ForceFold: An open-source AI platform for predicting single-protein mechanics and designing targeted mechanotherapies
CV (Salapaka) CV (Ervasti) CV (Muretta) CodeAbstract
Analysis of single-molecule data is vital to studying the force-bearing molecules driving diseases like muscular dystrophy and cancer. However, discovering targeted mechanotherapies remains severely bottlenecked because current experimental methods struggle with confounding multi-molecule noise and dauntingly slow manual analysis. Furthermore, traditional molecular dynamics simulations are computationally prohibitive for high-throughput screening. To overcome these critical barriers, we propose ForceFold: an open-source AI framework integrating polymer physics with machine learning to revolutionize force-bearing protein analysis and design. Our closed-loop pipeline automates single-molecule data extraction via a physics-augmented neural network (PemNN) and utilizes a novel physics-aware deep clustering architecture (Latent Unfold) to automatically decipher complex, heterogeneous protein domains. Furthermore, we deploy a physics-guided generative diffusion model (GenUnfold) to predict mechanical fingerprints directly from sequences and solve the inverse problem, engineering novel therapeutics for mechanobiological diseases. By open-sourcing our AI tools and massive single-molecule datasets, ForceFold aims to accelerate targeted therapeutic discovery for academic institutions and industry leaders worldwide.
Poster
Publications
Multiple modes of AFM reveal distinct mechanical properties for dystrophin and utrophin not manifest by small fragments
Using two modes of atomic force microscopy (AFM) and Monte Carlo simulations, this study reveals that full-length dystrophin exhibits brittle unfolding behavior, while full-length utrophin demonstrates a complex stiffening spring behavior. These findings provide critical insights into the potential efficacy of utrophin upregulation as a therapy for Duchenne muscular dystrophy.
PNAS 2026GenUnfold: Rapidly Predict Protein Mechanical Unfolding Trajectory via a Physics-Guided Diffusion Model
This work introduces GenUnfold, the first scalable generative diffusion framework for predicting full protein unfolding trajectories. By combining global coevolutionary context with local structural stiffness, the model achieves state-of-the-art performance in predicting physically consistent mechanical properties.
ICML 2026Resolving Heterogeneous Mechanical Domains via Physics-Aware Deep Clustering of Single-Molecule Force Spectroscopy Data
This work introduces Latent Unfold, the first automated framework to identify heterogeneous folding domains in SMFS data. By utilizing a physics-aware deep clustering architecture with dual autoencoders, the model resolves domain-level behavior in complex proteins like dystrophin and utrophin without prior knowledge.
Submitted to ACS NanoA Physics-Augmented Deep Learning Framework for Classifying Single Molecule Force Spectroscopy Data
A physics-augmented deep learning framework designed to automatically classify and filter single-molecule force spectroscopy data, providing an automation tool to ensure analysis is based on validated single-molecule events.
ICML 2025Distinct mechanical properties in homologous spectrin-like repeats of utrophin
This paper reports the first mechanical characterization of utrophin using AFM. The data indicates that the mechanical properties of utrophin's spectrin-like repeats are more similar to the stiff repeats of titin than those of spectrin or dystrophin, suggesting utrophin functions as a stiff elastic element at the myotendinous junction.
Scientific Reports 2019