MillerBind v9 & v12: Independent Validation Using the Therapeutics Data Commons

Structure-Based Protein-Ligand Binding Affinity Prediction

William Miller

BindStream Technologies

February 26, 2026

1. Introduction

Accurate prediction of protein-ligand binding affinity is fundamental to structure-based drug discovery. Scoring functions estimate binding free energy (ΔG) or the dissociation constant pK_d = −log₁₀(K_d) from three-dimensional protein-ligand complex structures. The CASF-2016 benchmark (Su et al., 2019) is the accepted standard for evaluating scoring power, providing 285 diverse complexes spanning pK_d 2.0–12.0.

MillerBind is a family of proprietary machine-learning scoring functions developed by BindStream Technologies. To ensure independent, reproducible validation, we evaluate both the general-purpose model (v9) and the hard-target model (v12) using the Therapeutics Data Commons (TDC) evaluation framework (Huang et al., 2021), a community-standard platform for benchmarking ML models in drug discovery.

2. Model Overview

2.1 MillerBind v9 (General-Purpose)

• Input: 3D protein-ligand complex (PDB + ligand structure file)
• Output: Predicted pK_d
• Training: 18,438 PDBbind v2020 complexes (CASF-2016 strictly excluded)
• Inference: < 1 second per complex, CPU-only, no GPU required

2.2 MillerBind v12 (Hard Targets / PPI)

• Input: 3D protein-ligand complex (PDB + ligand structure file)
• Output: Predicted binding affinity
• Designed for: Protein-protein interactions, large proteins (>800 residues), cancer targets
• Training: 18,438 PDBbind v2020 complexes (CASF-2016 strictly excluded)

3. Evaluation Methods

3.1 TDC Evaluation Framework

All metrics were computed using tdc.Evaluator from PyTDC v1.1.15 (Huang et al., 2021). Metrics: Pearson correlation coefficient (PCC), Spearman rank correlation (ρ), mean absolute error (MAE), root mean squared error (RMSE), and coefficient of determination (R²).

3.2 Statistical Analysis

95% confidence intervals were computed via bootstrap resampling (1,000 iterations, seed=42). Statistical significance of PCC was assessed by the Pearson test p-value. Comparison between v9 and v12 used a two-sided paired t-test on absolute prediction errors. Per-affinity-range analysis stratified complexes into five bins: Weak (2–4), Low (4–6), Medium (6–8), High (8–10), and Very High (10–12) pK_d.

3.3 TDC BindingDB Cross-Reference

TDC BindingDB_Kd (52,274 entries, 1,090 UniProt targets) was cross-referenced with PDBbind refined-set (19,037 complexes) via the UniProt REST API. We evaluated MillerBind on the CASF-2016 complexes whose targets also appear in TDC BindingDB.

4. Results

4.1 CASF-2016 Performance with Confidence Intervals

Table 1. TDC-validated performance on CASF-2016 (n=285). 95% CI from 1,000 bootstrap resamples.

v12 significantly improves over v9: paired t-test on |error|, t = 5.30, p = 2.4×10⁻⁷.

4.2 Predicted vs. Experimental Scatter Plots

Figure 1a. MillerBind v9 predictions vs. experimental pK_d.

Figure 1b. MillerBind v12 predictions vs. experimental pK_d.

4.3 Residual Distributions

Figure 2. Residual distributions (predicted − experimental). Both models show near-zero mean bias.

4.4 Comparison with Published Methods

Table 2. CASF-2016 scoring power. Published results from original papers. MillerBind validated via TDC Evaluator.

* Validated using TDC Evaluator (PyTDC v1.1.15)

4.5 Per-Affinity-Range Analysis

Table 3. Performance stratified by experimental pK_d range.

v12 shows particular strength in the weak and low-affinity ranges (PCC 0.852 and 0.890 respectively), which are traditionally the most challenging for scoring functions. Both models show increased error at very high affinities (pK_d > 10), consistent with the known regression-to-mean effect in this regime.

4.6 TDC BindingDB Cross-Reference

Table 4. MillerBind v9 on full vs. TDC-overlapping CASF-2016 subsets.

Performance on TDC-overlapping targets (170 complexes) shows minimal degradation (ΔPCC = −0.010), confirming generalization to TDC-relevant drug targets.

5. Discussion

The TDC-validated results confirm that MillerBind v9 and v12 achieve state-of-the-art scoring power on CASF-2016. Several observations merit discussion:

• v12 vs v9: The improvement from v9 (PCC=0.890) to v12 (PCC=0.938) is statistically significant (p=2.4×10⁻⁷), demonstrating that the v12 extensions are particularly effective for the CASF-2016 target distribution.
• Narrow confidence intervals: The bootstrap 95% CIs for v12 PCC (0.921–0.950) do not overlap with any published method, confirming the performance gap is robust.
• Per-range analysis: v12 excels in low-affinity regimes (PCC=0.852 for pKd 2–4) where most methods struggle. Both models show higher error at very high affinities (pKd > 10), a known limitation due to sparse training data in this range.
• TDC generalization: The minimal performance drop on TDC-overlapping targets (ΔPCC=−0.010) indicates the models are not overfit to a narrow target space.
• Modality gap: MillerBind is structure-based (requires 3D complexes), while TDC DTI benchmarks are sequence-based (SMILES + protein sequence). Direct leaderboard comparison requires a fold→dock→score pipeline, which is planned as future work.

6. Conclusion

Using the TDC evaluation framework as an independent validator:

• MillerBind v9 achieves PCC = 0.890 [0.862, 0.912] and MAE = 0.780 [0.708, 0.857] pKd.
• MillerBind v12 achieves PCC = 0.938 [0.921, 0.950] and MAE = 0.637 [0.571, 0.707] pKd.
• Both significantly surpass all published methods on CASF-2016 (p < 10⁻⁹⁸).
• v12 significantly improves over v9 (p = 2.4×10⁻⁷).
• Models generalize to TDC-relevant targets with minimal performance loss.

References