Molecular Scaffold Optimization Eval

---
name: molecular-scaffold-optimization-eval
description: Evaluates sample-efficient molecular scaffold optimization by testing how well a model can modify a given molecular scaffold to improve target properties (e.g., drug-likeness, docking scores) while preserving structural similarity, under a strict budget of oracle evaluations. Use when the user wants to benchmark on Gao et al. (2022) sample-efficiency benchmark, or asks about evaluating this task. Reports Top-10 average score.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2411.01423
  bibtex_key: boyar2024clasmo
  confidence: high
---

# molecular-scaffold-optimization-eval

> Conditional Latent Space Molecular Scaffold Optimization for Accelerated Molecular Design — Boyar et al. (2024) (arXiv:2411.01423, 2024)

## What this evaluates

Evaluates sample-efficient molecular scaffold optimization by testing how well a model can modify a given molecular scaffold to improve target properties (e.g., drug-likeness, docking scores) while preserving structural similarity, under a strict budget of oracle evaluations.

## Datasets

- **Gao et al. (2022) sample-efficiency benchmark** — total 20; splits: test (100)

## Metrics

- `Top-10 average score` **(primary)** — range: oracle-specific
  - Average of the top 10 oracle function values among the generated molecules after the optimization budget is exhausted.

## Input / output format

**Input**: Input molecular scaffold (SELFIES string) and a target oracle function to optimize, optionally conditioned on a 2D atomic environment embedding.

**Output**: Optimized molecular scaffold (SELFIES string) that improves the target oracle score while satisfying a Dice similarity threshold constraint (0.50, 0.25, or 0) relative to the input.

## Scoring recipe

```python
def top10_avg_score(predictions, oracle_fn):
    scores = [oracle_fn(mol) for mol in predictions]
    top_10 = sorted(scores, reverse=True)[:10]
    return sum(top_10) / len(top_10)
```

## Common pitfalls

- Using the full 10,000 oracle evaluation budget from the original Gao et al. (2022) benchmark instead of the paper's strict 100-evaluation budget.
- Ignoring the hard Dice similarity constraints (τ=0.50, 0.25, 0) that filter out molecules failing to preserve scaffold structure.
- Confusing the training dataset (QM9 BRICS fragments) with the evaluation dataset (100 scaffolds sampled from QM9 and ZINC250K).

## Evidence (verbatim from paper)

> Experiments were conducted with a fixed budget of 100 oracle evaluations per seed across 10 seeds, ensuring consistent comparisons. The choice of 100 oracle iterations (in contrast to the 10,000 iterations used in Gao et al. (2022)) reflects our focus on developing a methodology tailored for resource-constrained scenarios, which are more representative of real-world applications. ... Overall, under high constraints, CLaSMO proves to be the best approach, providing the highest Top-10 average scores in 13 out of 20 property optimization tasks.

## Citation

```bibtex
@misc{boyar2024clasmo,
  title={Conditional Latent Space Molecular Scaffold Optimization for Accelerated Molecular Design},
  author={Boyar et al. (2024)},
  year={2024},
  note={arXiv:2411.01423}
}
```

- arXiv: 2411.01423