Genome-wide prediction of disease variant effects with a deep protein language model.

Brandes N, Goldman G, Wang CH, Ye CJ, Ntranos V. · 2023

Read Genome-wide prediction of disease variant effects with a deep protein language model. by Brandes N, Goldman G, Wang CH, Ye CJ, Ntranos V. (2023) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/europmc-pmc10484790.txt.

Introduction

Brandes and co-authors used a deep AI model in 2023 to predict the effect of every gene variant in humans. A missense variant swaps one protein building block for another. Such swaps can be harmless or can cause disease. Past tools could score only some swaps, due to gaps in kin data or software caps.

Their model, ESM1b, learned from protein sequences alone. It scored all roughly 450 million possible missense swaps. Brandes found it beat past tools on 150,000 known disease-linked swaps. It also matched lab measures across 28 large test sets. About 2 million swaps harmed just one form of a protein, not its kin forms.

Catalog record

CREATORBrandes N, Goldman G, Wang CH, Ye CJ, Ntranos V.
YEAR2023
COLLECTIONPAPERS
FORMATTXT
FILE SIZE75 KB
LICENSECC BY 4.0
RIGHTSCC BY
SOURCEEurope PMC

Source: https://europepmc.org/articles/PMC10484790

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Genome-wide prediction of disease variant effects with a deep protein language model By Nadav Brandes, Grant Goldman, Charlotte H Wang, Chun Jimmie Ye, Vasilis Ntranos Published in: Nature Genetics DOI: 10.1038/s41588-023-01465-0 ABSTRACT Abstract Predicting the effects of coding variants is a major challenge. While recent deep-learning models have improved variant effect prediction accuracy, they cannot analyze all coding variants due to dependency on close homologs or software limitations. Here we developed a workflow using ESM1b, a 650-million-parameter protein language model, to predict all ~450 million possible missense variant effects in the human genome, and made all predictions available on a web portal. ESM1b outperformed existing methods in classifying ~150,000 ClinVar/HGMD missense variants as pathogenic or benign and predicting measurements across 28 deep mutational scan datasets. We further annotated ~2 million variants as damaging only in specific protein isoforms, demonstrating the importance of considering all isoforms when predicting variant effects. Our approach also generalizes to more complex coding variants such as in-frame indels and stop-gains. Together, these results establish protein language models as an effective, accurate and general approach to predicting variant effects. Subject terms: Functional genomics, Bioinformatics MAIN Determining the phenotypic consequences of genetic variants, known as variant effect prediction (VEP), is a key challenge in human genetics1–4. Coding variants altering the amino acid sequences of proteins are of special interest due to their enrichment in disease associations, better-understood mechanisms and therapeutic actionability5–8. Most naturally occurring coding variants are missense, substituting one amino acid with another9. Despite progress in functional genomics and genetic studies, distinguishing protein-disrupting damaging variants from neutral ones remains a challenge. Furthermore, most human genes are alternatively spliced, and the same variant may be damaging to some protein isoforms but neutral to others, depending on interactions with the rest of the protein. Thus, most missense variants remain as variants of uncertain significance (VUS), limiting the utility of exome sequencing in clinical diagnosis2,10. VEP is even more challenging for coding variants affecting multiple residues such as in-frame indels. Experimental approaches for VEP such as deep mutational scans (DMS)11 and Perturb-seq12 can measure molecular and cellular phenotypes across thousands of variants simultaneously. However, these endophenotypes are imperfect proxies for the relevant clinical phenotypes and remain difficult to scale genome-wide13,14. In contrast, computational methods that learn the biophysical properties or evolutionary constraints of proteins could theoretically cover all coding variants15–17. While most computational methods are trained on labeled data of pathogenic versus benign variants10, unsupervised homology-based methods predict variant effects directly from multiple sequence alignments (MSA) without training on labeled data. EVE, an unsupervised deep-learning method implementing a generative variational autoencoder, was recently shown to outperform supervised methods4. However, due to their reliance on MSA, homology-based methods provide predictions only for a subset of well-aligned proteins and residues. Moreover, because alternative isoforms of the same gene have identical homologs, it is unclear whether they can distinguish the effects of variants on different isoforms. Another deep-learning approach to VEP uses protein language models, a technique derived from natural language processing. These are deep neural networks trained to model the space of known protein sequences selected throughout evolution as captured by large protein datasets such as UniProt18 (Fig. 1a). Notably, protein language models do not require explicit homology and can estimate the likelihood of any possible amino acid sequence. They have been shown to implicitly learn how protein sequence determines many aspects of protein structure and function, including secondary structure, long-distance interactions, post-translational modifications and binding sites19–24. One of the largest protein language models is ESM1b, a publicly available 650-million-parameter model trained on ~250 million protein sequences20. It was demonstrated to predict, without further training, variant effects correlated with DMS experiment results25. [Figure: ESM1b predicts variant effects without homology coverage.a, ESM1b is a 650-million-parameter protein language model trained on 250 million protein sequences across all organisms. The model was trained via the masked language modeling task, where random residues are masked from input sequences and the model has to predict the correct amino acid at each position (including the missing residues). b, Illustration of the ESM1b model’s input (an amino acid sequence) and output (LLR of effect scores for all possible missense variants). c, The distribution of MSA coverage (that is, the fraction of a protein’s residues that are aligned) across ~3,000 disease-related proteins covered by EVE. d, Examples of the model’s capacity to detect protein domains and functional regions, including outside MSA coverage, across the following three human proteins: SPAST, SLC7A3 and ARX. Each heatmap visualizes the LLR scores across all 20 × L possible missense variants (where L is the protein length). Protein domains without MSA coverage are highlighted in orange.] However, several limitations have restricted the use of ESM1b for VEP. First, the model’s input sequence length is limited to 1,022 amino acids, excluding ~12% of human protein isoforms. Second, while evaluated on DMS data across 32 genes (10 from humans)25, it has remained unknown how the model performs at predicting the clinical impact of coding variants genome-wide. Finally, using ESM1b requires software engineering proficiency, deep-learning expertise and high-memory GPUs, which together create a technical barrier for widespread use. Here we implemented a workflow generalizing ESM1b to protein sequences of any length and used it to predict all ~450 million possible missense variant effects across all 42,336 protein isoforms in the human genome. We evaluated our workflow on three different benchmarks and compared it to 45 other VEP methods. Our workflow outperforms all compared methods in classifying variant pathogenicity (as annotated by ClinVar10 and HGMD26) and predicting DMS experiments. We further demonstrate the capacity of ESM1b to assess variant effects in the context of different protein isoforms, identifying isoform-sensitive variants in 85% of alternatively spliced genes. Finally, we present a scoring algorithm that generalizes ESM1b to variants affecting multiple residues and demonstrates the model’s accurate predictions over in-frame indels and stop-gain variants. We created a web portal allowing users to query, visualize and download missense VEPs for all human protein isoforms (accessible at https://huggingface.co/spaces/ntranoslab/esm_variants). RESULTS Predicting the effects of all possible missense variants in the human genome We developed a modified ESM1b workflow and applied it to obtain a complete catalog of all ~450 million missense variant effects on all 42,336 known human protein isoforms. Each variant’s effect score is the log-likelihood ratio (LLR) between the variant and wild-type (WT) residue (Fig. 1b). Unlike homology-based models currently available only for a subset of human proteins and residues with MSA coverage (for example, 84% of the residues in ~3,000 disease genes covered by EVE; Fig. 1c), ESM1b predicts the effects of every possible missense variant. Protein regions with many possible mutations predicted by ESM1b as damaging often align with known protein domains (Fig. 1d). As illustrated for SPAST, SLC7A3 and ARX, these domains may reside outside MSA coverage and be unsuitable for homology-based models (Fig. 1d), yet harbor disease-associated variants. For example, the microtubule-interacting and trafficking (MIT) domain in SPAST contains missense variants implicated in hereditary spastic paraplegias27, the CAT C domain in SLC7A3 contains an autism-linked variant (S589T)28 and multiple domains in ARX outside MSA coverage (highlighted in Fig. 1d) contain missense variants linked to intellectual disability29–32. ESM1b outperforms other VEP methods over clinical and experimental benchmarks To assess

Read the complete text