Pdbbind Lba Eval

---
name: pdbbind-lba-eval
description: Predicts the binding affinity between a protein pocket and a ligand from 3D structural data. It probes the model's ability to quantify molecular interaction strength and generalize across protein sequence identities. Use when the user wants to benchmark on PDBBind, or asks about evaluating this task. Reports RMSE.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2311.16160
  bibtex_key: feng2023proteinligand
  confidence: high
---

# pdbbind-lba-eval

> Protein-ligand binding representation learning from fine-grained interactions — Feng et al. (2023) (arXiv:2311.16160, 2023)

## What this evaluates

Predicts the binding affinity between a protein pocket and a ligand from 3D structural data. It probes the model's ability to quantify molecular interaction strength and generalize across protein sequence identities.

## Datasets

- **PDBBind** — total ?; splits: LBA 30% (-1), LBA 60% (-1)

## Metrics

- `RMSE` **(primary)** — range: other
  - Root Mean Square Error: sqrt(mean((y_pred - y_true)^2)). Lower values indicate better affinity prediction.
- `Pearson correlation` — range: [-1, 1]
  - Pearson correlation coefficient between predicted and true affinity values.
- `Spearman correlation` — range: [-1, 1]
  - Spearman rank correlation coefficient between predicted and true affinity values.

## Input / output format

**Input**: 3D coordinates of protein pocket residues and ligand atoms, processed into CLS token embeddings.

**Output**: Scalar binding affinity value (pKd/pKi).

## Scoring recipe

```python
rmse = sqrt(mean((pred - true)**2))
pearson = corr(pred, true)
spearman = rank_corr(pred, true)
```

## Common pitfalls

- Splits are based on protein sequence identity (30% vs 60%), not random, which significantly impacts generalization evaluation.
- Results are averaged over 3 runs with different random seeds to ensure robustness.

## Evidence (verbatim from paper)

> The dataset is partitioned using a protein sequence identity threshold, resulting in two distinct splits: LBA 30% (with a protein sequence identity threshold of 30%) and LBA 60% (with a protein sequence identity threshold of 60%). We employ RMSE (Root Mean Square Error), Pearson correlation coefficient, and Spearman correlation coefficient, to evaluate BindNet. To ensure robustness of evaluation, we conduct three runs with different random seeds and report the mean values for the aforementioned metrics.

## Citation

```bibtex
@misc{feng2023proteinligand,
  title={Protein-ligand binding representation learning from fine-grained interactions},
  author={Feng et al. (2023)},
  year={2023},
  note={arXiv:2311.16160}
}
```

- arXiv: 2311.16160