Rēs ipSAE loquunt: What′s wrong with AlphaFold′s ipTM score and how to fix it DOI Creative Commons
Roland L. Dunbrack

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 14, 2025

AlphaFold's ipTM metric is used to predict the accuracy of structural predictions protein-protein interactions (PPIs) and probability that two proteins interact. Many AF2/AF3 users have experienced phenomenon if they trim full-length sequence constructs (e.g. from UniProt) interacting domains (or domain+peptide), their scores go up, even though structure prediction interaction unchanged. The reason this happens due mathematical formulation in AF2/AF3, which whole chains. If both chains a PPI complex contain large amounts disorder or accessory do not form primary domain-domain domain/peptide interaction, score can be lowered significantly. then does accurately represent nor whether actually We solved problem by: 1) including only residue pairs good predicted aligned error ( PAE ) scores; 2) by adjusting d 0 parameter (a function length query sequences) TM equation include number residues with interchain s residue; 3) using value itself distributions over calculate pairwise residue-residue pTM values into calculation. first are crucial calculating high for domain-peptide presence many hundreds disordered regions and/or domains. third allows us require common output json files AF2 AF3 (including server output) without having change AlphaFold code affecting accuracy. show benchmark new score, called ipSAE (interaction Score Aligned Errors), able separate true false complexes more efficiently than AlphaFold2's score. resulting program freely available at https://github.com/dunbracklab/IPSAE .

Language: Английский

Using AlphaFold Multimer to discover interkingdom protein–protein interactions DOI Creative Commons
Felix Homma, Joy Lyu, Renier A. L. van der Hoorn

et al.

The Plant Journal, Journal Year: 2024, Volume and Issue: 120(1), P. 19 - 28

Published: Aug. 17, 2024

SUMMARY Structural prediction by artificial intelligence can be powerful new instruments to discover novel protein–protein interactions, but the community still grapples with implementation, opportunities and limitations. Here, we discuss re‐analyse our in silico screen for pathogen‐secreted inhibitors of immune hydrolases illustrate power limitations structural predictions. We strategies curating sequences, including controls, reusing sequence alignments highlight important caused different platforms, depth computing times. hope these experiences will support similar interactomic screens research community.

Language: Английский

Citations

4
Assessing scoring metrics for AlphaFold2 and AlphaFold3 protein complex predictions DOI Creative Commons

Luca R. Genz,

Sanjana Nair, Maya Topf

et al.

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2025, Volume and Issue: unknown

Published: April 17, 2025

Abstract The recent breakthroughs in AI-driven protein structure prediction have revolutionized structural biology, unlocking new possibilities to model complex biomolecular interactions. We evaluated widely-used scoring metrics for assessing such models predicted by ColabFold with templates, without and AlphaFold3. benchmarked the optimal cutoffs these assessment scores using a set of 325 heterodimeric high-resolution structures their predictions. Our results show that templates AlphaFold3 perform similarly both outperform templates. Furthermore, interface-specific are found be more reliable evaluating predictions compared corresponding global scores. Notably, ipTM confidence achieve best performance distinguishing correct from incorrect Based on our results, we developed weighted combined score, C2Qscore, improve quality assessment, which was used analyze dimers large assemblies solved cryoEM. This revealed potential limitations when multiple configurations heterodimers possible. C2Qscore has been integrated as tool into ChimeraX plug-in PICKLUSTER v2.0, facilitating interactive access metrics. is freely available download Toolshed https://gitlab.com/topf-lab/pickluster-v2.0.git . study provides insights strengths weaknesses current offers guidance improving assessment. Impact this work evaluates effectiveness assess computationally generated complexes, crucial understanding cellular functions. By analyzing comparing different methods, aim enhance reliability accuracy models. findings provide valuable researchers utilizing computational approaches biological processes applications biomedical research.

Language: Английский

Citations

0
Rēs ipSAE loquunt: What′s wrong with AlphaFold′s ipTM score and how to fix it DOI Creative Commons
Roland L. Dunbrack

bioRxiv (Cold Spring Harbor Laboratory), Journal Year: 2025, Volume and Issue: unknown

Published: Feb. 14, 2025

AlphaFold's ipTM metric is used to predict the accuracy of structural predictions protein-protein interactions (PPIs) and probability that two proteins interact. Many AF2/AF3 users have experienced phenomenon if they trim full-length sequence constructs (e.g. from UniProt) interacting domains (or domain+peptide), their scores go up, even though structure prediction interaction unchanged. The reason this happens due mathematical formulation in AF2/AF3, which whole chains. If both chains a PPI complex contain large amounts disorder or accessory do not form primary domain-domain domain/peptide interaction, score can be lowered significantly. then does accurately represent nor whether actually We solved problem by: 1) including only residue pairs good predicted aligned error ( PAE ) scores; 2) by adjusting d 0 parameter (a function length query sequences) TM equation include number residues with interchain s residue; 3) using value itself distributions over calculate pairwise residue-residue pTM values into calculation. first are crucial calculating high for domain-peptide presence many hundreds disordered regions and/or domains. third allows us require common output json files AF2 AF3 (including server output) without having change AlphaFold code affecting accuracy. show benchmark new score, called ipSAE (interaction Score Aligned Errors), able separate true false complexes more efficiently than AlphaFold2's score. resulting program freely available at https://github.com/dunbracklab/IPSAE .

Language: Английский

Citations

0