Benchmarking Binding-Site Fingerprinting — DB5.5 Validation of the AffiniBind Pipeline

Motivation

Computational binder design starts with one question: where on the target surface should we aim the generative engines?

Get this wrong, and every downstream design — no matter how sophisticated — will target the wrong interface. Get it right, and the generator starts with a biological pathway that nature already uses.

The AffiniBind surface fingerprinting stage answers this question using four complementary models plus a learned ensemble scorer. But how well does it work? And critically — is the ensemble worth the additional compute over the single best model?

To answer this, we ran the full fingerprinting pipeline against the gold-standard Docking Benchmark 5.5 (DB5.5) — 253 hand-curated protein-protein complexes with known co-crystal structures — and measured whether the pipeline correctly identifies the biologically relevant binding interface.

---

The Benchmark: DB5.5

The Docking Benchmark 5.5, curated by the ZLAB group, is the standard evaluation set for protein-protein docking and interface prediction. It spans five levels of difficulty (rigid-body to highly flexible) and nine functional categories:

Ground truth interface residues were defined by the PDBePISA/CAPRI standard: any receptor residue with a heavy atom within 5.0 Å of a ligand atom was labeled as interface. Of the 253 complexes, 223 were successfully processed through the full pipeline (30 were excluded due to missing PDBs or pipeline failures).

---

The Models Evaluated

Four individual models were benchmarked, plus the ensemble:

Individual Models

Ensemble

AffiniBind learned scoring model — operating on 24 features: the 9 base model scores (4 DL + electrostatics + SASA + residue depth + neighbor count) plus 15 multiplicative interaction terms that capture cross-model agreement patterns.

---

Primary Result: Patch Hit Rate

The central metric is patch hit rate: does the pipeline's top-ranked binding patch correspond to the known biological interface?

We measured this across four ranking modes. Here is the performance on all 223 complexes:

And on the 42 held-out test complexes (never seen during training):

Interpretation: The ensemble provides a 4.5 percentage point improvement in Top-1 hit rate over ScanNet alone (87.0% vs 82.5%) — and a 23.8 point improvement over the SASA baseline (63.2%). At Top-3, the ensemble reaches 96.4%, meaning the correct binding patch is almost always in the top three.

---

Per-Category Performance

Not all protein complexes are equally predictable. The benchmark stratifies results by functional category:

AffiniBind matches or outperforms ScanNet in every category except OR, where they are within 4.4 percentage points. The hardest categories — Enzyme-Substrate (69.2%) and Other-Receptor (73.9%) — represent transient, low-affinity interfaces where the binding patches are often smaller and less geometrically defined. The easiest categories — Antigen-Single (100%) and Antibody-Antigen (97.2%) — represent stable, high-affinity interfaces with strong geometric and electrostatic signatures.

---

Residue-Level Classification

Beyond patch-level accuracy, the ensemble was evaluated on per-residue classification: for each of 72,463 residues across 223 complexes (5,998 interface, 8.3%), how well can the model discriminate interface from non-interface?

On the held-out test set (42 complexes, not seen during training), the learned scoring model achieves 0.853 AUC, confirming that the ensemble generalizes beyond its training data.

---

Feature Importance: What Drives the Ensemble?

The top features by gain:

---

Hyperparameter Robustness

To assess stability, we ran a 25-configuration hyperparameter ablation sweep varying tree depth, learning rate, subsampling, and regularization. Every configuration achieved test AUC between 0.834 and 0.853 — a range of just 0.02 — confirming that the ensemble's performance is not sensitive to hyperparameter choice and that the feature set (not tuning) drives the result.

---

What This Means for the Pipeline

1. The fingerprinting stage works reliably

With a 96.4% Top-3 patch hit rate, the pipeline almost never misses the correct binding interface. In practice, the top 3–5 patches are enumerated as generator inputs, so the effective coverage for a real campaign is near 100%.

2. The ensemble is worth it

The 4.5 percentage point improvement over ScanNet-alone translates to approximately one fewer missed binding interface in every 22 targets. For campaigns against difficult targets — where the margin between a successful and failed campaign is thin — this improvement is material.

3. Hard targets need attention

Enzyme-Substrate (69.2%) and Other-Receptor (73.9%) are the weakest categories. For targets in these classes, we should consider: (a) lowering the patch rank cutoff to include more candidates; (b) incorporating additional structural information (e.g., conservation, experimental mutagenesis data) as priors; and (c) flagging the increased risk in the feasibility review.

4. Deployment is complete

The trained model has been deployed to the production pipeline. All new campaigns automatically use this model for residue-level scoring and patch ranking.

---

Data and Reproducibility

The full benchmark dataset, training/test splits, hyperparameter configurations, and evaluation code are available in the AffiniBind repository under experiments/db5/ and runs/benchmarks/db5/. The DB5.5 dataset is publicly available from the ZLAB benchmark archive. All figures in this article were generated by the 10_report.py script and are reproducible with a single command.

---

Explore the AffiniBind platform architecture to see how fingerprinting feeds into patch identification, binder generation, and developability ranking. View Platform

See the full pipeline evaluation metrics on real campaigns. View Open Targets