henhaobio
Sign in

ScienceBiophysicsPeer-reviewed journal article

Computational design of conformation-biasing mutations to alter protein functions

Peter E. CavanaghiD, Andrew G. XueiD, Shizhong A. DaiiD, Albert Qiang, Tsutomu MatsuiiD, Alice Y. TingiD

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

Abstract

Most natural proteins alternate between distinct conformational states, each associated with specific functions. Intentional manipulation of conformational equilibria could lead to improved or altered protein properties. Here we develop Conformational Biasing (CB), a rapid and streamlined computational method that utilizes contrastive scoring by inverse folding models to predict variants biased towards desired conformational states. We validated CB across seven diverse deep mutational scanning datasets, successfully predicting variants of K-Ras, SARS-CoV-2 spike, {beta}2 adrenergic receptor, and Src kinase with improved conformation-specific functions including enhanced effector binding or enzymatic activity. Furthermore, applying CB to lipoic acid ligase, a conformation-switching bacterial enzyme that has been used for the development of protein labeling technologies, revealed a previously unknown mechanism for conformational gating of sequence-specificity. Variants biased toward the "open" conformation were highly promiscuous, while "closed" conformation-biased variants were even more specific than wild-type, enhancing the utility of LplA for site-specific protein labeling with fluorophores in living cells. The speed, simplicity, and versatility of CB (available at: https://github.com/alicetinglab/ConformationalBiasing/) suggest that it may be broadly applicable for understanding and engineering protein conformational dynamics, with implications for basic research, biotechnology, and medicine.

In-depth summary

Claude Opus 5.5 · xhigh effort · 5 Nov 2026 · from the full text of version 3 (2 Jun 2025) · CC BY 4.0

Contrastive scoring of protein variants across alternative backbone structures with inverse folding models (Conformational Biasing) predicts mutations that shift conformational equilibria and tune effector binding, catalysis, and labeling specificity.

Scoring mutations on a single structure with inverse folding models mainly selects for expression rather than state-specific function, while molecular dynamics, AlphaFold, and ensemble samplers like BioEmu are computationally expensive at scale. Contrastive scoring across alternative conformations screens >3,000 variants in under a minute on one GPU without model fine-tuning.

Key results

  1. ShownConformational Biasing (CB) accurately predicts state-specific functional variants across seven deep mutational scanning datasets and 10 synthetic binder pairs, outperforming single-structure ProteinMPNN scoring.

    CB significantly separates state-biased variants (p < 0.0001) for K-Ras (>30,000 variants; 3320 single and 26761 double mutants), SARS-CoV-2 RBD (4022 mutants), beta2AR (5045 variants in text; 5263 in Fig. 2I), and Src kinase, plus B-Raf, FabZ, MurA, and 10 binder pairs (RMSD 0.23-2.6, p < 0.01), whereas single-structure scoring fails after expression normalization. · Results; Benchmarking CB; Figure legends

  2. ShownCB-predicted mutations in E. coli lipoic acid ligase (LplA, 38 kD) shift the solution equilibrium between closed and open adenylate-ester-bound conformations.

    In SEC-SAXS with non-hydrolyzable lipoyl-AMS, wild-type and W37V LplA shift only 22.3-22.4% to the open state, whereas CB-designed open-biased variants E265R^V and A48N^V reach 90% and 99% open occupancy and closed-biased variants C55M^V and T57I^V reach 12% and 16% open; across 24 variants, CB scores correlate with tryptophan fluorescence changes (Spearman r = -0.61). · CB predicts conformationally-biased variants of…; Benchmarking CB; Figure legends

  3. ShownBiasing LplA toward the open conformation increases promiscuous proteome labeling in living cells, whereas closed-biased variants increase sequence specificity by tethered C-terminal domain inhibition.

    Across 99 W37V-background variants assayed by flow cytometry in HEK293T cells (31 open-biased, 32 closed-biased, 36 neutral), 23/25 clash-free open-biased variants (>90%) increase promiscuous BCN labeling or Trp shift and 31/32 closed-biased variants (>95%) reduce promiscuity while retaining on-target E2p labeling, enabling higher signal-to-noise LAP tagging (T57I^V, p < 0.005). · CB predicts conformationally-biased variants of…; Benchmarking CB; Figure legends

  4. ShownCB generalizes across four inverse folding models and outperforms AFCluster and BioEmu in accuracy, sequence coverage, and computational speed on LplA.

    ProteinMPNN, ESM-IF1, Frame2Seq, and ThermoMPNN all correlate with LplA promiscuity (Spearman r = 0.63-0.7, or r = 0.67 for CB vs. r = 0.26 for AFCluster, which fails to score 1250 LplA mutants or 20% absent from the MSA), while BioEmu correlates worse with Trp fluorescence across 24 variants (Spearman r = -0.27 vs. -0.61 for CB) and takes days vs. <1 minute on an RTX 4090 GPU. · Benchmarking CB; Discussion; Figure legends

Methods, data and reproducibility

Conformational Biasing scores mutant libraries against two or more experimental or AlphaFold backbone conformations using inverse folding models (ProteinMPNN, ESM-IF1, Frame2Seq, ThermoMPNN) to identify variants with divergent state scores. Benchmarked on 7 deep mutational scanning datasets, AFCluster, and BioEmu, and validated on E. coli W37V LplA by SEC-SAXS, Trp fluorescence, and live-cell Click labeling.

  • Deep mutational scanning dataset of K-Ras single and double mutants bound to State 1 (PDB: 8T71) and State 2 (PDB: 6XHB) effectors (Weng et al.) · human · n = >30,000 mutants (3320 single and 26761 double mutants) · 8T71
  • Deep mutational scanning dataset of SARS-CoV-2 spike RBD expression and ACE2 binding in down (PDB: 7XIX) and up (PDB: 7XO8) conformations (Starr et al.) · SARS-CoV-2 · n = 4022 point mutants · 7XIX
  • Deep mutational scanning dataset of beta2-adrenergic receptor basal activation in inactive (PDB: 3SN6) and active (PDB: 2RH1) states (Jones et al.) · human · n = 5045 variants (5263 in Fig. 2I) · 3SN6
  • Deep mutational scanning datasets of Src kinase (PDB: 2SRC, 1Y57), B-Raf kinase, and E. coli FabZ and MurA enzymes, plus 10 synthetic protein binder pairs · multiple species · n = 4 enzyme DMS datasets and 10 synthetic binder pairs · 2SRC
  • E. coli lipoic acid ligase (LplA, PDB: 1X2H, 1X2G, 3A7R, 3A7A) variants assayed by SEC-SAXS, Trp fluorescence, and mammalian cell labeling · E. coli (expressed in E. coli, HEK293T, and HeLa cells) · n = 6 variants by SEC-SAXS, 24 by Trp fluorescence, 99 by flow cytometry · 3A7R
Methods & tools
Conformational Biasing (CB), ProteinMPNN, ESM-IF1, Frame2Seq, ThermoMPNN, AlphaFold2, AlphaFold3, AFCluster, BioEmu, Oligomer, DENSS
Software & environment
Google Colab notebook provided at https://github.com/alicetinglab/ConformationalBiasing/
Assay conditions & compute
>3000 variants analyzed in under 1 minute on a single consumer-grade RTX 4090 GPU
Limitations (2)
  • Conformational Biasing requires high-quality experimental or predicted backbone structures in at least two functionally relevant conformational states and has limited predictive power across the full set of all point mutants. · stated by the authors
  • Because current inverse folding models do not explicitly model bound small-molecule substrates, several of the 31 open-biased LplA variants (leaving 25 clash-free) had low activity due to steric clashes with the BCN substrate. · stated by the authors

Read this if you engineer proteins, receptors, or enzymes and want a fast inverse-folding workflow to design point mutations that shift conformational equilibria without molecular dynamics simulations.

Citations
31
FWCI
60.67
Field percentile
top 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
THISDaniel '26Sauer '19Biel '16Mitrovic '23Stern '25Boer '20
60% semantic matchBioinformatics
Julie Daniel, lucas vitoriano, Diego Javier Zea·bioRxiv·Aug 22, 2026