Constraint-Guided Protein Structure Reconstruction Using ESM-2 Representations and Multi-Task Deep Learning

Native-free protein-structure reconstruction requires candidate generation, refinement, physical assessment, and ranking without access to native coordinates. The proposed work presents an ESM-2-guided framework that combines multi-task prediction of inter-residue distances, contacts, and backbone torsion angles with training-only fragment retrieval, stochastic assembly, differentiable refinement, OpenMM cleanup, and frozen multi-criteria ranking. A leakage-controlled CASP12 dataset was used for model development, with proteins restricted to 40–250 residues and exact sequence overlap removed across partitions. The frozen pipeline was externally applied to 15 eligible single-chain CASP15 targets after metadata and training-homology screening. Fifty candidates were generated and refined for each target, producing 750 candidates; 150 shortlisted structures underwent OpenMM minimization, and one physically valid model was selected per target without native-coordinate access. Post-hoc evaluation was performed for 10 targets with resolved native structures. The selected models achieved mean C-alpha RMSD of 14.5366 Å, backbone RMSD of 14.4318 Å, pairwise distance MAE of 6.8380 Å, and contact F1 of 0.1032. The results establish the operational feasibility and auditability of the pipeline but indicate limited structural accuracy. The framework is therefore presented as a reproducible proof of concept rather than a high-accuracy or state-of-the-art folding system.