henhao
Sign in
BioinformaticsbioRxiv
Preprint

Modeling the structure-conditioned sequence landscape for large-scale protein design with TriFlow

Harish SrinivasaniD, Rongqing YuaniD, Jing ZhangiD, Qian CongiD, Jian ZhouiD

Abstract

Generative models have revolutionized computational protein design, and the design of high-quality sequences given backbone structure is a critical component for success. Current state-of-the-art design pipelines utilize sequence design methods with local structural context and autoregressive generation. To improve efficiency and quality of sequence design, we developed TriFlow, a model that combines a RoseTTAFold-like three-track architecture for global structural context with discrete flow-matching for efficient few-step sequence generation. We trained TriFlow on a large dataset of interacting protein chains from Protein Data Bank and interacting domains from AlphaFold protein structure Database to enrich its knowledge of natural protein and domain interfaces. TriFlow outperforms existing sequence design methods like ProteinMPNN across diverse benchmarks, including de novo binder design, where it boosts the in silico success rate of state-of-the-art design pipelines such as BindCraft. We demonstrated this by conducting a large-scale benchmark, generating and computationally validating binders for over 500 diverse protein targets. Experimental validation on a small set of targets also suggests that the performance of TriFlow is on par with BindCraft. By leveraging the model to explore the designed sequence landscape, we discovered that we can effectively highlight functional sites, by contrasting designed sequences that reflect structure constraints with natural evolutionary profiles. As a practical demonstration of its capabilities, we applied our pipeline to systematically design specific binders against human class I cytokines, computationally optimizing for on-target affinity while minimizing off-target interactions, demonstrating that specificity also scales with inference time and computational budget. TriFlow thus provides a robust framework both for large-scale protein engineering and for exploring the fundamental principles of the structure-conditioned sequence landscape. TriFlow software and generated resources are available at https://triflow.zhoulab.io/.

PDFDOI
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
2
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Sonnet 4.6 · xhigh effort · 7 Nov 2026 · from the full text of version 2 (1 Aug 2026) · CC BY 4.0

TriFlow, discrete flow-matching sequence designer with global three-track structure attention, raises in silico binder success over ProteinMPNN in RFdiffusion and BindCraft-style pipelines on 500 targets and flags functional sites.

Binder design pipelines almost all use ProteinMPNN as the sequence designer, and its local receptive field limits capture of long-range structural dependencies. TriFlow instead conditions on global structure and decodes in about ten parallel steps, and the authors also build a 500-target binder benchmark.

Key results

  1. ShownOn monomers, TriFlow recovered native sequences better than ProteinMPNN, ESM-IF and PiFold, and its sequences refolded better, especially at 600 to 1000 residues. Its designed sequences also showed coevolution signals closer to true contacts.

    1392 shared PDB test structures; refoldability is the best AlphaFold3 single-sequence TM-score for backbones of 100 to 1000 residues; Gremlin coevolution on 30 proteins with 1,000 sequences each, versus ProteinMPNN and ESM-IF. No recovery percentages are given in the text. · TriFlow improves protein sequence design; Sequence Recovery, Refoldability and…

  2. ShownIn a 24-target binder panel, TriFlow gave higher AlphaFold3 structure success than ProteinMPNN (RFdiffusion backbones) and SolubleMPNN (BindCraft backbones). It also won on sequence success for most targets.

    8 sequences per backbone; 21 of 24 targets evaluable (TGFβ, H3, Der f21 excluded); sequence success better for 19 of 21 targets; BindCraft benchmark improved for all targets but PD-1; full redesign including the interface gave the highest structure success. · TriFlow improves state-of-the-art de novo binder…; Binder Design Benchmarks

  3. ShownOn a new 500-target benchmark built from binary PDB complexes, TriFlow beat the MPNN baselines on more targets in both pipelines. Success rose with overlap with the native interface and with helical, negatively charged target interfaces, and fell with larger native interfaces.

    16,216 backbones (12,500 RFdiffusion, 3,716 ColabDesign), 129,728 sequences; structure success better for TriFlow on 143 targets vs 68 for ProteinMPNN (RFdiffusion) and 200 vs 93 for SolubleMPNN (ColabDesign); logistic regression of target features. · Large-scale binder design benchmark; Large-scale Binder Design Benchmark

  4. SuggestedFor human class I cytokines, ColabDesign backbones with TriFlow sequences gave AlphaFold3-confident binders for all targets except IL7. Best predicted specificity rose with the number of sequences designed per target, then plateaued for most.

    31 cytokines; up to 10 backbone trials and 40 sequences per backbone; specificity score from on-target vs best off-target ipAE_min; largest gain going from 1 to 30 sequences; GH2 stayed moderate; IL15 example on-target ipAE_min 1.04. All values are AlphaFold3 predictions. · Designing specific binders for all class I…; Design Specific Binders for Class I Cytokines

  5. ShownContrasting natural conservation with conservation in TriFlow-designed profiles highlights functional sites. Designed profiles are barely conserved at active sites, so subtracting them removes structural constraint and improves active-site and ligand-site detection.

    Designed 500 sequences for each of 43,659 ECOD domains (temperature 0.5); 19,264 domains with at least 100 homologs analysed; precision-recall at 1:10 positive-to-negative ratio improved by the difference score over natural conservation alone (no numeric values given). · Designed sequence profiles informing functional…; Comparison of Natural and Designed Profiles and…

Methods, data and reproducibility

A three-track (single, pair, 3D) invariant-point-attention network is trained with a discrete flow-matching objective to generate sequences from backbone coordinates in ten steps. It is evaluated on sequence recovery, refoldability, binder design pipelines scored by AlphaFold3, and profile comparisons with natural MSAs.

  • PDB assemblies up to August 2021 with ProteinMPNN splits (30% identity clusters); test set of 1392 structures · n = 1392 test structures
  • Extended training set adding representative AlphaFold Database (AFDB) models and predicted interacting domain pairs
  • Large-scale binder design benchmark: binary PDB complexes released August 2021 to April 2025, filtered for low similarity to training data · n = 500 targets
  • Therapeutic targets from the RFdiffusion and BindCraft papers · n = 24 targets
  • Human class I cytokines (AFDB structures, signal peptides removed) · human · n = 31 cytokines
  • Evolutionary Classification of Domains (ECOD) domains with UniProt annotations and HHblits MSAs against UniRef30 · n = 43,659 designed; 19,264 analysed
Methods & tools
TriFlow, AlphaFold3, RFdiffusion, ColabDesign, BindCraft, ProteinMPNN, SolubleMPNN, ESM-IF, PiFold, Gremlin, HHblits, AL2CO
Assay conditions & compute
Trained on 4 A100 GPUs (batch size 8 per GPU); BindCraft run for 2 days per target on one A100; TriFlow designs a sequence in seconds, versus hours per backbone for ColabDesign.
Reproducibility gaps
An open-source repository with tutorials and a public set of designed cytokine binder sequences are mentioned, but no URL is given in the text.
Limitations (6)
  • Success is judged with AlphaFold-based metrics as a stand-in for experiments; no experimental validation, so the real-world gain from higher in silico success is unquantified. · stated by the authors
  • Only 8 sequences per backbone were designed in the large benchmark, so per-target success rates vary stochastically. · stated by the authors
  • The AlphaFold3 wrapper handled only single-chain targets, so TGFβ, H3 and Der f21 could not be evaluated. · stated by the authors
  • Designed binders are overwhelmingly α-helical, and current pipelines assume rigid backbones, which inflates conservation in loops; structure and sequence are designed separately. · stated by the authors
  • The BindCraft comparison uses a modified ColabDesign-only version (fixed trial count, no Rosetta or AlphaFold2 filtering), not the full pipeline; the benchmark was built by the same authors. · noted in this summary
  • The functional-site analysis keeps only 19,264 of 43,659 designed domains, those with deep MSAs, and relies on UniProt annotations for positives. · noted in this summary

Who should read:Read this if you run binder design pipelines such as RFdiffusion or BindCraft and want a faster ProteinMPNN replacement with a 500-target in silico benchmark.

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
THISYan '25Li '25Akpinaroglu '23Chacón-Sánc '25Team '25Lai '22
78% semantic matchBioengineering
Junyu Yan, Zibo Cui, Wenqing Yan+4 more·bioRxiv·Apr 29, 2025· 4 citations