures","description":"","filename":"HasupplementaryinformationGofLofDetermination.docx","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/e2262bc0f71f86e38508953a.docx"},{"id":84059349,"identity":"00253909-8375-40f3-84e6-c762fd553263","added_by":"auto","created_at":"2025-06-06 09:56:41","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":1996987,"visible":true,"origin":"","legend":"Supplementary Figures","description":"","filename":"HasupplementaryinformationGofLofDeterminationforreportingsummarychecklist.docx","url":"https://assets-eu.researchsquare.com/files/rs-6705195/v1/acf7484c8fb65a94233b5d34.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"Learning sequence to predict gain- or loss-of-function variants","fulltext":[{"header":"Introduction","content":"\u003cp\u003eElucidating the phenotypic consequences of genetic mutations is a fundamental goal in genetics and biomedical research\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Since the effects of all mutational alterations on the human genome cannot be observed or confirmed, variant effect predictors (VEPs) are used to predict the functional impacts of genetic mutations. The American College of Medical Genetics and Genomics (ACMG) Standards and Guidelines, aimed at improving genetic diagnosis and clinical decision-making, have provided recommendations for the use of computational methods to assess the pathogenicity of genetic variants\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Recently, several studies have reported the effectiveness of performing in silico predictions using multiple VEPs and proposed a higher level of evidence strength\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e,\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eRecent progress in artificial intelligence has led to the development of deep learning models. Cheng \u003cem\u003eet al\u003c/em\u003e. reported that the performance scores of their deep learning model assessing missense variants within multiple sequence alignment (MSA) coverage were the highest across numerous clinical and experimental pathogenic benchmarks\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. The model was introduced as an adaptation of AlphaFold, which is a neural network-based model previously built for predicting 3D protein structures with approximately 93\u0026nbsp;million trainable parameters. Moreover, Brandes \u003cem\u003eet al\u003c/em\u003e. used evolutionary scale modelling (ESM1b), a protein language model with 650\u0026nbsp;million parameters, to predict the effects of not only missense variants but also coding variants affecting multiple residues, including in-frame indel and nonsense mutations\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eHowever, the mechanisms of pathogenic variants are heterogeneous. Most VEP tools are limited in their ability to predict whether a mutation is pathogenic or benign\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. As a recent review article reported, none of the widely used computational tools used to predict variant effects, except for MutPred2, can predict mutational effects on gene function\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. This limitation leads to the next step in computational mutation analysis: identifying whether the mutations are either gain-of-function (GOF) mutations or loss-of-function (LOF) mutations. Tools that are used to predict the mutational effects on a variant of interest could be called directional variant effect predictors (D-VEPs).\u003c/p\u003e \u003cp\u003eThe classification of mutational effects on gene function is important for obtaining a clear understanding of disease mechanisms. GOF mutations typically confer proteins with enhanced or novel functions that can affect cellular processes whereas LOF mutations result in complete or partial loss of normal protein function\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. For example, the PCSK9 D374Y mutation, labelled as a GOF mutation, increases the binding affinity of PCSK9 to low-density lipoprotein receptors (LDL-Rs) and is thus considered a causal determinant of hypercholesterolemia, a disease characterized by elevated levels of LDL cholesterol\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. In contrast, PCSK9 LOF mutations, such as R46L and G106R, are associated with hypocholesterolemia, which is a disease characterized by low levels of LDL cholesterol\u003csup\u003e\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e,\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. Importantly, elucidating mutational effects on the functions of genes such as PCSK9, LDL-R, APOB, and ANGPTL3 has facilitated sophisticated diagnostic strategies and the identification of drug targets for cholesterol-related diseases\u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. The biological mechanisms of LOF mutations are well established; however, the mechanisms underlying functional enhancement or the acquisition of a new function are more complicated\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Existing D-VEP methods, including MutPred2 and LoGoFunc, exhibit poor GOF mutation prediction performance due to an insufficient amount of biological information for models to generate a specific inference\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. The low amount of data on GOF also makes it difficult to identify new GOF mutations\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eThe ability to precisely predict GOF mutations is connected to understanding protein language, as it relies on recognizing the nuanced relationships between a protein's sequence and its function\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e,\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e. Like how written words form coherent meanings in natural languages, functional sequence regions, including specific motifs and domains within the gene sequence, dictate the functions of the encoded proteins in biological systems\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Protein language models (PLMs) depend heavily on the broad sequence context, which might distinguish between GOF and LOF mutations. The increasing collection of annotations on variant functionality has supported deep learning training with mutated protein sequences. ClinVar and the Human Gene Mutation Database (HGMD) harbour information on mutational effects and their phenotypic consequences\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Multiplexed assays of variant effects (MAVEs) systematically measure protein variant effects, and the data have been organized as large-scale experimental benchmarks\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Furthermore, approximately 12\u0026nbsp;million reference sequences across over a million species have been collected over time, providing the opportunity to train models with wild-type and variant protein languages\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. With increasingly large genetic datasets, PLM training operates at the amino acid level by embedding vectors for each residue\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. By employing attention mechanisms, these models enable the interpretation of residue importance, specifically for identifying which residues are key to determining the effects of mutations.\u003c/p\u003e \u003cp\u003eHere, we present ClearVariant, a model for classifying GOF and LOF mutations. The model explicitly integrates both reference and variant sequences to facilitate direct comparisons of mutational effects and leverages a self-attention mechanism to enhance interpretability. Evaluations using datasets on pathogenic mutations across diverse genes revealed that ClearVariant outperforms existing models, achieving a GOF prediction precision of 0.814 and an F1 score of 0.778, which are much higher than the GOF prediction precision of 0.578 and F1 score of 0.670 achieved by the previous state-of-the-art (SOTA) model. Furthermore, experimentally measured single-gene mutational effects demonstrate that its inference outcomes and decision-making process align well with established biological insights, underscoring its reliability for functional variant prediction.\u003c/p\u003e"},{"header":"Results","content":"\u003cp\u003eLearning protein sequences to predict GOF and LOF variants\u003c/p\u003e \u003cp\u003eWe developed ClearVariant, an attention-based deep learning model, to predict GOF and LOF variants on the basis of ESM-2 (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ea; Methods)\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. ESM-2 was trained using 250\u0026nbsp;million protein sequences for protein structure prediction. We restructured the model architecture of ESM-2 for the following two principles. First, a protein\u0026rsquo;s reference sequence and variant sequence were jointly considered since the direction of the mutational effects is assigned by comparing consequences of mutated genes to those of wild-type genes. Therefore, it is necessary to explicitly introduce the reference sequence to provide context for direct comparison. Proteins with GOF mutations have enhanced functions or new functions, whereas proteins with LOF mutations present either partial or complete loss of function relative to the WT protein\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u003c/sup\u003e. Second, the model architecture was structured to allow interpretability by incorporating a self-attention mechanism within the classification layer. Attention refers to the weighting of input features on the basis of their relevance to the model\u0026rsquo;s decision-making process\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e. The CLS (classification) vector is subsequently used as the input of the classification head.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eTo validate the predictive performance of ClearVariant regarding mutational effects, we conducted training and evaluation experiments using two distinct datasets. The first dataset comprised pathogenic GOF and LOF mutations spanning a diverse array of genes from the ClinVar database and the HGMD (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003eb)\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. The other dataset consisted of experimental measurements quantifying the effects of mutations in a single gene obtained from ProteinGym (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ec)\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. To ensure the reliability of the model's predictions, we performed \u0026lsquo;explanation\u0026rsquo; and \u0026lsquo;interpretation\u0026rsquo; analyses. In the \u0026lsquo;explanation\u0026rsquo; analysis, we verified that the model's predictions aligned with established biological patterns such as amino acid chemical properties (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ed)\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. In the \u0026lsquo;interpretation\u0026rsquo; analysis, we evaluated whether the underlying rationale of the model's decisions was consistent with biological information that was not explicitly provided during training (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003ee). By analysing the attention placed between residue pairs within the model, we can interpret the extent to which each residue interaction provides crucial information for GOF and LOF prediction. For example, we assessed whether residues associated through coevolution were recognized as key determinants in predicting mutational effects\u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003ePrecise prediction of mutational effects on gene function\u003c/p\u003e \u003cp\u003eTo assess the performance of ClearVariant in predicting mutational effects, we compared basic controls and previous D-VEP models in all available pathogenic missense variants with GOF and LOF labels from the ClinVar database and the HGMD\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. ClearVariant achieved a GOF precision of 0.814 and a GOF F1 score of 0.778, outperforming all the other controls, including the previous SOTA model, LoGoFunc, with a GOF precision of 0.578 and a GOF F1 score of 0.670 (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea and b)\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. In LOF prediction, which is a relatively easy task where performance has converged, ClearVariant performs similarly to the other basic models (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ec and d). To more accurately evaluate the overall performance of the model, we examined the precision‒recall curve with positive labels of GOF and LOF (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ee and f), and also assessed the receiver operating characteristic (ROC) curve for comprehensive evaluation (Supplementary Figs.\u0026nbsp;1 and 2). Thus, the model revealed a greater proportion of correct answers in the GOF prediction (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eg; odds ratio\u0026thinsp;=\u0026thinsp;99.24 and \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2.43 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;87\u003c/sup\u003e (ClearVariant), odds ratio\u0026thinsp;=\u0026thinsp;34.31 and \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2.55 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;58\u003c/sup\u003e (LoGoFunc)). ClearVariant exhibited better performance than a single ESM-2 encoder, achieving higher MCC values in the GOF and LOF predictions (Supplementary Fig.\u0026nbsp;3). Basic controls included i) \u003cem\u003erandom\u003c/em\u003e, ii) \u003cem\u003eall one label\u003c/em\u003e, and iii) \u003cem\u003egene-biased\u003c/em\u003e prediction (gray bars in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea-d; Methods). The control D-VEP models included LoGoFunc, MutPred2, and Gerasimavicius, etc. (salmon bars in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea-d; Methods)\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e,\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e,\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. Notably, the \u003cem\u003egene-biased\u003c/em\u003e prediction had high prediction performance comparable with those of the previous models, although it is a simple rule-based model where inference is decided by other labels on variants of the gene.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eIt is not surprising that no significant difference between GOF and LOF variants was observed in the pathogenic scores calculated by AlphaMissense or 3Cnet (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eh and Supplementary Fig.\u0026nbsp;4; \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;N.S., Mann\u0026ndash;Whitney \u003cem\u003eU\u003c/em\u003e test)\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e,\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. Similarly, the pathogenicity scores of the predicted GOF and LOF variants were similar (Supplementary Fig.\u0026nbsp;5; \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;N.S., Mann\u0026ndash;Whitney \u003cem\u003eU\u003c/em\u003e test). Previous VEPs have not been developed to predict the direction of the mutation effect; thus, most VEPs are not suitable for predicting GOF and LOF variants.\u003c/p\u003e \u003cp\u003eSmall biological datasets often present challenges in AI due to inherent characteristics such as high measurement noise and complex data structures, which can lead to greater variability in model performance during cross-validation\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. Driven by this concern, we proceeded k-fold cross validation (k\u0026thinsp;=\u0026thinsp;5) (blue lines in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ei; Methods). An ESM-2 model with randomized parameters showed far lower performance between the initial and final epochs, highlighting the importance of pretraining the model with protein language (gray line in Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ei). Furthermore, we evaluated the GOF and LOF prediction performance of ClearVariant for variants classified as likely pathogenic and likely benign by AlphaMissense in the test set of the 5-fold cross-validation (Supplementary Fig.\u0026nbsp;6). The results revealed that ClearVariant exhibited greater predictive performance across all evaluation metrics for variants classified as likely pathogenic by AlphaMissense. This might be due to relatively larger data numbers in AM-likely pathogenic variants (Supplementary Fig.\u0026nbsp;7), and correct prediction by ClearVariant is achieved with a larger dataset (Supplementary Fig.\u0026nbsp;8).\u003c/p\u003e \u003cp\u003eBiological explanation of inference outcomes\u003c/p\u003e \u003cp\u003eHuman-understandable explanations are required for the results of biological analyses with in silico tools, leading to their application in subsequent research and medical fields\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e,\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e\u003c/sup\u003e. In studies involving DNA or protein sequences, the results can be explained in terms of amino acid residues, which biological researchers intuitively understand\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. First, in the collection of benchmarks for mutational effects, changes to amino acid groups by residues with similar characteristics are more likely to have similar mutational effects (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea-d, Supplementary Fig.\u0026nbsp;9)\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. \u003cem\u003eA\u003c/em\u003emino acids can be categorized into four groups on the basis of their side chain properties: \u003cem\u003echarged positive (CP), charged negative (CN), polar uncharged (PU), and hydrophobic (HP)\u003c/em\u003e. Specifically, one residual change to \u003cem\u003eCP\u003c/em\u003e groups has a similar mutational effect with other residual changes to \u003cem\u003eCP\u003c/em\u003e groups compared with changes to \u003cem\u003eCN\u003c/em\u003e, \u003cem\u003ePU\u003c/em\u003e, or \u003cem\u003eHP\u003c/em\u003e groups (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ea). The similarity in mutational effects driven by the chemical properties of amino acids is consistently maintained for wild-type and variant amino acid pairs (Supplementary Figs.\u0026nbsp;10 and 11). For example, in interleukin enhancer binding factor 3 (\u003cem\u003eILF3\u003c/em\u003e), mutational changes to \u003cem\u003eCN\u003c/em\u003e amino acids (Asp and Glu) at residues 3 and 4 increased the stability of the protein, whereas changes to \u003cem\u003eHP\u003c/em\u003e amino acids (Ala, Val, Ile, Leu, Met, Phe, Tyr, and Trp) decreased the stability of both the observed values in ProteinGym and the values predicted by ClearVariant (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ee and f). The predicted and observed values of ILF3 mutations were highly significantly correlated (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eg; \u003cem\u003eρ\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.936 and \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;5.00 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;324\u003c/sup\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAmino acid residues form interresidue interactions within a protein to construct wild-type structures of a protein; thus, those interaction residues essentially contribute to the protein\u0026rsquo;s functions and mechanisms\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e\u003c/sup\u003e. Compared with variants at non-interaction residues, variants at interaction residues exhibit a greater reduction in predicted values relative to that of the WT (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eh and Supplementary Fig.\u0026nbsp;12). Two standards were utilized for the detection of residual interactions: Ca5 and Min3 (Methods)\u003csup\u003e\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. We analysed the within-protein interactions using 32 proteins that had AlphaFold protein structures and WT experimental values. Under Ca5 standard conditions, 4 out of 32 proteins (12.5%) presented higher values for their interaction residues than did the WT, whereas 11 out of 32 proteins (34.4%) presented elevated values for their non-interaction residues than did the WT. For example, the mutational effects predicted by ClearVariant for peroxisome proliferator-activated receptor γ (PPARG) were significantly different between interaction and non-interaction residues (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ei). Surprisingly, variants with non-interaction residues presented higher activity values than the WT did. Residues located in the protein interior are more likely to engage in residue‒residue interactions\u003csup\u003e\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e,\u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e35\u003c/span\u003e\u003c/sup\u003e. When the average score of the predicted activity index values for all 19 possible missense variants at each residue were mapped onto the protein structure, we observed that residues in the protein interior tended to have lower predicted activity index values than the wild type did, whereas residues with higher predicted values were predominantly located on the protein surface. The PPARG mutation M280I, located on the protein surface, is related to overexpression, which is a gain-of-function mechanism (green colour in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003ei)\u003csup\u003e\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e. These findings are consistent with those of previous studies, suggesting functional differences between the core and surface of proteins\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eClearVariant had a Spearman median of 0.73 and a Pearson median of 0.78 across all 92 human proteins (Supplementary Fig.\u0026nbsp;13). However, the standard deviation was large, which means that the prediction performance varied across proteins. To investigate this, we measured performance across different experimental categories and assessed data coverage. As a result, we found that only the data coverage of the variant ratio of stability, highest prediction performance, was greater than that of the other experimental categories (Supplementary Figs.\u0026nbsp;14 and 15).\u003c/p\u003e \u003cp\u003eWithin-model interpretation at amino acid residue resolution\u003c/p\u003e \u003cp\u003eUnderstanding the specific information an AI model uses as the basis for its decisions is essential for improving the model's reliability\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. In ClearVariant, the self-attention layer of the interpretation module plays a critical role in comparing the mutated protein sequence with the reference protein sequence at the residue level, determining which residue's information should be prioritized for prediction (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ea). Residues receiving higher attention weights are deemed more significant and contribute more substantially to the formation of downstream representations (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eb). Since the final prediction is made using the CLS vector, analysing the attention weights that contribute to forming this vector allows us to identify the residues that ClearVariant considers crucial for predicting the direction of mutational effects (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ec).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eAttention analysis revealed that coevolution residues provide key information for predicting the direction of mutational effects. 44 mutational effect datasets with information on protein coevolution were utilized for attention analyses\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. We obtained the attention distribution for four residue groups: 1) the residue where the variant occurs, referred to as the residue of interest (ROI), 2) the coevolved residues with the ROI, and 3) all other residues. In the ornithine transcarbamylase (OTC) protein enzymatic activity dataset, coevolution residues presented a greater distribution of attention scores than did all other residues, both when all missense variants across the entire protein were considered and when only those within a specific domain were analysed (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ed and e). This finding suggests that, for each missense variant, coevolving residues consistently receive high attention scores (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003ef and g).\u003c/p\u003e \u003cp\u003eThe median attention weights indicated that, in most cases, the ROIs and coevolved residues had higher attention weights than the other residues did (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eh and i; coevolution of in-protein, \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;7.33 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;6\u003c/sup\u003e, coevolution of in-domain; \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;4.51 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;5\u003c/sup\u003e, ROI of in-protein; \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;3.09 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;6\u003c/sup\u003e, ROI of in-domain; and \u003cem\u003eP\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1.05 \u0026times; 10\u003csup\u003e\u0026minus;\u0026thinsp;5\u003c/sup\u003e Wilcoxon signed rank test; Supplementary Fig.\u0026nbsp;16). Notably, when ROIs or coevolution residues had higher attention weights, the differences were often substantial, exceeding 10\u003csup\u003e4\u003c/sup\u003e-fold in some instances. Conversely, in cases where other residues had higher attention weights, the differences were marginal, typically close to onefold. The median attention weight for coevolution residues in 33 models was greater than that for other residues, with the highest cases showing a 10\u003csup\u003e3.5\u003c/sup\u003e-fold difference. This suggests that the model effectively learns the biological significance of interactions involving coevolution residues, which are well known to play a critical role in protein functions. Additionally, to understand whether high performance in the gene-biased model (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003ea-d) is derived from research bias or the evolution of protein language, we scrutinized the degree of expression, stability and activity changes caused by a missense mutation. If the proportions are distributed at both extremes, most genes might be considered affected by research bias. We observed a wide distribution in the proportion of variants with experimental scores higher than those of the WT, meaning that each protein could obtain divergent protein languages by evolving in separate paths to different proportions of GOFs and LOFs (Supplementary Fig.\u0026nbsp;17, Methods).\u003c/p\u003e \u003cp\u003eGOF/LOF inference dataset for the science community\u003c/p\u003e \u003cp\u003eWe provide the integrated resource of GOF/LOF inference on pathogenic variants by ClearVariant, and inference is conducted on pathogenic variants (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ea). Four additional annotations are tagged: i) ClinVar pathogenicity annotation, ii) AlphaMissense prediction, iii) ClearVariant prediction, and iv) whether the variant belongs to a gene encountered during training. Notably, ClearVariant predicted the KCNQ1 p.Arg231His variant\u0026mdash;reported in multiple families with Familial Atrial Fibrillation (FAF)\u0026mdash;as a GOF variant\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. This prediction aligns with previous studies suggesting that GOF mutations in \u003cem\u003eKCNQ1\u003c/em\u003e are associated with the development of FAF\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e,\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. The number of GOF variants was 723 (14.5%) for the ClinVar/HGMD variants mapped with GOF/LOF (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003eb and Supplementary Table\u0026nbsp;1), 20,831 (14.2%) for the ClearVariant\u0026rsquo;s prediction of ClinVar pathogenic variants (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ec), and 7.7 M (9.4%) for the ClearVariant\u0026rsquo;s prediction of AM-likely pathogenic variants (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ed). The GOF percentages are between 8.7% and 15.4% across diverse intervals of the minor allele frequency (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003ee).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe development of D-VEPs, defined as VEPs with a determined direction of mutational effects, could lead to new insights into various biological fields from the mechanistic study of diseases to genetic testing for diagnosis and drug development. The diagnostic yield of Mendelian disorders has steadily increased over time. For example, an after-five-year reanalysis of paediatric neurological diseases increased the yield from 24.8\u0026ndash;46.8%\u003csup\u003e41\u003c/sup\u003e. One of the core issues for undiagnosed cases is the inability to identify causal variant candidates. The progressive development of VEPs in recent years has contributed to molecular diagnosis and reanalysis by facilitating the selection of causal variants\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e\u003c/sup\u003e. By enabling molecular diagnostic specialists to identify the alignment between the patient's symptoms and mutation effects, D-VEP could increase the diagnostic rates of rare diseases and decreases the diagnostic turnaround time, enhancing both patient quality of life and survival outcomes.\u003c/p\u003e \u003cp\u003eTrained on human genomes and clinical annotations, D-VEPs could yield a significant contribution to drug development by accelerating the entire pipeline. Human genetic evidence significantly increases the probability of FDA approval, making drugs with genetic evidence 2.6 times more likely to receive approval than those without\u003csup\u003e\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e,\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e. Genetics-based approaches aid in the selection of appropriate drug targets. Identifying diseases driven by GOF mutations and subsequently targeting the affected proteins with inhibitory compounds is a representative strategy for drug development. Furthermore, D-VEPs can find a variant with a reverse effect on the causal variant using data on protein functional annotations. These mutations, provided that they result in LOF mechanisms, offer a basis for developing inhibitory therapies that replicate the protective effects. For example, the first genetic evidence that ANGPTL3 is associated with hypobetalipoproteinaemia was reported in 2010\u003csup\u003e44\u003c/sup\u003e. After 11 years, by mimicking the beneficial effects of LOF variants in target genes, EVINACUMAB, an angiopoietin-related protein 3 inhibitor, was approved for this disease. This strategy, with the assistance of D-VEPs, will revolutionize the identification of targets for drug development.\u003c/p\u003e \u003cp\u003eThe prediction performance of an in silico tool alone does not guarantee its adoption in medical applications\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. The reliability of such tools, especially deep learning models, in predicting pathogenicity or mutational effects can be determined by two factors: i) the outputs of the tool are aligned with established biological knowledge, and ii) the prediction mechanism of the tool is human-understandable. To evaluate the former, which we call \u0026lsquo;explanation\u0026rsquo;, we analysed the amino acid characteristics of the inference outputs from ClearVariant (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). We observed that alterations to the same chemical property tended to have similar mutational effects on a single residue. Interestingly, alterations in intraprotein interaction residues are more likely to be LOF mutations. For the latter, which we call \u0026lsquo;interpretation\u0026rsquo;, we scrutinized the internal representations of ClearVariant and observed that coevolution residues place considerable attention on the residue of interest (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e). This observation means that the inference process of ClearVariant is largely guided by the coevolutionary relationship with the residue of interest. Beyond prediction performance, \u0026lsquo;explanation\u0026rsquo; and \u0026lsquo;interpretation\u0026rsquo; analyses could make the model more usable in real-world applications, providing biological insights for further discussion.\u003c/p\u003e \u003cp\u003eRather than starting from scratch, ClearVariant was built on the basis of models reported in prior works in both artificial intelligence and life sciences. Vaswani \u003cem\u003eet al\u003c/em\u003e., whose work introduced the Transformer enabled parallel processing and improved efficiency by training a large amount of text sequence data with self-attention mechanisms\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e. By inheriting the structure and methodology of the Transformer, advanced PLMs like evolutionary scale modelling foundation models are refined to learn and interpret protein language within its broader biological context, through training a large volume of reference sequences across over 138\u0026nbsp;million proteins\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Pathogenic variant data mapped to GOFs and LOFs have been collected and imported into various databases, including the ClinVar database and the HGMD, over time\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. The determination of LOFs has been achieved by many tools based on simple standards, and even their prediction precision reached approximately 0.95 in a previous study (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e)\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. Unlike LOFs, GOFs are not easily identifiable, and the amount of available data on GOFs is significantly smaller, making GOF prediction highly challenging. Leveraging these sophisticated PLMs and the increasing availability of extensive biological data, ClearVariant\u0026mdash;a model fine-tuned with a large amount of labelled genomic data\u0026mdash;is poised to predict GOFs precisely.\u003c/p\u003e \u003cp\u003eTo unlock the full potential of D-VEP and broaden its applications, effectively addressing its current limitations is paramount. Indeed, the current scope of D-VEP often focuses on GOF and LOF, while the broader spectrum of mutational effects, including dominant-negative effects and haploinsufficiency, presents further complexities\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e. Encouragingly, the ongoing efforts to collect and annotate data on such diverse mutational effects are paving the way for more sophisticated and comprehensive predictive models\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e. Another recognized challenge lies in maintaining high prediction performance for variants emerging after a model's training period or for mutations in novel genes (Supplementary Figs.\u0026nbsp;2 and 18). This performance discrepancy often highlights the limitations of employed datasets; for instance, the dataset used in this study, while substantial, covered only 975 proteins with pathogenic GOF or LOF labels (Supplementary Table\u0026nbsp;1), representing a small fraction of the human proteome and thus a potentially biased subset (Supplementary Fig.\u0026nbsp;19)\u003csup\u003e\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e. Nevertheless, these hurdles, although their resolution requires considerable time, are surmountable. Significant improvements are anticipated through the strategic integration of diverse data types beyond gene sequences. For example, incorporating allele frequency data, phenotypic information from large population cohorts such as the UK Biobank, and detailed knowledge of signalling pathways and functional processes holds considerable promise for more accurately determining the direction and impact of mutational effects\u003csup\u003e\u003cspan additionalcitationids=\"CR48\" citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e. Ultimately, adaptive evolution of D-VEP, building upon methodologies such as those demonstrated herein, is anticipated to be pivotal for the future of genetics and biomedical research.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003eData Curation\u003c/h2\u003e \u003cp\u003eProtein reference sequence\u003c/p\u003e \u003cp\u003eWe collected protein reference sequences, used them as inputs for deep learning, and generated variant sequences. A total of 20,359 reviewed human protein reference sequences were downloaded on March 7, 2024, from the UniProtKB website (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://www.uniprot.org/uniprotkb?query=*\u003c/span\u003e\u003cspan address=\"https://www.uniprot.org/uniprotkb?query=*\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). Protein sequences and their corresponding gene names and UniProt and RefSeq IDs (NM_, NP_) were obtained. Gene names and IDs were used to match reference sequences with variants across multiple databases. ID versions were not considered.\u003c/p\u003e \u003cp\u003eTraining and test sets for LOF/GOF prediction\u003c/p\u003e \u003cp\u003eWe obtained pathogenic GOF and LOF variants from Bayrak \u003cem\u003eet al\u003c/em\u003e. The datasets were generated on the basis of known pathogenic variants from the HGMD and the ClinVar database. For each pathogenic variant, they searched for GOF- and LOF-related nomenclature in all relevant publications using natural language processing libraries. They curated each dataset using the 2019.1 professional version of the HGMD and the 2021.06 version of ClinVar. The GOF and LOF variant datasets were downloaded from their website (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://itanlab.shinyapps.io/goflof/\u003c/span\u003e\u003cspan address=\"https://itanlab.shinyapps.io/goflof/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). A total of 9,619 variants were obtained from the HGMD-derived dataset (goflof_HGMD2019_v032021_allfeat.csv), and 5,085 variants were obtained from the ClinVar-derived dataset (goflof_ClinVar_v062021.csv). The numbers of missense variants were 5,436 and 3,103, respectively.\u003c/p\u003e \u003cp\u003eInformation for each variant, including the GOF/LOF label, gene name, mRNA RefSeq ID (NM_), and HGVSp, was extracted from the two datasets. Using the gene names and RefSeq IDs, we retrieved each variant's corresponding protein reference sequence from the UniProt database. These protein reference sequences served as direct inputs to ClearVariant and, when combined with HGVSp, were used to derive variant sequences\u0026mdash;another input to ClearVariant. Finally, duplicates were removed on the basis of the variant sequence. A total of 8,270 variants remained, of which 4,984 were missense variants. A total of 7,316 (4,656 missense) variants were from the HGMD-based dataset, and 3,795 (2,483 missense) variants were from the ClinVar-based dataset. There were 2,841 (2,155 missense) duplicate variants between the two datasets. No duplicate variants had both LOF and GOF labels.\u003c/p\u003e \u003cp\u003eAdditional LOF/GOF dataset\u003c/p\u003e \u003cp\u003eWe obtained 1,665 labelled pathogenic GOF and LOF missense variants from Stein \u003cem\u003eet al\u003c/em\u003e.\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e. We constructed an LOF/GOF variant dataset using the same method used by Bayrak \u003cem\u003eet al\u003c/em\u003e. along with chronological updates of known pathogenic variants. The dataset was generated on the basis of variants from the 2021.3 professional version HGMD and the 2023.08.13 version of ClinVar.\u003c/p\u003e \u003cp\u003eWe downloaded the training and test sets of LoGoFunc from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://gitlab.com/itan-lab/logofunc\u003c/span\u003e\u003cspan address=\"https://gitlab.com/itan-lab/logofunc\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. The training and test sets contained LOF/GOF labels and LoGoFunc input features for each variant but did not contain variant IDs such as HGVSp. To match the variant IDs with LOF/GOF labels for each variant, we downloaded LoGoFunc predictions from their website (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://itanlab.shinyapps.io/goflof/\u003c/span\u003e\u003cspan address=\"https://itanlab.shinyapps.io/goflof/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). It contains LoGoFunc predictions for canonical missense variants in the human genome, providing both the variant ID and LoGoFunc features for each variant. Since the variant IDs corresponded to DNA-level variants in VCF format, we used VEP (v104) (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://asia.ensembl.org/info/docs/tools/vep/\u003c/span\u003e\u003cspan address=\"https://asia.ensembl.org/info/docs/tools/vep/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) with the \u0026ndash;hgvs option to annotate the protein-level variants, HGVSp.\u003c/p\u003e \u003cp\u003eProteinGym\u003c/p\u003e \u003cp\u003eWe downloaded deep mutational scanning (DMS) substitution datasets from the ProteinGym website on 2024-08-23 (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://proteingym.org/\u003c/span\u003e\u003cspan address=\"https://proteingym.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003csup\u003e20\u003c/sup\u003e. We used this dataset to test whether our model could predict the direction of protein mutation within a gene. Each DMS substitution dataset in ProteinGym contains substitution mutations for one gene and corresponding DMS assay scores. The DMS assays conducted for each dataset can be broadly categorized into one of the following five types: activity, binding, expression, organismal fitness, or stability. We used 92 datasets for human proteins among 217 datasets for various species. Detailed information about the datasets used is available in Supplementary Table\u0026nbsp;2.\u003c/p\u003e \u003cp\u003eThe DMS assay scores were min\u0026ndash;max scaled to a value between 0 and 1 for efficient and stable training of the deep learning model. For the test set, the scaling parameters were determined exclusively from the training set and then applied to the test set.\u003c/p\u003e \u003cp\u003eCoevolution\u003c/p\u003e \u003cp\u003eCoevolution refers to the accumulation of interdependencies between specific residues during the evolutionary of protein sequences. These interdependencies, shaped by natural selection, reflect the structural constraints and functional properties of proteins, providing crucial insights for predicting protein 3D structures and the effects of mutations. We downloaded the evolutionary coupling dataset from the Debbie Marks laboratory website (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://marks.hms.harvard.edu/evmutation/\u003c/span\u003e\u003cspan address=\"https://marks.hms.harvard.edu/evmutation/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e)\u003csup\u003e26\u003c/sup\u003e. This dataset contains evolutionary coupling scores for each pair of residues within the domain of each protein. When the domain length was L, the top L residue pairs with the highest scores were defined as coevolution pairs. A higher score indicates greater interdependency between two residues, and a score of 0 might indicate that no coupling was detected in multiple sequence alignments. Coevolution data were obtained for 44 out of the 92 proteins in the ProteinGym dataset.\u003c/p\u003e \u003cp\u003eAlphaFold2 protein structure\u003c/p\u003e \u003cp\u003eAlphaFold2 is a SOTA model for predicting protein structures from amino acid sequences. A total of 90 human protein structures included in AlphaFold2 protein structure v4 (UP000005640_9606_HUMAN_v4) were downloaded from the AlphaFold protein structure database (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://alphafold.ebi.ac.uk/\u003c/span\u003e\u003cspan address=\"https://alphafold.ebi.ac.uk/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eAlphaMissense\u003c/p\u003e \u003cp\u003eAlphaMissense is a model for predicting the pathogenicity of missense variants and is fine-tuned from the protein structure prediction model AlphaFold2. By utilizing the structural information and evolutionary conservation data learned by AlphaFold2, it has achieved state-of-the-art performance across a wide range of genes. We downloaded the predicted pathogenic scores calculated with AlphaMissense from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://console.cloud.google.com/storage/browser/dm_alphamissense;tab=objects?pli=1\u0026amp;inv=1\u0026amp;invt=AbjufA\u0026amp;prefix=\u0026amp;forceOnObjectsSortingFiltering=false\u003c/span\u003e\u003cspan address=\"https://console.cloud.google.com/storage/browser/dm_alphamissense;tab=objects?pli=1\u0026amp;inv=1\u0026amp;invt=AbjufA\u0026amp;prefix=\u0026amp;forceOnObjectsSortingFiltering=false\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. A prediction score is allocated for all possible human missense variants by amino acid substitutions.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003e3Cnet\u003c/h3\u003e\n\u003cp\u003e3Cnet is a model for predicting the pathogenicity of all types of genetic variants, including missense variants. By utilizing the amino acid context of human variants, it has achieved state-of-the-art performance across various types of variants. We downloaded 3Cnet from its github (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/KyoungYeulLee/3Cnet.git\u003c/span\u003e\u003cspan address=\"https://github.com/KyoungYeulLee/3Cnet.git\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e). All possible missense variants were generated for 67,584 human RefSeq sequences using 21 amino acids, including selenocysteine. 3Cnet prediction scores were measured for a total of 779,456,981 variants.\u003c/p\u003e\n\u003ch3\u003eClearVariant model architecture\u003c/h3\u003e\n\u003cp\u003eThe key principle of ClearVariant\u0026rsquo;s architecture consists of two key components: i) the use of ESM2, which has learned protein evolutionary information, and ii) leveraging ESM2 to compare reference protein sequences with their mutated counterparts. ESM2 has been trained on nearly all protein sequences found in nature (UniRef50) with masked language modelling (MLM). As a result, ESM2 can embed protein sequences into vectors that encapsulate key aspects of protein structure and function. The direction of mutational effects can be measured by comparing the function of the reference protein with the function of the mutated protein. We used two ESM2 modules to embed the reference protein sequence and the mutated protein sequence. The information contained in these two vectors effectively serves as information for predicting the direction of mutational effects.\u003c/p\u003e \u003cp\u003eWe designed a classification module to extract key information for distinguishing GOF/LOF mutations from ESM2-embedded vectors. To predict the direction of mutational effects, we hypothesized that the functionality of the reference protein, the functionality of the mutated protein, and the changes between these two states are crucial. To analyse these three components collectively, we concatenated the embedded vectors of the reference protein and the mutated protein at the same protein positions. This approach allows for a comprehensive analysis of the values and differences between the reference and mutated proteins for each residue. The concatenated vectors are subsequently passed through an interpretation layer, which incorporates a self-attention mechanism, enabling the model to identify residue-specific patterns relevant to GOF and LOF prediction. We anticipate that certain changes in the embedded vectors of individual residues caused by mutations will provide more critical information for GOF/LOF prediction, and we expect the interpretation layer to effectively learn and prioritize this information. Finally, the CLS vector, which represents the overall protein sequence and variant information, was passed through a dense layer to predict the GOF/LOF label. This architecture aims to capture and utilize the most informative signals for accurate mutational effect direction classification.\u003c/p\u003e \u003cp\u003eThe parameters of the ESM2 pretrained model were downloaded from Hugging Face using a Python module. The classification module was designed using the PyTorch Python module.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eLearning protein sequences to predict GOF and LOF variants\u003c/h2\u003e \u003cp\u003eModel training\u003c/p\u003e \u003cp\u003eThe missense variants with pathogenic GOF/LOF labels downloaded from Bayrak \u003cem\u003eet al\u003c/em\u003e. were used for training. The input features included the protein reference sequence and the mutated protein sequence, whereas the output labels were binary (GOF/LOF).\u003c/p\u003e \u003cp\u003eThe input features were tokenized using the pretrained ESM tokenizer, downloaded via the Hugging Face Python module, and used as inputs to the model. The labels were one-hot encoded. To address the class imbalance in the dataset (with LOF mutations being approximately nine times more frequent than GOF mutations), we utilized the imbalanced data sampler from Python\u0026rsquo;s torchsampler module. This sampler oversamples the smaller class (i.e., GOF) and undersamples the larger class (i.e., LOF) to ensure that the model learns equally from both labels during training (Supplementary Fig.\u0026nbsp;20). In the test set, the 9:1 ratio was maintained because it reflects the real-world scenarios encountered in medical practice.\u003c/p\u003e \u003cp\u003eThe model was trained using BCEWithLogitsLoss as the loss function, and AdamW as the optimizer. Performance was evaluated through 5-fold cross-validation, employing stratified random splitting on the basis of the GOF/LOF labels.\u003c/p\u003e \u003cp\u003eThe performances on both the test set and the chronological test set were evaluated, with the mean and 95% confidence interval for each metric calculated from the results of 5-fold cross-validation.\u003c/p\u003e \u003cp\u003eWe evaluated whether learning the directional effect of mutations by comparing variant sequences with reference sequences improved predictive performance compared with learning from variant sequences alone. We built a model using a single ESM-2 encoder trained on variant sequences and compared its performance with that of ClearVariant.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eControls (other tools and basic controls)\u003c/h3\u003e\n\u003cp\u003eWe evaluated the performance of random baselines or existing tools. All controls were assessed using the same 5-split dataset previously employed for performance evaluation in ClearVariant.\u003c/p\u003e \u003cp\u003eBasic controls\u003c/p\u003e \u003cp\u003eWe evaluated the performance of our model using four basic control methods: random, all-GOF, all-LOF, and gene bias. The random control method predicts each variant in the test set as either a GOF or an LOF with an equal probability of 50:50. All the GOF controls assume that all the variants in the test set are GOF mutations, whereas all the LOF controls predict all the variants as LOF mutations. The gene-biased control predicts test set labels on the basis of the label distribution of each gene in the training set. For genes where only GOF or only LOF variants are observed in the training set, the test set variants for those genes are assigned the same label. In cases where both GOF and LOF variants are present for a gene in the training set, test set predictions are made randomly. These controls provide baseline metrics for evaluating the performance of ClearVariant.\u003c/p\u003e \u003cp\u003eLoGoFunc\u003c/p\u003e \u003cp\u003eLoGoFunc is a state-of-the-art model designed to predict LOF and GOF variants. To compare its performance with that of ClearVariant, we downloaded the training and evaluation code provided in LoGoFunc and assessed its performance on a comparable dataset. The comparable dataset was defined by two criteria: first, the use of the same 5-split dataset; second, training solely on genetic features.\u003c/p\u003e \u003cp\u003eTo ensure consistency with the 5-split dataset, we identified the features of variants in the ClearVariant dataset from the LoGoFunc dataset and used them for training. Among the 4,984 variants in the ClearVariant dataset, 4,301 variants were available in the LoGoFunc dataset and used for training and evaluation. LoGoFunc provides a wide range of training features, including population data and other annotations. For this comparison, only genetic features were used as training inputs. Out of the 501 features offered by LoGoFunc, 399 genetic features were selected for the analysis (detailed in Supplementary Table\u0026nbsp;3).\u003c/p\u003e \u003cp\u003eGerasimavicius \u003cem\u003eet al\u003c/em\u003e.\u003c/p\u003e \u003cp\u003eGerasimavicius \u003cem\u003eet al\u003c/em\u003e. reported that the absolute change in protein stability caused by variants is significantly greater for LOF variants than for GOF variants. Since this finding is based on analysis rather than a predictive model, we directly used the dataset provided in their study for performance evaluation. This approach may represent a more lenient method of performance evaluation than models that divide the data into distinct training and test sets.\u003c/p\u003e \u003cp\u003eWe utilized data from the \u0026ldquo;HGMD_variant_level\u0026rdquo; sheet in the \u0026ldquo;DiseaseMech_Stability_VEPS.xlsx\u0026rdquo; file, which was downloaded from \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://osf.io/h62fq/\u003c/span\u003e\u003cspan address=\"https://osf.io/h62fq/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e. This dataset includes information on the functional classification of each variant (GOF or LOF, indicated by the \u0026ldquo;Disease_mechanism\u0026rdquo; field) and the corresponding absolute change in protein structural stability (\u0026ldquo;abs_FoldX_Monomer\u0026rdquo;). Using the absolute change in protein stability as the predictive feature and the functional classification as the label, we determined the threshold that achieved the highest F1 score to differentiate between the GOF and LOF variants with the training set. Variants with \u0026ldquo;abs_FoldX_Monomer\u0026rdquo; values below the threshold were predicted as GOF variants, whereas those with values above the threshold were predicted as LOF variants in the test set.\u003c/p\u003e \u003cp\u003eMutPred2\u003c/p\u003e \u003cp\u003eMutPred2 is a widely used tool for predicting the pathogenicity of missense variants and classifying molecular mechanisms, such as GOF or LOF. We downloaded the standalone tool executable from the MutPred2 website (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://mutpred.mutdb.org/index.html\u003c/span\u003e\u003cspan address=\"http://mutpred.mutdb.org/index.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) and evaluated its performance on our test sets. The datasets used in ClearVariant were converted into input formats compatible with MutPred2. The default command provided on the MutPred2 help page (\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttp://mutpred.mutdb.org/help.html\u003c/span\u003e\u003cspan address=\"http://mutpred.mutdb.org/help.html\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e) was used for analysis without modifications.\u003c/p\u003e \u003cp\u003eTo assess the GOF and LOF prediction performance, we analysed the \"molecular mechanism\" terms in the MutPred2 output. Variants in the test set were classified as GOF variants if the highest probability term contained \"gain\" and as LOF variants if it contained \"loss\". However, since variants in the test set might have already been included in MutPred2\u0026rsquo;s training data, these performance results could overestimate the tool\u0026rsquo;s true predictive ability compared with experiments with a strict separation of training and test sets.\u003c/p\u003e\n\u003ch3\u003eLearning protein DMS assay\u003c/h3\u003e\n\u003cp\u003eModel training\u003c/p\u003e \u003cp\u003eAmong the ProteinGym DMS substitution datasets, 92 human protein datasets were used for training and evaluation. The input features included the protein reference sequence and the mutated protein sequence, whereas the output labels had continuous values between 0 and 1 (min\u0026ndash;max scaled DMS score).\u003c/p\u003e \u003cp\u003eThe input features were tokenized using the pretrained ESM tokenizer, downloaded via the Hugging Face Python module, and used as inputs to the model. The labels are scaled between 0 and 1.\u003c/p\u003e \u003cp\u003eThe model was trained using MSELoss as the loss function, and AdamW as the optimizer. Performance was measured by randomly dividing the training and test sets at a 4:1 ratio. The model's performance was evaluated at the epoch with the highest r2 score on the test set.\u003c/p\u003e \u003cp\u003eBiological explanation of inference outcomes\u003c/p\u003e \u003cp\u003eTo evaluate whether the model reflects the biological characteristics, we analysed the relationships between the predicted scores and amino acid properties. The impact of a substitution is often proportional to the chemical disparity between the original and substituted amino acids. We examined whether this biological knowledge is reflected in ClearVariant's predictions.\u003c/p\u003e \u003cp\u003eFor each of the models trained on one of the 92 datasets, we generated predictions for all possible single substitution variants. Each variant was created by substituting one position in the protein reference sequence with one of the 20 standard amino acids. For each protein position, the similarity between the predicted DMSs for different substitutions was calculated.\u003cdiv id=\"Equa\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e\n$$\\:Similarity\\left(pos,\\:aa1,\\:aa2\\right)=1-|Prediction\\left(pos,\\:aa1\\right)-Prediciton\\left(pos,\\:aa2\\right)|$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003ewhere \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:Prediction\\left(10,\\:aa1\\right)\\)\u003c/span\u003e\u003c/span\u003e represents the predicted DMS when the amino acid at the 10th position of the protein is substituted with aa1. As a result, substitutions with the same amino acid were assigned a similarity score of 1, and similarity scores were always\u0026thinsp;\u0026le;\u0026thinsp;1.\u003c/p\u003e \u003cp\u003eEach data point in the boxplot represents the median of the similarity values calculated for each dataset. Amino acids are grouped according to the characteristics of their side chains, including positive charge, negative charge, polar uncharged, and hydrophobicity.\u003cdiv id=\"Equb\" class=\"Equation\"\u003e\u003cdiv format=\"TEX\" class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e\n$$\\:DataPoint\\left(Dataset,aa1,\\:aa2\\right)=\\:{Median}_{\\:pos}\\left(Similarity\\left(pos,aa1,aa2\\right)\\right)$$\u003c/div\u003e\u003c/div\u003e\u003c/p\u003e \u003cp\u003eTo evaluate whether the model reflects biological characteristics, we analysed the relationship between whether each residue has residue‒residue interactions and the predicted scores. The presence of interactions for each residue was determined using the protein structure predicted by AlphaFold. To achieve this goal, we conducted separate experiments on the basis of two independent criteria. The first criterion considers a residue to have an interaction if the shortest atomic distance between the residue and any other residue is less than 3 \u0026Aring; (Min3), which is usually used for potential strong interactions such as hydrogen bonds, ionic bonds, or π‒π stacking. The second criterion considered a residue to interact if the C-alpha distance between the residue and any other residue was less than 5 Angstroms (Ca5). The interaction status of each residue was determined separately for each criterion and then compared with the prediction scores. For 32 proteins from the 92 proteins ProteinGym\u003csup\u003e\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e expression, stability, and activity collections, whose wild-type experimental values could be identified from the original publications, those wild-type values were used in the analysis.\u003c/p\u003e \u003cp\u003eWithin-model interpretation at amino acid residue resolution\u003c/p\u003e \u003cp\u003eWe evaluated whether the features the model focuses on during prediction align with biologically well-established knowledge. In the classification module of ClearVariant, the concatenated ESM-embedded vector is first processed by the interpretation layer. Subsequently, only the embedded vector corresponding to the CLS token (CLS vector) is passed through a dense layer to produce the final prediction.\u003c/p\u003e \u003cp\u003eThe CLS token is a placeholder positioned at the beginning of the input sequence, containing no intrinsic information. During fine-tuning, it is specifically utilized to aggregate information relevant to prediction. In our case, the interpretation layer learns to encode information that is critical for determining the direction of mutational effects into the CLS vector. More specifically, it is trained to assign higher attention weights to residues critical for prediction, thereby ensuring that their information is effectively captured in the CLS vector\u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e \u003cp\u003eWe identified residues that are critical for determining the direction of mutational effects. These include the residue of interest (ROI) where the substitution mutation occurs and residues coevolving with the ROI. For each mutated sequence, we examined whether these critical residues received higher attention weights than did all other residues that are not ROIs or that coevolve with ROIs. For each dataset, we generated all possible single substitution variants and extracted the attention values to generate the CLS vector in the interpretation layer for each variant. We subsequently analysed the distribution of attention scores between the ROI, coevolving residues, and all other residues, both at the domain level and across the entire protein.\u003c/p\u003e \u003cp\u003eGOF/LOF inference dataset for the science community\u003c/p\u003e \u003cp\u003eFor the characterisation of predictions, we utilised proteome-wide predictions of pathogenic or likely pathogenic missense variants. We applied ClearVariant inference to two sets of missense variants: (i) those annotated as pathogenic or likely pathogenic with a review status of at least one star in ClinVar (release date 2024-09-04), and (ii) those predicted to be pathogenic by AlphaMissense. Each variant corresponds to a specific nucleotide substitution resulting in an altered amino acid sequence from the RefSeq reference.\u003c/p\u003e \u003cp\u003eAllele frequency (AF) data were obtained from gnomAD version 4.1.0 (non-UK Biobank subset). For the minor allele frequency (MAF), we used the global allele frequency. Variants in gnomAD were linked to our altered amino acid sequences using the corresponding protein RefSeq ID (NP_).\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eACKNOWLEDGMENTS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe thank all the members of the artificial intelligence and bioinformatics team at 3billion for their helpful discussions. S.K. and D.H. conceived and designed the experiments. S.K. and D.H. performed the experiments. S.K., K.K., and D.H. prepared the data. S.K., W.C., J. H., and D.H. analysed the results. S.K. and D.H. wrote the paper.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAUTHOR INFORMATION\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAuthors and Affiliations\u003c/p\u003e\n\u003cp\u003eDepartment of Artificial Intelligence, Research and Development Center, 3billion, Inc., Seoul, Republic of Korea\u003c/p\u003e\n\u003cp\u003eSungnam Kim, Wonseok Chung \u0026amp; Doyeon Ha\u003c/p\u003e\n\u003cp\u003eDepartment of Bioinformatics, Research and Development Center, 3billion, Inc., Seoul, Republic of Korea\u003c/p\u003e\n\u003cp\u003eKisang Kwon \u0026amp; Joohyun Han\u003c/p\u003e\n\u003cp\u003eInterdisciplinary Program of Medical Informatics, College of Medicine, Seoul National University, Seoul, Republic of Korea\u003c/p\u003e\n\u003cp\u003eJoohyun Han\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eSUPPLEMENTARY DATA\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eSupplementary data are available at Nature Genetics online.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDATA AVAILABILITY\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe source code of ClearVariant is available at public GitHub (https://github.com/Doyeon-Ha/ClearVariant). Predictions for all pathogenic human missense variants and amino acid substitutions are available at Zenodo (https://doi.org/10.5281/zenodo.15429133)\u003csup\u003e50\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eETHICS DECLARATIONS\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCompeting interests\u003c/p\u003e\n\u003cp\u003eFinancial Support: none declared.\u003c/p\u003e\n\u003cp\u003eConflict of Interest: none declared.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eLek, M. \u003cem\u003eet al.\u003c/em\u003e Analysis of protein-coding genetic variation in 60,706 humans. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e536\u003c/strong\u003e, 285\u0026ndash;291 (2016).\u003c/li\u003e\n\u003cli\u003eRichards, S. \u003cem\u003eet al.\u003c/em\u003e Standards and guidelines for the interpretation of sequence variants: a joint consensus recommendation of the American College of Medical Genetics and Genomics and the Association for Molecular Pathology. \u003cem\u003eGenetics in Medicine\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 405\u0026ndash;424 (2015).\u003c/li\u003e\n\u003cli\u003eRastogi, R. \u003cem\u003eet al.\u003c/em\u003e Critical assessment of missense variant effect predictors on disease-relevant variant data. \u003cem\u003eHum Genet\u003c/em\u003e \u003cstrong\u003e144\u003c/strong\u003e, 281\u0026ndash;293 (2025).\u003c/li\u003e\n\u003cli\u003eLivesey, B. J. \u003cem\u003eet al.\u003c/em\u003e Guidelines for releasing a variant effect predictor. \u003cem\u003eGenome Biol\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 97 (2025).\u003c/li\u003e\n\u003cli\u003eCheng, J. \u003cem\u003eet al.\u003c/em\u003e Accurate proteome-wide missense variant effect prediction with AlphaMissense. \u003cem\u003eScience (1979)\u003c/em\u003e \u003cstrong\u003e381\u003c/strong\u003e, eadg7492 (2023).\u003c/li\u003e\n\u003cli\u003eBrandes, N., Goldman, G., Wang, C. H., Ye, C. J. \u0026amp; Ntranos, V. Genome-wide prediction of disease variant effects with a deep protein language model. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e55\u003c/strong\u003e, 1512\u0026ndash;1522 (2023).\u003c/li\u003e\n\u003cli\u003eLin, Y. J., Menon, A. S., Hu, Z. \u0026amp; Brenner, S. E. Variant Impact Predictor database (VIPdb), version 2: trends from three decades of genetic variant impact predictors. \u003cem\u003eHum Genomics\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e, 90 (2024).\u003c/li\u003e\n\u003cli\u003eKarchin, R. \u003cem\u003eet al.\u003c/em\u003e Improving transparency of computational tools for variant effect prediction. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e56\u003c/strong\u003e, 1324\u0026ndash;1326 (2024).\u003c/li\u003e\n\u003cli\u003ePejaver, V. \u003cem\u003eet al.\u003c/em\u003e Inferring the molecular and phenotypic impact of amino acid variants with MutPred2. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 5918 (2020).\u003c/li\u003e\n\u003cli\u003eBoisson-Dupuis, S. \u003cem\u003eet al.\u003c/em\u003e Inborn errors of human STAT1: allelic heterogeneity governs the diversity of immunological and infectious phenotypes. \u003cem\u003eCurr Opin Immunol\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 364\u0026ndash;378 (2012).\u003c/li\u003e\n\u003cli\u003eGuo, Q., Feng, X. \u0026amp; Zhou, Y. PCSK9 Variants in Familial Hypercholesterolemia: A Comprehensive Synopsis. \u003cem\u003eFront Genet\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 1020 (2020).\u003c/li\u003e\n\u003cli\u003eBenjannet, S., Hamelin, J., Chr\u0026eacute;tien, M. \u0026amp; Seidah, N. G. Loss- and gain-of-function PCSK9 variants: Cleavage specificity, dominant negative effects, and low density lipoprotein receptor (LDLR) degradation. \u003cem\u003eJournal of Biological Chemistry\u003c/em\u003e \u003cstrong\u003e287\u003c/strong\u003e, 33745\u0026ndash;33755 (2012).\u003c/li\u003e\n\u003cli\u003eSrivastava, R. A. K. A., Ajit, R. \u0026amp; Srivastava, K. A Review of Progress on Targeting LDL Receptor-Dependent and -Independent Pathways for the Treatment of Hypercholesterolemia, a Major Risk Factor of ASCVD. \u003cem\u003eCells\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 1648 (2023).\u003c/li\u003e\n\u003cli\u003eGerasimavicius, L., Livesey, B. J. \u0026amp; Marsh, J. A. Loss-of-function, gain-of-function and dominant-negative mutations have profoundly different effects on protein structure. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 3895 (2022).\u003c/li\u003e\n\u003cli\u003eStein, D. \u003cem\u003eet al.\u003c/em\u003e Genome-wide prediction of pathogenic gain- and loss-of-function variants from ensemble learning of a diverse feature set. \u003cem\u003eGenome Med\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 1 (2023).\u003c/li\u003e\n\u003cli\u003eVu, M. H. \u003cem\u003eet al.\u003c/em\u003e Linguistically inspired roadmap for building biologically reliable protein language models. \u003cem\u003eNat Mach Intell\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 485\u0026ndash;496 (2023).\u003c/li\u003e\n\u003cli\u003eUnsal, S. \u003cem\u003eet al.\u003c/em\u003e Learning functional properties of proteins with language models. \u003cem\u003eNat Mach Intell\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 227\u0026ndash;245 (2022).\u003c/li\u003e\n\u003cli\u003ePaysan-Lafosse, T. \u003cem\u003eet al.\u003c/em\u003e InterPro in 2022. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, D418\u0026ndash;D427 (2023).\u003c/li\u003e\n\u003cli\u003eLandrum, M. J. \u003cem\u003eet al.\u003c/em\u003e ClinVar: improvements to accessing data. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e48\u003c/strong\u003e, D835\u0026ndash;D844 (2020).\u003c/li\u003e\n\u003cli\u003eNotin, P. \u003cem\u003eet al.\u003c/em\u003e ProteinGym: Large-Scale Benchmarks for Protein Fitness Prediction and Design. \u003cem\u003eAdv Neural Inf Process Syst\u003c/em\u003e \u003cstrong\u003e36\u003c/strong\u003e, 64331\u0026ndash;64379 (2023).\u003c/li\u003e\n\u003cli\u003eO\u0026rsquo;Leary, N. A. \u003cem\u003eet al.\u003c/em\u003e Reference sequence (RefSeq) database at NCBI: current status, taxonomic expansion, and functional annotation. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e44\u003c/strong\u003e, D733\u0026ndash;D745 (2016).\u003c/li\u003e\n\u003cli\u003eXiao, Y. \u003cem\u003eet al.\u003c/em\u003e Protein Large Language Models: A Comprehensive Survey. \u003cem\u003eArXiv\u003c/em\u003e (2025).\u003c/li\u003e\n\u003cli\u003eWiegreffe, S. \u0026amp; Pinter, Y. Attention is not not Explanation. \u003cem\u003eArXiv\u003c/em\u003e (2019).\u003c/li\u003e\n\u003cli\u003eStenson, P. D. \u003cem\u003eet al.\u003c/em\u003e The Human Gene Mutation Database (HGMD\u0026reg;): optimizing its use in a clinical diagnostic or research setting. \u003cem\u003eHum Genet\u003c/em\u003e \u003cstrong\u003e139\u003c/strong\u003e, 1197\u0026ndash;1207 (2020).\u003c/li\u003e\n\u003cli\u003eGrantham, R. Amino Acid Difference Formula to Help Explain Protein Evolution. \u003cem\u003eScience (1979)\u003c/em\u003e \u003cstrong\u003e185\u003c/strong\u003e, 862\u0026ndash;864 (1974).\u003c/li\u003e\n\u003cli\u003eHopf, T. A. \u003cem\u003eet al.\u003c/em\u003e Mutation effects predicted from sequence co-variation. \u003cem\u003eNat Biotechnol\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 128\u0026ndash;135 (2017).\u003c/li\u003e\n\u003cli\u003eWon, D. G., Kim, D. W., Woo, J. \u0026amp; Lee, K. 3Cnet: pathogenicity prediction of human variants using multitask learning with evolutionary constraints. \u003cem\u003eBioinformatics\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 4626\u0026ndash;4634 (2021).\u003c/li\u003e\n\u003cli\u003eMaleki, F., Ovens, K., McQuillan, I. \u0026amp; Kusalik, A. J. Size matters: how sample size affects the reproducibility and specificity of gene set analysis. \u003cem\u003eHum Genomics\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 42 (2019).\u003c/li\u003e\n\u003cli\u003eChen, V. \u003cem\u003eet al.\u003c/em\u003e Applying interpretable machine learning in computational biology\u0026mdash;pitfalls, recommendations and opportunities for new developments. \u003cem\u003eNat Methods\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 1454\u0026ndash;1461 (2024).\u003c/li\u003e\n\u003cli\u003eLiu, Y. \u003cem\u003eet al.\u003c/em\u003e Rotamer-free protein sequence design based on deep learning and self-consistency. \u003cem\u003eNat Comput Sci\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 451\u0026ndash;462 (2022).\u003c/li\u003e\n\u003cli\u003eHenikoff, S. \u0026amp; Henikoff, J. G. Amino acid substitution matrices from protein blocks. \u003cem\u003eProc Natl Acad Sci U S A\u003c/em\u003e \u003cstrong\u003e89\u003c/strong\u003e, 10915 (1992).\u003c/li\u003e\n\u003cli\u003eGromiha, M. M. \u0026amp; Selvaraj, S. Inter-residue interactions in protein folding and stability. \u003cem\u003eProg Biophys Mol Biol\u003c/em\u003e \u003cstrong\u003e86\u003c/strong\u003e, 235\u0026ndash;277 (2004).\u003c/li\u003e\n\u003cli\u003eMcDonald, I. K. \u0026amp; Thornton, J. M. Satisfying Hydrogen Bonding Potential in Proteins. \u003cem\u003eJ Mol Biol\u003c/em\u003e \u003cstrong\u003e238\u003c/strong\u003e, 777\u0026ndash;793 (1994).\u003c/li\u003e\n\u003cli\u003eRichards, F. M. Areas, volumes, packing and protein structure. \u003cem\u003eAnnu Rev Biophys Bioeng\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, 151\u0026ndash;176 (1977).\u003c/li\u003e\n\u003cli\u003eMajithia, A. R. \u003cem\u003eet al.\u003c/em\u003e Prospective functional classification of all possible missense variants in PPARG. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e48\u003c/strong\u003e, 1570\u0026ndash;1575 (2016).\u003c/li\u003e\n\u003cli\u003eRochel, N. \u003cem\u003eet al.\u003c/em\u003e Recurrent activating mutations of PPAR\u0026gamma; associated with luminal bladder tumors. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 253 (2019).\u003c/li\u003e\n\u003cli\u003eGilpin, L. H. \u003cem\u003eet al.\u003c/em\u003e Explaining explanations: An overview of interpretability of machine learning. \u003cem\u003e2018 IEEE 5th International Conference on data science and advanced analytics (DSAA)\u003c/em\u003e \u003cstrong\u003eIEEE\u003c/strong\u003e, 80\u0026ndash;89 (2018).\u003c/li\u003e\n\u003cli\u003eBartos, D. C. \u003cem\u003eet al.\u003c/em\u003e A KCNQ1 Mutation Causes a High Penetrance for Familial Atrial Fibrillation. \u003cem\u003eJ Cardiovasc Electrophysiol\u003c/em\u003e \u003cstrong\u003e24\u003c/strong\u003e, 562 (2013).\u003c/li\u003e\n\u003cli\u003eChen, Y. H. \u003cem\u003eet al.\u003c/em\u003e KCNQ1 gain-of-function mutation in familial atrial fibrillation. \u003cem\u003eScience (1979)\u003c/em\u003e \u003cstrong\u003e299\u003c/strong\u003e, 251\u0026ndash;254 (2003).\u003c/li\u003e\n\u003cli\u003eOlesen, M. S. \u003cem\u003eet al.\u003c/em\u003e Mutations in the potassium channel subunit KCNE1 are associated with early-onset familial atrial fibrillation. \u003cem\u003eBMC Med Genet\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 24 (2012).\u003c/li\u003e\n\u003cli\u003eLiu, P. \u003cem\u003eet al.\u003c/em\u003e Reanalysis of Clinical Exome Sequencing Data. \u003cem\u003eNew England Journal of Medicine\u003c/em\u003e \u003cstrong\u003e380\u003c/strong\u003e, 2478\u0026ndash;2480 (2019).\u003c/li\u003e\n\u003cli\u003eSalfati, E. L. \u003cem\u003eet al.\u003c/em\u003e Re-analysis of whole-exome sequencing data uncovers novel diagnostic variants and improves molecular diagnostic yields for sudden death and idiopathic diseases. \u003cem\u003eGenome Med\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 1\u0026ndash;8 (2019).\u003c/li\u003e\n\u003cli\u003eMinikel, E., Painter, J., Dong, C. \u0026amp; Nelson, M. Refining the impact of genetic evidence on clinical success. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e629\u003c/strong\u003e, 624\u0026ndash;629 (2024).\u003c/li\u003e\n\u003cli\u003eTrajanoska, K. \u003cem\u003eet al.\u003c/em\u003e From target discovery to clinical drug development with human genetics. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e620\u003c/strong\u003e, 737\u0026ndash;745 (2023).\u003c/li\u003e\n\u003cli\u003eVaswani, A. \u003cem\u003eet al.\u003c/em\u003e Attention is All you Need. \u003cem\u003eAdv Neural Inf Process Syst\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, (2017).\u003c/li\u003e\n\u003cli\u003ePonomarenko, E. A. \u003cem\u003eet al.\u003c/em\u003e The size of the human proteome: the width and depth. \u003cem\u003eInt J Anal Chem\u003c/em\u003e \u003cstrong\u003e2016\u003c/strong\u003e, 7436849 (2016).\u003c/li\u003e\n\u003cli\u003eRajeevan, H., Soundararajan, U., Kidd, J. R., Pakstis, A. J. \u0026amp; Kidd, K. K. ALFRED: an allele frequency resource for research and teaching. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, D1010\u0026ndash;D1015 (2012).\u003c/li\u003e\n\u003cli\u003eGene Ontology Consortium. The gene ontology resource: 20 years and still GOing strong. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e47\u003c/strong\u003e, D330\u0026ndash;D338 (2019).\u003c/li\u003e\n\u003cli\u003eHalldorsson, B. V \u003cem\u003eet al.\u003c/em\u003e The sequences of 150,119 genomes in the UK Biobank. \u003cem\u003eNature\u003c/em\u003e \u003cstrong\u003e607\u003c/strong\u003e, 732\u0026ndash;740 (2022).\u003c/li\u003e\n\u003cli\u003eLearning sequence to predict gain- or loss-of-function variants. https://zenodo.org/records/15429133.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-6705195/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6705195/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eA clear understanding of mutational effects can advance genetics and biomedical research by providing valuable insights into gene functions, disease mechanisms, and therapeutic approaches. However, methods to determine the pathogenicity of genetic variants are limited by the absence of information on the direction of mutational effects. Here, we present ClearVariant, a deep learning system to classify pathogenic variants into gain- or loss-of-function, achieving state-of-the-art performance validated with data from ClinVar and Human Gene Mutation Database (HGMD). The model contains protein language models (PLMs) for training mutated sequences alongside their reference counterparts, showing similar predicted outcomes when a residue changed to another amino acid belonging to the same property group. We evaluated its ability to learn the protein language by observing high attention scores on coevolutionary relationships. To support advancements in biomedicine, we provide a database of pathogenic human missense variants labelled with their predicted mutational effects.\u003c/p\u003e","manuscriptTitle":"Learning sequence to predict gain- or loss-of-function variants","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-06 09:56:36","doi":"10.21203/rs.3.rs-6705195/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"cd22f301-1d95-4c28-b9bd-3de2653a57ba","owner":[],"postedDate":"June 6th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":49592307,"name":"Biological sciences/Computational biology and bioinformatics/Genome informatics"},{"id":49592308,"name":"Biological sciences/Genetics/Functional genomics/Mutagenesis"},{"id":49592309,"name":"Biological sciences/Computational biology and bioinformatics/Sequence annotation"},{"id":49592310,"name":"Biological sciences/Genetics/Clinical genetics"}],"tags":[],"updatedAt":"2026-04-29T01:11:38+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-06 09:56:36","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6705195","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6705195","identity":"rs-6705195","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}