Biology subjects

Stenson, P. D.

Publications and source records attributed to Stenson, P. D..

8 recordsLinked to original sources

bioRxiv2022

Genome-wide prediction of pathogenic gain- and loss-of-function variants from ensemble learning of diverse feature set

Gain-of-function (GOF) variants give rise to increased or novel protein functions whereas loss-of-function (LOF) variants lead to diminished protein function. GOF and LOF variants can result in markedly varying phenotypes, even when occurring in the same gene. However, experimental approaches for identifying GOF and LOF are generally slow and costly, whilst currently available computational methods have not been optimized to discriminate between GOF and LOF variants. We have developed LoGoFunc, an ensemble machine learning method for predicting pathogenic GOF, pathogenic LOF, and neutral genetic variants. LoGoFunc was trained on a broad range of gene-, protein-, and variant-level features describing diverse biological characteristics, as well as network features summarizing the protein-protein interactome and structural features calculated from AlphaFold2 protein models. We analyzed GOF, LOF, and neutral variants in terms of local protein structure and function, splicing disruption, and phenotypic associations, thereby revealing previously unreported relationships between various biological phenomena and variant functional outcomes. For example, GOF and LOF variants exhibit contrasting enrichments in protein structural and functional regions, whilst LOF variants are more likely to disrupt canonical splicing as indicated by splicing-related features employed by the model. Further, by performing phenome-wide association studies (PheWAS), we identified strong associations between relevant phenotypes and high-confidence predicted GOF and LOF variants. LoGoFunc outperforms other tools trained solely to predict pathogenicity or general variant impact for the identification of pathogenic GOF and LOF variants.

genetics↗

bioRxiv2022

Genome-wide detection of human variants that disrupt intronic branchpoints

Pre-mRNA splicing is initiated with the recognition of a single-nucleotide intronic branchpoint (BP) within a BP motif by spliceosome elements. Fifty-six rare variants in 44 human genes have been reported to alter splicing and cause disease by disrupting BP. However, until now, no computational approach has been available to efficiently detect such variants in next-generation sequencing (NGS) data. We established a comprehensive human genome-wide BP database by integrating existing BP data, and by generating new BP data from RNA-seq of lariat debranching enzyme DBR1-mutated patients and from machine-learning predictions. We in-depth characterize multiple features of BP in major and minor introns, and find that BP and BP-2 (two-nucleotides upstream of BP) positions exhibit a lower rate of variation in human populations and higher evolutionary conservation than the intronic background, whilst being comparable to the exonic background. We develop BPHunter as a genome-wide computational approach to systematically and efficiently detect intronic variants that may disrupt BP recognition in NGS data. BPHunter retrospectively identifies 48 of the 56 known pathogenic BP mutations in which we summarize a strategy for prioritizing BP mutation candidates, and the remaining 8 all create AG dinucleotides between BP and acceptor site which is probably the reason for mis-splicing. We demonstrate the utility of BPHunter prospectively by using it to identify a novel germline heterozygous BP variant of STAT2 in a patient with critical COVID-19 pneumonia, and a novel somatic intronic 59-nucleotide deletion of ITPKB in a lymphoma patient, both of which we validate experimentally. BPHunter is publicly available from https://hgidsoft.rockefeller.edu/BPHunter and https://github.com/casanova-lab/BPHunter.

genetics↗

bioRxiv2020

PIVOTAL: Prioritizing variants of uncertain significance with spatial genomic patterns in the 3D proteome

Variants of uncertain significance (VUS) have posed an increasingly prominent challenge to clinicians due to their growing numbers and difficulties in making clinical responses to them. Currently there are no existing methods that leverage the spatial relationship of known disease mutations and genomic properties for prioritizing variants of uncertain significance. More importantly, disease genes often associate with multiple clinically distinct diseases, but none of the existing variant prioritization methods provide clues as to the specific type of disease potentially associated with a given variant. We present PIVOTAL, a spatial neighborhood-based method using three-dimensional structural models of proteins, that significantly improves current variant prioritization tools and identifies potential disease etiology of candidate variants on a proteome scale. Using PIVOTAL, we made pathogenicity predictions for over 140,000 VUS and deployed a web application (http://pivotal.yulab.org) that enables users both to explore these data and to perform custom calculations.

genomics↗

bioRxiv2019

AMELIE 2 speeds up Mendelian diagnosis by matching patient phenotype & genotype to primary literature

The diagnosis of Mendelian disorders requires labor-intensive literature research. Trained clinicians can spend hours looking for the right publication/s supporting a single gene that best explains a patients disease. AMELIE (Automatic Mendelian Literature Evaluation) greatly accelerates this process. AMELIE parses all 29 million PubMed abstracts, downloads and further parses hundreds of thousands of full text articles in search of information supporting the causality and associated phenotypes of any published genetic variant. AMELIE then prioritizes patient candidate variants for their likelihood of explaining any patients given set of phenotypes. Diagnosis of singleton patients (without relatives exomes) is the most time-consuming scenario. AMELIE ranked the causative gene at the very top in 2/3 of 215 diagnosed singleton Mendelian patients. Evaluating only the top 11 AMELIE scored genes of 127 (median) candidate genes per patient results in rapid diagnosis for 90+% of cases. AMELIE-based evaluation of all cases is 3-19x more efficient than hand-curated database-based approaches. We replicate these results on a cohort of clinical cases from Stanford Childrens Health and the Manton Center for Orphan Disease Research. An analysis web portal with our most recent update, programmatic interface and code will be available at AMELIE.stanford.edu. A pilot run of the web portal has already served many thousands of job submissions from dozens of countries.

genetics↗

bioRxiv2019

Common homozygosity for predicted loss-of-function variants reveals both redundant and advantageous effects of dispensable human genes

Humans homozygous or hemizygous for variants predicted to cause a loss of function of the corresponding protein do not necessarily present with overt clinical phenotypes. However, the set of effectively dispensable genes in the human genome has not yet been fully characterized. We report here 190 autosomal genes with 207 predicted loss-of-function variants, for which the frequency of homozygous individuals exceeds 1% in at least one human population from five major ancestry groups. No such genes were identified on the X and Y chromosomes. Manual curation revealed that 28 variants (15%) had been misannotated as loss-of-function, mainly due to linkage disequilibrium with different compensatory variants. Of the 179 remaining variants in 166 genes (0.82% of 20,232 genes), only 11 alleles in 11 genes had previously been confirmed experimentally to be loss-of-function. The set of 166 dispensable genes was enriched in olfactory receptor genes (41 genes), but depleted of genes expressed in a wide range of organs and in leukocytes. The 125 dispensable non-olfactory receptor genes displayed a relaxation of selective constraints both between species and within humans, consistent with greater redundancy. In total, 62 of these 125 genes were found to be dispensable in at least three human populations, suggesting possible evolution toward pseudogenes. Out of the 179 common loss-of-function variants, 72 could be tested for two neutrality selection statistics, and eight displayed robust signals of positive selection. These variants included the known FUT2 mutant allele conferring resistance to intestinal viruses and an APOL3 variant involved in resistance to parasitic infections. Finally, the 41 dispensable olfactory receptor genes also displayed a strong relaxation of selective constraints similar to that observed for the 341 non-dispensable olfactory receptor genes. Overall, the identification of 166 genes for which a sizeable proportion of humans are homozygous for predicted loss-of-function alleles reveals both redundancies and advantages of such deficiencies for human survival.\n\nSignificance statementHuman genes homozygous for seemingly loss of function (LoF) variants are increasingly reported in a sizeable proportion of individuals without overt clinical phenotypes. Here, we found 166 genes with 179 predicted LoF variants for which the frequency of homozygous individuals exceeds 1% in at least one of the populations present in databases ExAC and gnomAD. This set of putatively dispensable genes showed relaxation of selective constraints suggesting that a large number of these genes are undergoing pseudogenization. Eight of the common LoF variants displayed robust signals of positive selection including two variants located in genes involved in resistance to infectious diseases. The identification of dispensable genes will allow identifying functions that are, at least nowadays, redundant, or possibly advantageous, for human survival.

genetics↗

bioRxiv2018

AVADA Enables Automated Genetic Variant Curation Directly from the Full Text Literature

PurposeThe primary literature on human genetic diseases includes descriptions of pathogenic variants that are essential for clinical diagnosis. Variant databases such as ClinVar and HGMD collect pathogenic variants by manual curation. We aimed to automatically construct a freely accessible database of pathogenic variants directly from full-text articles about genetic disease.\n\nMethodsAVADA (Automatically curated VAriant DAtabase) is a novel machine learning tool that uses natural language processing to automatically identify pathogenic variants and genes in full text of primary literature and converts them to genomic coordinates for rapid downstream use.\n\nResultsAVADA automatically curated almost 60% of pathogenic variants deposited in HGMD, a 4.4-fold improvement over the current state of the art in automated variant extraction. AVADA also contains more than 60,000 pathogenic variants that are in HGMD, but not in ClinVar. In a cohort of 245 diagnosed patients, AVADA correctly annotated 38 previously described diagnostic variants, compared to 43 using HGMD, 20 using ClinVar and only 13 (wholly subsumed by AVADA and ClinVars) using the best automated abstracts-only based approach.\n\nConclusionAVADA is the first machine learning tool that automatically curates a variants database directly from full text literature. AVADA is available upon publication at http://bejerano.stanford.edu/AVADA.

genetics↗

bioRxiv2018

S-CAP extends clinical-grade pathogenicity prediction to genetic variants that affect RNA splicing

There are over 15,000 known variants that cause human inherited disease by disrupting RNA splicing. While several in silico methods such as CADD, EIGEN and LINSIGHT are commonly used to predict the pathogenicity of noncoding variants, we introduce S-CAP, a tool developed specially for splicing which is better able to effectively distinguish pathogenic splicing-relevant variants from benign variants. S-CAP is a novel splicing pathogenicity predictor that reduces the number of splicing-relevant variants of uncertain significance in patient exomes by 41%, a nearly 3-fold improvement over existing noncoding pathogenicity measures while correctly classifying known pathogenic splicing-relevant variants with a clinical-grade 95% sensitivity.

genomics↗

bioRxiv2017

AMELIE accelerates Mendelian patient diagnosis directly from the primary literature

The diagnosis of Mendelian disorders requires labor-intensive literature research. Our software system AMELIE (Automatic Mendelian Literature Evaluation) greatly automates this process. AMELIE parses hundreds of thousands of full text articles to find an underlying diagnosis to explain a patients phenotypes given the patients exome. AMELIE prioritizes patient candidate genes for their likelihood of causing the patients phenotypes. Diagnosis of singleton patients (without relatives exomes) is the most time-consuming scenario. AMELIEs gene ranking method was tested on 215 singleton Mendelian patients with a clinical diagnosis. AMELIE ranked the causal gene among the top 2 in the majority (63%) of cases. Examining AMELIEs top 10 genes, amounting to 8% of 124 candidate genes with rare functional variants per patient, results in diagnosis for 95% of cases. Strikingly, training only on gene pathogenicity knowledge from 2011 leads to identical performance compared to training on current data. An accompanying analysis web portal has launched at AMELIE.stanford.edu.

genetics↗