ConforFold

Many proteins are not defined by a single fixed shape but interconvert between two or more functionally distinct conformations. Standard structure predictors such as AlphaFold2 and OpenFold typically collapse onto one dominant state, and the common workaround of subsampling the multiple sequence alignment (MSA) recovers alternative states unreliably. ConforFold tackles this problem directly by injecting structural diversity through secondary structure rather than through the MSA.

ConforFold is a two-stage framework. The first component, Confor-PSSP, is a transformer trained on multi-residue fragments to generate diverse eight-state protein secondary structure predictions (PSSPs) for a given sequence. These alternative secondary-structure hypotheses are then used to condition a retrained OpenFold model, ConforFold, which folds the sequence toward each hypothesized state. By steering folding with varied but physically plausible secondary-structure inputs, the pipeline surfaces multiple conformers that MSA subsampling misses.

ConforFold was developed by Reza Syrlybaeva and Eva-Maria Strauch at Washington University in St. Louis. It was posted as a preprint in October 2025 and published in Protein Science in 2026.

#Key Features

• Secondary-structure-driven conformer sampling: Diversity is introduced through eight-state secondary structure predictions rather than MSA subsampling, giving a more controllable route to alternative states.
• Confor-PSSP generator: A transformer trained on multi-residue fragments produces a range of plausible secondary-structure assignments to seed folding.
• Retrained OpenFold backbone: The folding stage is an OpenFold model retrained to accept secondary-structure conditioning and resolve distinct conformers.
• Strong alternative-conformer recovery: On proteins with two known conformations, ConforFold correctly recovers both states more often than competing ensemble methods.

#Technical Details

The framework couples two deep learning models. Confor-PSSP is a transformer trained on multi-residue fragments to emit diverse eight-state PSSPs, which serve as conditioning inputs to ConforFold, a retrained OpenFold structure predictor. On a benchmark of proteins that adopt two alternative conformations, ConforFold correctly identified both conformers in 84% of cases at TM-scores of at least 0.8. This surpasses AlphaFlow (75.4%), a diffusion-based ensemble generator, and Cfold, which relies on MSA clustering to enumerate states. Code for both the Confor-PSSP and ConforFold stages is released on GitHub.

#Applications

ConforFold is useful for structural biologists and protein engineers studying proteins with multiple functional states, such as receptors, transporters, and fold-switching proteins. By producing structural models for distinct conformers, it supports mechanistic hypotheses about conformational change, guides the design of state-selective binders and stabilizers, and complements experimental methods that capture only one state at a time.

#Impact

By reframing conformational sampling as a secondary-structure problem, ConforFold offers a more reliable alternative to MSA subsampling for recovering the multiple states that many proteins occupy. Its measured gains over AlphaFlow and Cfold on two-state proteins, together with peer-reviewed publication in Protein Science and open-source code, make it a practical addition to the growing toolkit for protein conformational ensemble prediction.