Co-folding model guided by structural proteomics

Researchers introduce AIMS-Fold, a guided-diffusion framework that integrates structural proteomics data (XL-MS and HDX-MS measurements) with protein structure prediction models to improve accuracy in predicting protein complex conformations. The approach outperforms unguided computational models on challenging induced proximity drug targets, advancing structure-based drug design capabilities.

AIMS-Fold represents a meaningful convergence of experimental biology and machine learning, addressing a persistent gap in protein structure prediction. While generative models like AlphaFold excel at predicting individual protein structures from sequence alone, they struggle with protein complexes in specific conformational states—a critical limitation for rational drug design, particularly for induced proximity modalities like PROTACs and antibody therapeutics. This work demonstrates how sparse experimental constraints from structural proteomics can guide neural network sampling without requiring complete structural information.

The technical innovation lies in transforming heterogeneous experimental data (cross-linking distances and solvent exposure profiles) into differentiable physical potentials that steer diffusion model trajectories during inference. Rather than retraining models, this approach operates at inference time, making it broadly applicable to existing pretrained architectures. The synergistic improvement from combining XL-MS and HDX-MS data suggests these techniques capture complementary spatial information.

For the biotech and pharmaceutical industry, this framework accelerates the computational validation pipeline for protein-based therapeutics, reducing experimental cycles needed to validate target engagement predictions. The demonstrated superiority over Boltz-2 on induced proximity targets indicates practical utility for PROTAC and bispecific antibody design, where conformational accuracy directly impacts drug efficacy and specificity.

The commitment to releasing evaluation code upon publication suggests genuine scientific transparency and positions this work as a reproducible benchmark. Future development likely involves scaling to larger complexes and integrating additional structural proteomics modalities, potentially establishing this integrative approach as standard practice in structure-guided drug discovery workflows.

• →AIMS-Fold combines experimental proteomics data with diffusion models to improve protein complex structure prediction accuracy.
• →The framework outperforms unguided computational models on induced proximity drug targets critical for PROTAC and antibody design.
• →Integration of XL-MS and HDX-MS data yields synergistic improvements, suggesting complementary information capture.
• →Inference-time guided diffusion enables application to existing pretrained models without requiring retraining.
• →Framework advancement reduces computational validation cycles in structure-based drug design workflows.