In silico saturation mutagenesis of cancer genes

Muiños, Ferran; Martínez-Jiménez, Francisco; Pich, Oriol; González-Pérez, Abel; López‐Bigas, Núria

doi:10.1038/s41586-021-03771-1

Cited by 87 publications

(92 citation statements)

References 62 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…For mutations predicted to be oncogenic in these genes[15], we already had experimental structures for 65% of them, and AlphaFold only added 3%. We observed the same trend for another recent set of predicted oncogenic mutations by in silico mutagenesis of cancer driver genes[39], with experimental structural data from PDB covering 52% of all oncogenic mutations and AlphaFold only adding 4% ( Figure 3d ). It should be noted, though, that the algorithms predicting the oncogenicity of both sets of somatic mutations use in part structural information, so the results are likely biased towards regions with pre-existing structural data.…”

Section: Resultssupporting

confidence: 81%

See 1 more Smart Citation

The structural coverage of the human proteome before and after AlphaFold

Porta-Pardo

Ruiz-Serra

Valencia

2021

Preprint

View full text Add to dashboard Cite

The protein structure field is experiencing a revolution. From the increased throughput of techniques to determine experimental structures, to developments such as cryo-EM that allow us to find the structures of large protein complexes or, more recently, the development of artificial intelligence tools, such as AlphaFold, that can predict with high accuracy the folding of proteins for which the availability of homology templates is limited. Here we quantify the effect of the recently released AlphaFold database of protein structural models in our knowledge on human proteins. Our results indicate that our current baseline for structural coverage of 47%, considering experimentally-derived or template-based homology models, elevates up to 75% when including AlphaFold predictions, reducing the fraction of dark proteome from 22% to just 7% and the number of proteins without structural information from 4.832 to just 29. Furthermore, although the coverage of disease-associated genes and mutations was near complete before AlphaFold release (70% of ClinVar pathogenic mutations and 74% of oncogenic mutations), AlphaFold models still provide an additional coverage of 2% to 14% of these critically important sets of biomedical genes and mutations. We also provide several examples of disease-associated proteins where AlphaFold provides critical new insights. Overall, our results show that the sequence-structure gap of human proteins has almost disappeared, an outstanding success of direct consequences for the knowledge on the human genome and the derived medical applications.

show abstract

Section: Resultssupporting

confidence: 81%

“…The list of driver genes used in Figure 3d is from OncoKB. Mutations from BoostDM[39] were obtained from the IntoGen website (www.intogen.org).…”

Section: Methodsmentioning

confidence: 99%

The structural coverage of the human proteome before and after AlphaFold

Porta-Pardo

Ruiz-Serra

Valencia

2021

Preprint

View full text Add to dashboard Cite

show abstract

“…Gradient boosting classifier is chosen both because of its capability for feature interpretation and its performance in previously reported variant impact prediction tasks (e.g. boostDM [26]). We frame this as a binary classification task by converting the variant impact scores obtained in individual DMS experiments into binary (damaging/not damaging) labels, and annotating variants with four features: (1) wild-type and mutant amino acids; (2) trinucleotide motif surrounding the DNA substitution (hereafter referred to as DNA mutational signature); (3) conservation of each position of the trinucleotide motif (using PhyloP [34]); (4) solvent accessibility (the quotient solvent accesisble surface area, or Q(SASA), calculated using POPS [35]; see Methods for detailed description of these metrics).…”

Section: Resultsmentioning

confidence: 99%

Section: Dms Data Can Be Utilised For Probing the Mutational "Dark Ma...mentioning

confidence: 99%

“…One way to address this issue is to perform “Deep Mutational Scanning” (DMS) [17, 18] where every position in the protein is mutated to any other 19 amino acids either in vitro or in silico , followed by assays to measure the stability and/or activity of the mutants. Experimental DMS have been applied on several cancer-related proteins and domains [19, 20, 21, 22]; computationally, variant impact predictors which incorporate sequence conservation (large protein/domain multiple sequence alignments), structural/physicochemical features (protein secondary/tertiary structures) and consequential representations of protein motions (imposing network models onto three-dimensional structures to account for molecular vibrations) have been developed, applied and evaluated in a DMS context [23, 24, 25, 26]. These computational and experimental data allows us to probe and access features of variants which reside in this mutational “dark matter”.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

The “dark matter” of protein variants carries a distinct DNA signature and predicts damaging variant effects

Fraternali

2021

Preprint

View full text Add to dashboard Cite

Signatures of DNA motifs associated with distinct mutagenic exposures have been defined for somatic variants, but little is known about the consequences different mutational processes pose to the cancer cell, particularly the distribution of the resulting variants in the implied proteins and their structural regions (surface, core, interacting interface). Here we first compare the protein-level consequences of six mutational signatures (Aging, APOBEC, POLE, UV, 5-FU and Platinum) characterised by clear DNA motif preferences. By mapping individual substitution events observed in tumours to three-dimensional protein structures, we show that these common somatic mutational signatures are biased against the protein core, consistent with the lower tolerability of substitutions at such structurally important regions. On the other hand, deep mutational scanning (DMS) data allow us to probe the "dark matter" of somatic mutational landscape, exploring variants which are otherwise removed in purifying selection. A computational DMS analysis identifies mutational contexts (5'-G/C[T>G]A/G-3') which are associated with damaging mutations, by altering physicochemical characteristics of amino acids at the protein core. We argue that comprehensive DMS analysis can contribute to classification of variants according to their true impact to the stability/activity of the affected protein, decoupling this from pathogenicity prediction offered by conventional variant impact classifiers.

show abstract

A Learning Program for Treatment Recommendations by Molecular Tumor Boards and Artificial Intelligence

Sunami,

Naito,

Saigusa

et al. 2024

JAMA Oncol

View full text Add to dashboard Cite

ImportanceSubstantial heterogeneity exists in treatment recommendations across molecular tumor boards (MTBs), especially for biomarkers with low evidence levels; therefore, the learning program is essential.ObjectiveTo determine whether a learning program sharing treatment recommendations for biomarkers with low evidence levels contributes to the standardization of MTBs and to investigate the efficacy of an artificial intelligence (AI)–based annotation system.Design, Setting, and ParticipantsThis prospective quality improvement study used 50 simulated cases to assess concordance of treatment recommendations between a central committee and participants. Forty-seven participants applied from April 7 to May 13, 2021. Fifty simulated cases were randomly divided into prelearning and postlearning evaluation groups to assess similar concordance based on previous investigations. Participants included MTBs at hub hospitals, treating physicians at core hospitals, and AI systems. Each participant made treatment recommendations for each prelearning case from registration to June 30, 2021; participated in the learning program on July 18, 2021; and made treatment recommendations for each postlearning case from August 3 to September 30, 2021. Data were analyzed from September 2 to December 10, 2021.ExposuresThe learning program shared the methodology of making appropriate treatment recommendations, especially for biomarkers with low evidence levels.Main Outcomes and MeasuresThe primary end point was the proportion of MTBs that met prespecified accreditation criteria for postlearning evaluations (approximately 90% concordance with high evidence levels and approximately 40% with low evidence levels). Key secondary end points were chronological enhancements in the concordance of treatment recommendations on postlearning evaluations from prelearning evaluations. Concordance of treatment recommendations by an AI system was an exploratory end point.ResultsOf the 47 participants who applied, 42 were eligible. The accreditation rate of the MTBs was 55.6% (95% CI, 35.3%-74.5%; P &lt; .001). Concordance in MTBs increased from 58.7% (95% CI, 52.8%-64.4%) to 67.9% (95% CI, 61.0%-74.1%) (odds ratio, 1.40 [95% CI, 1.06-1.86]; P = .02). In postlearning evaluations, the concordance of treatment recommendations by the AI system was significantly higher than that of MTBs (88.0% [95% CI, 68.7%-96.1%]; P = .03).Conclusions and RelevanceThe findings of this quality improvement study suggest that use of a learning program improved the concordance of treatment recommendations provided by MTBs to central ones. Treatment recommendations made by an AI system showed higher concordance than that for MTBs, indicating the potential clinical utility of the AI system.

show abstract

In silico saturation mutagenesis of cancer genes

Cited by 87 publications

References 62 publications

The structural coverage of the human proteome before and after AlphaFold

The structural coverage of the human proteome before and after AlphaFold

The “dark matter” of protein variants carries a distinct DNA signature and predicts damaging variant effects

A Learning Program for Treatment Recommendations by Molecular Tumor Boards and Artificial Intelligence

Contact Info

Product

Resources

About