Skip to content

Ground Truth Philosophy: Geometry vs. Physics

This document outlines the philosophical and scientific rationale behind how synth-pdb defines "Ground Truth" for protein structures, particularly in the context of training and benchmarking AI models.

The Central Tension: Idealism vs. Realism

In structural biology, there is a fundamental tension between two types of "Truth":

  1. Geometric Truth (The Ideal): A structure defined by precise mathematical internal coordinates (torsions, bond lengths, angles).
  2. Physical Truth (The Reality): A structure that has reached an energy minimum within a specific forcefield and solvent environment.

synth-pdb prioritizes Geometric Truth as the primary "Ground Truth" label.

Why Non-Minimized Structures are "Truth"

For a structural biologist, a non-minimized structure with steric clashes looks "wrong." However, for a Machine Learning engineer, this structure is the most "honest" label possible.

1. Mathematical Determinism

The synth-pdb generator is a deterministic mathematical engine. When you request an alpha helix, it places atoms exactly at the Ramachandran centers (e.g., \(\phi = -60^\circ, \psi = -45^\circ\)). * The Signal: The "Ground Truth" coordinates represent the intent of the generation. * The Noise: Steric clashes are "noise" resulting from the high-dimensional complexity of protein packing.

If we minimize the structure before calling it "truth," we allow a physics engine (like OpenMM) to rewrite our intent. We are no longer testing if an AI can learn to build an alpha helix; we are testing if it can learn the specific preferences of the Amber14 forcefield.

2. Forcefield Agnosticism

Physics engines are models, not reality. Every forcefield (Amber, CHARMM, OPLS) has its own biases. * If we define ground truth as the output of an Amber14 minimization, our dataset is "biased" toward Amber14. * An AI trained on this data might perform poorly on experimental data that doesn't perfectly match Amber's specific energetic curves.

By using the NeRF-generated ideal geometry as ground truth, we provide a stable, forcefield-independent reference point.

3. Measuring "Conformational Strain"

By preserving the non-minimized state as the reference, we can calculate a critical metric: Strain-RMSD. * Formula: \(RMSD(GroundTruth, Minimized)\) * Interpretation: If the RMSD is low (< 0.5 Å), the "ideal" geometric fold was physically plausible. If the RMSD is high (> 2.0 Å), the intended fold was physically impossible (a "clash-trap").

The "Best of N" Strategy: A Bridge to Plausibility

To satisfy both the Geometer and the Physicist, synth-pdb recommends a Best of N selection strategy rather than a Minimize-to-Relax strategy.

The Algorithm

Instead of taking one structure and moving its atoms to resolve clashes, we: 1. Generate \(N\) independent versions of the structure using small stochastic "drifts" in torsion angles. 2. Calculate the number of steric violations for each. 3. Select the one with zero or minimum violations.

Why this is superior for AI

  • No Warp: The selected structure still uses "ideal" bond lengths and angles. No atoms were "pushed" by a forcefield.
  • Physically Valid: The structure is clash-free and exists in a low-energy state.
  • Clean Labels: The AI learns to associate a sequence with a physically plausible but still mathematically ideal conformation.

Practical Recommendations for ML Research

Use Case Recommended Ground Truth Rationale
Pre-training Non-minimized Maximizes signal/noise ratio for geometric patterns.
Fine-tuning Best of N (Selected) Introduces physical plausibility without forcefield bias.
Physical Audit Minimized Use as a secondary check to see if the AI's "ideal" prediction is physically stable.

Scientific Conclusion

In synth-pdb, the Ground Truth is the Source of Intent. We treat the structure as a "Label" that an AI should strive to recover. While minimization is a powerful tool for refinement, it should be treated as an analytical step rather than a definition of truth.

See Also