henhao
Sign in
BiophysicsbioRxiv
Preprint

FKSFold: Improving AlphaFold3-Type Predictions of Molecular Glue-Induced Ternary Complexes with Feynman-Kac-Steered Diffusion

Jian SheniD, Shengmin ZhouiD, Xing CheiD

Abstract

We introduce an AI model, FKSFold, that uses Feynman-Kac steered diffusion as an inference-time strategy to improve AlphaFold3-type predictions of molecular-glue induced ternary complexes. FKSFold augments reverse diffusion with a Feynman-Kac derived steering term and uses the interface predicted TM-score (ipTM) as the guiding potential, coupled with particle-based sampling and adaptive resampling to bias trajectories toward high-quality interfaces without retraining the base models. Implemented on Chai-1r and Boltz-2 (named FKSFold-Chai and FKSFold-Boltz) - two open-source AF3-type architectures - FKSFold was benchmarked on eight ternary systems. FKSFold-Chai recovered three challenging complexes with sub-3 [A] accuracy: VHL:MG:CDO1 (DockQ 0.922, iRMSD 0.629, fnat 0.963), FKBP12:MG:mTOR-FRB (DockQ 0.590, iRMSD 1.813, fnat 0.519), and FKBP12:MG:BRD9 (DockQ 0.841, iRMSD 0.801, fnat 0.913), while unmodified baselines failed. Other cases highlighted challenges - flexible loop rearrangements (e.g., CRBN:MG:VAV1-SH3c) and large conformational search spaces (e.g., NEK7, QDPR, HDAC1) - and, empirically, using few particles, short resampling intervals, and moderate-to-high lambda (steering strength) gave the best exploration-exploitation balance. Although the strategy is theoretically model-agnostic, on Boltz-2 ipTM head instability currently limits observed gains. The code for the two implementations, FKSFold-Chai and FKSFold-Boltz, is available at https://github.com/YDS-Pharmatech/FKSFold-Chai and https://github.com/YDS-Pharmatech/FKSFold-Boltz.

PDFDOICode
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
4
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (14 Oct 2025) · CC BY 4.0

Feynman-Kac steering with an ipTM reward at inference time lets Chai-1r recover three of eight molecular-glue ternary complexes below 3 Å without retraining.

Molecular glue discovery is largely serendipitous, and AlphaFold3-type models struggle with the energetic landscape of three-body ternary complexes. This work adds training-free, ipTM-guided steering of the diffusion sampler, avoiding classifier guidance that needs extra training.

Key results

  1. ShownOn VHL:MG:CDO1, FKSFold-Chai turned a failed Chai baseline prediction into a near-native interface.

    Chai baseline vs FKSFold-Chai: DockQ 0.012 to 0.922; iRMSD 15.899 Å to 0.629 Å; fnat 0.000 to 0.963. · VHL:MG:CDO1: A Novel VHL-based Molecular Glue…; Abstract

  2. ShownFKSFold-Chai also improved the two non-E3-ligase glue systems FKBP12:MG:mTOR-FRB and FKBP12:MG:BRD9 over the unsteered Chai-1 baseline.

    mTOR-FRB: DockQ 0.590, iRMSD 1.813 Å vs baseline DockQ 0.076, iRMSD 9.172 Å. BRD9: DockQ 0.841, iRMSD 0.801 Å, fnat 0.913 vs baseline DockQ 0.258, iRMSD 2.569 Å, fnat 0.250. · FKBP12:MG:mTOR-FRB: A Challenging Non-E3 Ligase…; FKBP12:MG:BRD9: Another Challenging Non-E3…

  3. SuggestedSteering reached RMSD below 3 Å in 3 of 8 systems. For four unsolved cases the authors suggest flexible loops (CRBN:MG:VAV1-SH3c) and large proteins (NEK7, QDPR, HDAC1) as likely causes.

    3 of 8 cases reached RMSD below 3 Å. Input sizes: NEK7 283, QDPR 230, HDAC1 369 residues vs CDO1 185 and BRD9 98 in the successful cases. The VAV1 RT-loop residues D797 and R798 were mispositioned. · Performance on Molecular Glue Complexes; Analysis of Remaining Challenging Cases

  4. SuggestedPerformance depends strongly on particle number, path length, lambda and potential type. Few particles, short resampling intervals and moderate-to-high lambda worked best, with system-specific optima.

    [2, 5, 20.0, vanilla] gave RMSD 1.44 Å (VHL:MG:CDO1) and 1.45 Å (mTOR-FRB). For BRD9 with [5, 5, 10.0], max potential gave 2.14 Å vs vanilla 2.52 Å; [5, 10, 10.0, vanilla] gave 10.55 Å. · Hyperparameter Sensitivity in FK Steering; Lambda Weight; Potential Type; Particle Numbers and Diffusion Path Length

  5. ClaimedThe strategy is theoretically model-agnostic, but on Boltz-2 gains are limited by an unstable ipTM confidence head. Chai-1r gave consistently better results.

    The text reports that Boltz-2 gives higher but more unstable confidence scores and greater variability, and that Chai-1r performed better. Boltz-2 numeric metrics (Table 2) are not printed in the text. · Abstract; Introduction; VHL:MG:CDO1: A Novel VHL-based Molecular Glue…; FKBP12:MG:mTOR-FRB: A Challenging Non-E3 Ligase…

Methods, data and reproducibility

Reverse diffusion in Chai-1r and Boltz-2 gets a Feynman-Kac steering term. Parallel particles are scored by confidence-head ipTM on partially denoised states, then periodically resampled multinomially with weights from an exponential potential of lambda times the scores.

  • Eight molecular glue-induced ternary complexes; references are PDB entries or YDS-GlueFold predictions · n = 8 ternary systems
  • FKBP12:MG:mTOR-FRB reference structure · PDB 8PPZ
Methods & tools
FKSFold-Chai, FKSFold-Boltz, Chai-1r, Boltz-2, ipTM confidence score, Feynman-Kac steering with multinomial resampling, YDS-GlueFold (reference predictions), DockQ, iRMSD, fnat metrics
Code
https://github.com/YDS-Pharmatech/FKSFold-Chaihttps://github.com/YDS-Pharmatech/FKSFold-Boltz
Reproducibility gaps
Some reference structures come from YDS-GlueFold predictions, not experiment. No dataset links, software versions or hardware details are given in the text.
Limitations (5)
  • Only 3 of 8 systems reached RMSD below 3 Å. The authors link four unsolved cases to flexible loops (CRBN:MG:VAV1-SH3c) and larger proteins (NEK7, QDPR, HDAC1). · stated by the authors
  • On Boltz-2 the confidence head gives unstable scores that limit the gains from ipTM-based steering. · stated by the authors
  • Results are sensitive to particle number, path length, lambda and potential type, so per-system tuning is needed. · stated by the authors
  • Some reference structures are YDS-GlueFold predictions rather than experimental structures, so the accuracy comparison is partly against model output. · noted in this summary
  • Hyperparameters were explored on the same three solved systems used for reporting, and no variance estimates are given, so the reported gains may be optimistic. · noted in this summary

Who should read:Read this if you work on inference-time steering of AlphaFold3-type diffusion models or want an honest failure analysis for molecular-glue ternary complex prediction.

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
THISChe '24Elliott '24Feldman '25Neupane '26Chen '26Liao '25
75% semantic matchBiophysics
Xing Che, The Statistical Mechanics Inference Team·bioRxiv·Dec 23, 2024· 8 citations