henhaobio
Sign in

bioRxivBioinformaticsPreprint

Large-scale structure prediction of DUF-containing protein-protein interactions

Lino RiepenhauseniD, Francesco CostaiD, Antonina E. AndreevaiD, Alex G. BatemaniD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Motivation: Continuing advances in genome and metagenome sequencing expand the number of identified conserved protein families that remain functionally uncharacterized and contain domains of unknown function (DUFs). Functional-association resources such as STRING provide biological context, but mostly do not distinguish indirect association from physical interaction. We assessed whether AlphaFold 3 complex prediction, combined with STRING evidence and domain-level analysis of interfaces and interaction partners, can help identify and characterize DUF-containing proteins. Results: We generated four structural-prediction cohorts from STRING associations involving DUF-containing proteins and evaluated the predicted complexes using interface ipSAE, average pLDDT and buried surface area. An L2-regularized logistic regression model was trained on an initial cohort of predictions from high-confidence STRING associations to prioritize DUF-containing candidates likely to produce structurally confident AlphaFold 3 complexes. The model was then applied across all 12,535 organisms represented in STRING v12.0, followed by grouping into DUF-family and partner-architecture modules, covering 2,076 unique DUF families. The final L2-model screen contained 12,298 successfully modelled protein pairs, including 1,208 (9.82%) complexes meeting a strict-confidence criterion and 2,433 (19.78%) meeting a more liberal confidence criterion. Two examples suggest roles for DUF4130 in nucleic-acid-associated radical-SAM biology and DUF5819 in a bacterial system related to vitamin-K-dependent carboxylation. Availability and implementation: Predicted structures and associated metadata are available through Zenodo at https://doi.org/10.5281/zenodo.21875362. The model implementation and code used to generate the analyses and figures are available at https://github.com/linoriep/Proteome-scale-structure-prediction-of-DUF-containing-protein-protein-interactions.

In-depth summary

Claude Opus 5.5 · xhigh effort · 8 Nov 2026 · from the full text of version 1 (20 Aug 2026) · CC BY 4.0

A logistic regression on STRING evidence picked DUF (domain of unknown function) protein pairs for AlphaFold 3; of 12,298 modelled pairs, 1,208 met strict and 2,433 liberal confidence.

STRING links many DUF-containing proteins to partners but mostly cannot tell physical contact from indirect association, and the AlphaFold Protein Structure Database (AFDB) heterodimers cover only 46 selected proteomes and few DUFs. This work runs AlphaFold 3 on STRING-derived pairs from all 12,535 STRING v12.0 organisms, adds domain-level interface analysis, and finds confident complexes for 1,027 DUF families versus 321 in AFDB (liberal confidence).

Key results

  1. ShownThe model-selected screen across all STRING v12.0 organisms gave higher AlphaFold 3 confidence yields than the three rule-based cohorts and the AFDB DUF heterodimer set, and yield rose further at a stricter model-score cutoff.

    Liberal (ipSAE ≥0.2) / strict (ipSAE ≥0.6), both average pLDDT ≥70: L2 screen 19.78%/9.82% (2,433 and 1,208 of 12,298 modelled pairs); model score ≥0.75 subset (4,323 structures) 27.90%/15.22%; STRING ≥900, ≥10 taxa cohort 13.58%/7.19%; STRING ≥700, ≥50 taxa 9.58%/4.83%; fusion cohort 5.62%/2.76%; AFDB DUF heterodimers 1.41%/0.37% (2,992 and 791 of 212,084). · AlphaFold 3 Predictions and Analysis; 3 Results

  2. ShownConfident complexes from the L2 screen cover many more DUF families than AFDB does, and in most of them the DUF domain itself lies at the predicted interface.

    Of 2,076 DUF families screened, 1,027 had a liberal-confidence and 681 a strict-confidence complex, versus 321 and 80 in AFDB. By buried surface area (interface ≥300 Ų, DUF share ≥15%), the DUF was at the interface in 1,391 of 2,433 liberal (57.17%) and 796 of 1,208 strict (65.89%) complexes; 649 and 348 of these had a functionally annotated non-DUF partner. · AlphaFold 3 Predictions and Analysis; 3 Results; 4 Discussion

  3. ShownAn L2-regularized logistic regression on STRING evidence channels ranked pairs by likely AlphaFold 3 confidence better than the combined STRING score, co-occurrence or neighborhood alone.

    Five-fold stratified cross-validation on 4,523 bacterial modules, with liberal confidence as the positive label: average precision (AP) 0.372 and AUROC 0.800, versus 0.316/0.781 for co-occurrence, 0.244/0.737 for neighborhood and 0.142/0.524 for the combined STRING score. XGBoost improved slightly but was not adopted because it is harder to interpret. · Cohorts; 3 Results; 4 Discussion

  4. SuggestedFor DUF families that Pfam later annotated, the partners predicted here often matched the new function or the biological system it belongs to.

    Of 38 Pfam 38.0 DUF families that no longer count as DUFs in Pfam 38.2, 15 showed specific agreement, 15 broader-process agreement and 8 no clear relationship. Examples: XusB (formerly DUF4374) with TonB-dependent iron receptors; GH116 catalytic region (formerly DUF608) with beta-glucosidase partners. · 3 Results; 4 Discussion

  5. SuggestedTwo strict-confidence complexes give testable hypotheses: DUF4130 as a nucleic-acid-binding module of a radical S-adenosyl methionine (SAM) enzyme, and DUF5819 as part of a bacterial system resembling vitamin K-dependent gamma-glutamyl carboxylase (GGCX).

    DUF4130 pair (Lachnospiraceae bacterium 3-1): ipSAE 0.804, pLDDT 86.3, 3,252 of 3,666 Ų interface from DUF4130; cofolded SAM sits by a Cys70-X3-Cys74-X2-Cys77 motif; DALI Z 8.8 to 6.0 against a DHH-DHHA1 arrangement. DUF5819 pair (Melghirimyces thermohalophilus): ipSAE 0.806, pLDDT 85.2; resembles human GGCX (PDB 9L54); cofolded menaquinone-7 hydroquinone lies near Lys189, Trp195 and Phe273. · 3 Results; 4 Discussion

Methods, data and reproducibility

STRING v12.0 pairs involving Pfam DUF proteins were grouped into modules (DUF family plus partner Pfam architecture; fusion cohort: both proteins' full architectures); one pair per module was modelled with AlphaFold 3 in four cohorts, scored by ipSAE, average pLDDT and DUF buried surface area. An L2-regularized logistic regression on STRING channels, trained on cohort 1, selected the final all-organism cohort.

  • STRING v12.0 full scored-link database (local copy), mapped to UniProt via alias files · multiple species · n = 12,535 organisms
  • Pfam 38.0 family assignments (DUF = name or description matching DUF[0-9]+); Pfam 38.2 used for the retrospective check · n = 25,545 families, 5,951 DUFs
  • Cohort 1: bacterial STRING pairs with combined score ≥900, module present in ≥10 bacterial taxa (also the L2 training set) · bacteria · n = 4,523 modules (4,520 modelled)
  • Cohort 2: bacterial STRING pairs with combined score ≥700, module present in ≥50 taxa · bacteria · n = 4,017 modules
  • Cohort 3: bacterial pairs with STRING fusion score >1, grouped by full Pfam architectures, present in ≥5 organisms · bacteria · n = 2,135 modules
  • Cohort 4: L2-model-selected pairs (STRING combined score ≥900, model score ≥0.5) from all STRING organisms; 10,422 bacterial, 1,680 eukaryotic, 272 archaeal pairs · bacteria, eukaryotes, archaea · n = 12,374 modules (12,298 modelled)
  • AFDB heterodimer predictions containing at least one DUF protein, used as reference · multiple species (46 selected proteomes) · n = 212,084 complexes
  • Experimental structures for comparison: human GGCX with protein S, SAM-bound MiaB, DNA-bound RecJ · PDB 9L54, 7MJV, 5F55
Methods & tools
STRING v12.0, Pfam 38.0, AlphaFold 3, AlphaJudge v1.0.2, Biopython (Shrake-Rupley), UniProt / Swiss-Prot, Gene Ontology, AlphaFold Protein Structure Database, Protein Data Bank, DALI, XGBoost, L2-regularized logistic regression
Software & environment
Local AlphaFold 3 (1 diffusion sample, 10 recycles, seed=1, other settings as in the official installation); AlphaJudge v1.0.2; Biopython Shrake-Rupley solvent-accessible area; local STRING v12.0 and Pfam 38.0. No container or dependency files described.
Reproducibility gaps
Hardware, GPU type and runtime are not reported. Each pair was run with a single seed and diffusion sample. Model features, coefficients and the classifier benchmark are in supplementary Tables S1 and S2, which are not in the main text.
Limitations (6)
  • The L2 model is not a general protein-protein interaction classifier: its features are correlated STRING channels, its labels are AlphaFold 3 outcomes rather than validated interactions, and predicting another model's output limits its performance. · stated by the authors
  • Each pair was modelled with one diffusion sample and one seed, so variation between AlphaFold 3 runs was not assessed. · stated by the authors
  • Predictions do not establish in vivo binding, stoichiometry or affinity, and functional assignments such as those for DUF4130 and DUF5819 still need experiments; confidence metrics depend on protein and interface size, and membrane proteins can give confident but spurious hydrophobic interfaces. · stated by the authors
  • Gene distance came from locus-tag numbering and is only approximate; picking one representative per module may drop informative interactions while keeping redundant ones. · stated by the authors
  • The model was cross-validated only within the bacterial STRING ≥900/≥10-taxa cohort, yet it also selected the eukaryotic (1,680) and archaeal (272) pairs in the final screen. · noted in this summary
  • The Pfam 38.2 retrospective check covers only 38 families, and its three agreement categories are qualitative definitions applied by the authors, with no blinded or independent assignment reported. · noted in this summary

Read this if you work on uncharacterized protein families or want precomputed AlphaFold 3 complexes and a STRING-based model for choosing which protein pairs are worth modelling.

Citations
1

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
THISHan '26Bryant '22Pannu '26Feldman '25Goodwin '25Salmoral '26
69% semantic matchSynthetic Biology
Yewon Han, Maxim I. Tsenkov, Niccolò Alberto Elia Venanzi+37 more·bioRxiv·Jul 4, 2026· 11 citations