henhao
Sign in
BioinformaticsbioRxiv
Preprint

Enhanced Protein-Protein Interaction Discovery via AlphaFold-Multimer

Ah‐Ram KimiD, Yanhui HuiD, Aram ComjeaniD, Jonathan RodigeriD, Stephanie Elizabeth MohriD, Norbert PerrimoniD

Abstract

Accurately mapping protein-protein interactions (PPIs) is critical for elucidating cellular functions and has significant implications for health and disease. Conventional experimental approaches, while foundational, often fall short in capturing direct, dynamic interactions, especially those with transient or small interfaces. Our study leverages AlphaFold-Multimer (AFM) to re-evaluate high-confidence PPI datasets from Drosophila and human. Our analysis uncovers a significant limitation of the AFM-derived interface pTM (ipTM) metric, which, while reflective of structural integrity, can miss physiologically relevant interactions at small interfaces or within flexible regions. To bridge this gap, we introduce the Local Interaction Score (LIS), derived from AFMs Predicted Aligned Error (PAE), focusing on areas with low PAE values, indicative of the high confidence in interaction predictions. The LIS method demonstrates enhanced sensitivity in detecting PPIs, particularly among those that involve flexible and small interfaces. By applying LIS to large-scale Drosophila datasets, we enhance the detection of direct interactions. Moreover, we present FlyPredictome, an online platform that integrates our AFM-based predictions with additional information such as gene expression correlations and subcellular localization predictions. This study not only improves upon AFMs utility in PPI prediction but also highlights the potential of computational methods to complement and enhance experimental approaches in the identification of PPI networks.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
76
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (21 Feb 2024) · CC BY 4.0

A Local Interaction Score (LIS), computed from AlphaFold-Multimer predicted aligned error, separates curated fly and human protein interactions from controls better than ipTM, recovering small or flexible interfaces.

The interface pTM (ipTM) score from AlphaFold-Multimer tracks the structural accuracy of a predicted complex, so pairs that bind through small interfaces, linear motifs or disordered regions can score low despite experimental support. Unlike pDockQ and pDockQ2, which rely on residue proximity, LIS uses only the confidence in interface regions with low predicted aligned error (PAE), and it is applied here to screen more than 80,000 fly protein pairs.

Key results

  1. ShownLIS, an average of inverted PAE over low-error interface regions, separates positive reference PPIs from controls more strongly than ipTM. It also flags many positive pairs with low ipTM but high LIS.

    Fly and human PRS vs RRS, PRS-GFP and PRS-Wg controls, t-test p = 4.35e-201 (average LIS) and 9.40e-164 (best LIS), vs 1.38e-175 (average ipTM) and 8.60e-122 (best ipTM). A PAE cutoff of 12 gave the highest ROC AUC across cutoffs 1 to 30 (values reported as 0.911 for average and 0.891 for best LIS). · Limitations of assessing PPIs using…; Enhancing PPI detection using Local Interaction…; Calculation of Local Interaction Score and Local…; Figure legends

  2. ShownCompared with AFM model confidence (0.2*ipTM + 0.8*pTM), pDockQ, pDockQ2, ipTM and LIA, LIS ranked first in ROC analysis for both average and best scores. It was also the metric least correlated with pLDDT and pTM.

    AUC ranking for average scores: LIS first, then ipTM, then LIA. For best scores: LIS first, then LIA. LIS had the lowest correlation with pLDDT and pTM of all metrics and the most significant positive-vs-control p-values. The main text does not give numeric AUCs for this comparison. · Comparison of the LIS with existing AFM-derived…; Figure legends

  3. ShownThresholds set on fly and human data carry over to yeast and to short-linear-motif complexes from ELM, where many true pairs have low ipTM but high LIS.

    Over 70% of yeast PRS pairs and ELM-annotated PPIs passed the average and best LIS thresholds, vs about 5% of yeast negative controls (RRS, PRS-GFP, PRS-Wg), similar to fly and human negatives. The examples shown have best LIS above the 0.21 threshold and best ipTM below the 0.38 threshold. · Predicting direct interactions among yeast PPIs…; Discussion; Figure legends

  4. SuggestedApplied to large Drosophila collections, LIS calls a minority of reported pairs direct. The rate is lowest for IP-MS and highest for multiply validated literature PPIs, and each collection includes over 100 positives with low ipTM (over 200 in FlyBi, over 300 in the literature set).

    Average/best LIS thresholds were met by about 25%/19% of FlyBi Y2H PPIs (a high-confidence set of 8,723 interactions), about 16%/13% of DPiM IP-MS PPIs and 28%/24% of over 8,000 literature-curated PPIs. Positives with low ipTM: over 200, over 100 and over 300 respectively. Over 5,000 direct PPIs were predicted from more than 30,000 evaluated pairs. · Introduction; Computational analysis of a large-scale…

  5. SuggestedThree rounds of iterative AFM screening on a Drosophila m6A methyltransferase complex (MTC) IP-MS dataset rebuilt known contacts such as METTL3-METTL14 and proposed Kap-alpha3 and 14-3-3 epsilon as network hubs.

    Pairs passing LIS and LIA thresholds (average or best) were called direct. CCT7 was first predicted to bind METTL3; later rounds added CCT3, CCT4, CCT8, CCT2 and CCT6, giving binary interactions for six of eight CCT subunits. Three baits were predicted to homodimerize. · Expanding the network from an IP-MS dataset via…; Sequential AFM analysis; Figure legends

Methods, data and reproducibility

Protein pairs were folded with AlphaFold-Multimer 2.3.1 in LocalColabFold (5 models, 5 recycles). LIS averages PAE values inverted to a 0 to 1 scale over interface residue pairs within a PAE cutoff of 12, with thresholds set by Youden's index on fly/human reference vs control sets; LIS was then applied to yeast, ELM, IP-MS, Y2H and literature fly PPIs, and over 80,000 fly predictions populate FlyPredictome.

  • Positive and random reference sets (PRS/RRS) of PPIs, plus PRS proteins paired with GFP or Drosophila Wingless as extra negative controls · Drosophila, human, S. cerevisiae
  • ELM protein-linear peptide complexes, using full-length UniProt sequences for PDB entries mapping to exactly two UniProt IDs
  • m6A methyltransferase complex (MTC) IP-MS interactome from the authors' earlier work · Drosophila
  • FlyBi high-confidence interactions from all-by-all Y2H screens · Drosophila · n = 8,723 binary interactions spanning 2,939 proteins
  • DPiM IP-MS interactions from baits expressed in S2R+ cells · Drosophila · n = over 10,000 PPIs from roughly 5,000 baits
  • Literature-curated PPIs supported by multiple methods or groups (publications up to 2021) · Drosophila · n = >8,000 PPIs
  • Candidate kinase-substrate pairs from MIST and NetPhorest, with substrates trimmed to 101-residue windows around phosphosites · Drosophila · n = 49 kinases
  • Proteins in major signaling pathways (FlyPhoneDB) and KEGG metabolic pathways · Drosophila
Methods & tools
LocalColabFold 1.5.2, AlphaFold-Multimer 2.3.1, MMseqs2, pDockQ / pDockQ2, scikit-learn (sklearn), ChimeraX, Cytoscape, DeepLoc 2.0, MIST, NetPhorest, UniProt, 3Dmol.js
Code
https://github.com/flyark/AFM-LIS
Data & accessions
https://www.flyrnai.org/tools/fly_predictome/http://elm.eu.org/pdbs/
Software & environment
LocalColabFold 1.5.2 with AlphaFold-Multimer 2.3.1 and MMseqs2 alignments; colabfold_batch with --num-recycle 5 --num-models 5 --model-type auto --rank multimer. Python analysis used sklearn, pandas, NumPy, Matplotlib and networkx (no versions stated).
Assay conditions & compute
Mixed Nvidia GPUs (Tesla M40, Tesla V100, RTX6000, RTX8000, A40, A100, L40) on the Harvard O2 cluster, allocated by complex size. Results were consistent across GPUs except for large complexes on low-memory cards. Runtimes are not reported.
Reproducibility gaps
Only the LIS/LIA Jupyter notebook is linked. Other analysis scripts (partly drafted with ChatGPT, per S22) and Python package versions are not provided. Thresholds other than best LIS 0.21 and best ipTM 0.38, including the LIA cutoffs, appear only in supplementary figures.
Limitations (6)
  • The optimal LIS/LIA thresholds come from fly and human reference sets that may favour stable interactions. Pairs whose interaction area does not match their LIS can be misclassified, and the authors call for more tailored thresholds. · stated by the authors
  • Some pairs that pass the LIS and LIA thresholds conflict with predicted subcellular localization, e.g. FlyBi Y2H pairs between extracellular and nuclear proteins. Without biological context, false positives remain. · stated by the authors
  • Only about 25% (average) and 19% (best) of high-confidence FlyBi pairs pass the LIS thresholds. The authors cannot tell how much of this reflects errors in the dataset versus limited LIS sensitivity. · stated by the authors
  • Covering every pair in the proteome would need more than a hundred million predictions, which is beyond the authors' compute. FlyPredictome therefore covers prioritized pathways, kinases and existing datasets. · stated by the authors
  • For fly and human, the PAE cutoff, thresholds, AUCs and p-values all come from the same reference sets. The negatives are random pairs or GFP/Wingless pairings, and only the yeast and ELM sets were not used for tuning. · noted in this summary
  • None of the newly predicted direct interactions, such as the Kap-alpha3 and 14-3-3 epsilon hubs in the MTC network, are tested experimentally in this work. · noted in this summary

Who should read:Read this if you screen protein pairs with AlphaFold-Multimer and want a PAE-based score that catches motif- or disorder-mediated binding that ipTM misses, or you want precomputed Drosophila PPI predictions.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISHu '24Peng '26Goodwin '25Connesson '25Bellinzona '24Kim '26
71% semantic matchBioinformatics
Wenxing Hu, Masahito Ohue·bioRxiv·Oct 25, 2024· 3 citations