Pkad R Eval

pkad-r-eval

Hybrid Quantum-Classical Encoding for Accurate Residue-Level pKa Prediction — Van Le et al. (2026) (arXiv:2603.11061, 2026)

What this evaluates

This evaluation probes the ability of hybrid quantum-classical models to accurately predict residue-level pKa values across diverse protein microenvironments. It tests whether entanglement-aware quantum feature mappings generalize beyond the training distribution to experimental datasets and capture subtle electronic and geometric correlations in flexible peptide regions.

Datasets

• PKAD-R — total ?; splits: train (-1), test (-1)
• Aβ40 — total 3; splits: test (3)

Metrics

• RMSE (primary) — range: other
  • Root Mean Square Error: sqrt(mean((y_true - y_pred)^2)). Lower values indicate higher predictive accuracy.
• MAE — range: other
  • Mean Absolute Error: mean(|y_true - y_pred|). Measures average magnitude of prediction errors in pKa units.
• MaxErr — range: other
  • Maximum Absolute Error: max(|y_true - y_pred|). Captures the worst-case prediction deviation.
• Pearson correlation (R) — range: other
  • Pearson correlation coefficient between true and predicted pKa values. Higher values indicate stronger linear agreement with experiment.
• Regression slope (m) — range: other
  • Slope of the linear regression line fitting predicted values to true values. Values closer to 1 reflect stronger linear agreement.

Input / output format

Input: Residue-level protein structures processed through an entanglement-aware quantum feature mapping to produce hybrid quantum-classical embeddings.

Output: Continuous pKa value prediction (float) per residue.

Scoring recipe

def compute_metrics(y_true, y_pred):
    import numpy as np
    from scipy.stats import pearsonr
    errors = y_true - y_pred
    rmse = np.sqrt(np.mean(errors**2))
    mae = np.mean(np.abs(errors))
    max_err = np.max(np.abs(errors))
    r, _ = pearsonr(y_true, y_pred)
    slope = np.polyfit(y_pred, y_true, 1)[0]
    return {'RMSE': rmse, 'MAE': mae, 'MaxErr': max_err, 'R': r, 'Slope': slope}

Common pitfalls

• Overfitting to high-dimensional quantum feature space: models like GradientBoosting achieve near-zero training error but suffer severe degradation on test data due to fitting noise rather than transferable structure-function relationships.
• Sensitivity to structural flexibility and underrepresented regions: residues in highly dynamic or solvent-exposed environments yield higher prediction errors and variance due to limited training distribution coverage and higher experimental uncertainty.
• Misinterpreting correlation-based quantum features as capturing all local biochemical effects: the encoding emphasizes nonlocal electronic/geometric correlations, potentially missing local descriptors like solvent accessibility or backbone dihedrals.

Evidence (verbatim from paper)

Table I reports the performance of four representative models—DQNN, GradientBoosting, GPR_SE, and kNN—across RMSE, MAE, maximum absolute error, Pearson correlation (R), and regression slope (m). Lower RMSE, MAE, and MaxErr values indicate higher predictive accuracy, while higher R and slopes closer to 1 reflect stronger linear agreement with experimental measurements.

Citation

@misc{vanle2026hybrid,
  title={Hybrid Quantum-Classical Encoding for Accurate Residue-Level pKa Prediction},
  author={Van Le et al. (2026)},
  year={2026},
  note={arXiv:2603.11061}
}

• arXiv: 2603.11061