Skip to content

TorsionTuner

The TorsionTuner is a specialized machine learning tool designed to bridge the gap between "idealized" static protein structures (like those from AlphaFold) and "dynamic" solution-state experimental data (SAXS/NMR).

By utilizing a JAX-based Graph Neural Network (GNN) and a differentiable kinematics layer, the model applies subtle adjustments to the backbone dihedral angles (\(\phi/\psi\)) to minimize the discrepancy between the predicted structure and real experimental observables.

πŸ”¬ Key Features

  • Torsional Prediction Strategy: Predicts \(\Delta\phi/\Delta\psi\) to maintain chemical constraints.
  • Differentiable Kinematics: Natural Extension Reference Frame (NeRF) implemented in JAX.
  • Multi-Objective Loss: Integrates SAXS, NMR, and Ramachandran geometry regularization.
  • GNN-Based: Captures spatial and sequential relationships via an Equinox-based GNN.

πŸ§ͺ Scientific Validation

TorsionTuner is validated at three levels:

  1. Internal parity β€” physics engines (CS predictor, Debye SAXS) verified against independent implementations
  2. NESG benchmark β€” CΞ± CSRMSD reduction on NESG targets using authentic BMRB experimental shifts
  3. External quality β€” structural quality tracked via MolProbity, PSVS, and ANSURR (planned)

Current benchmark results:

Target BMRB Residues Initial CSRMSD Final CSRMSD Ξ”
2KHD (Ξ±-helical) 16238 17 0.430 ppm 0.287 ppm βˆ’33%
2RN7 (mixed Ξ±/Ξ²) 11017 91 1.819 ppm β€” see noteΒΉ

ΒΉ 2RN7's CSRMSD improvement is limited by the residue-agnostic CS predictor floor (~1.8 ppm); see the roadmap for details.

For the full validation plan, benchmark methodology, and status of each item see docs/VALIDATION_ROADMAP.md.

πŸ“– Key References

  • NeRF Kinematics: Parsons, J., et al. (2005). J. Comput. Chem., 26(10), 1063–1068.
  • Chemical Shift Index (CSI): Wishart, D. S., & Sykes, B. D. (1994). J. Biomol. NMR, 4(2), 171–180.
  • Ramachandran Statistics: Lovell, S. C., et al. (2003). Proteins, 50(3), 437–450.
  • RPF Scores: Huang, Y. J., et al. (2005). J. Am. Chem. Soc., 127(5), 1665–1674.
  • Rosetta Refinement: Mao, B., et al. (2014). J. Am. Chem. Soc., 136(5), 1893–1906.
  • AlphaFold-NMR Assessment: Li, E. H., et al. (2023). J. Magn. Reson., 352, 107481.
  • Debye Formula for SAXS: Debye, P. (1915). Annalen der Physik, 351(6), 809-876.