# Mol Air Goal Directed Generation Eval

> Evaluates a reinforcement learning model's ability to generate molecules that optimize specific target chemical or biological properties. It probes the model's exploration-exploitation balance in navigating chemical space to find high-scoring structures for penalized LogP, drug-likeness (QED), structural similarity, and kinase inhibition targets. Use when the user wants to benchmark on Mol-AIR Goal-Directed Generation Tasks, or asks about evaluating this task. Reports Best Property Score.

- Skill: `qhjqhj00/mol-air-goal-directed-generation-eval` (Agent Skill)
- Install (CLI): `npx skillmds add qhjqhj00/mol-air-goal-directed-generation-eval`
- Raw SKILL.md: https://api.skillmd.com/api/skills/qhjqhj00/mol-air-goal-directed-generation-eval/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: qhjqhj00 (https://skillmd.com/u/qhjqhj00)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/qhjqhj00/mol-air-goal-directed-generation-eval

---


# mol-air-goal-directed-generation-eval

> Mol-AIR: Molecular Reinforcement Learning with Adaptive Intrinsic Rewards for Goal-directed Molecular Generation — Park et al. (2024) (arXiv:2403.20109, 2024)

## What this evaluates

Evaluates a reinforcement learning model's ability to generate molecules that optimize specific target chemical or biological properties. It probes the model's exploration-exploitation balance in navigating chemical space to find high-scoring structures for penalized LogP, drug-likeness (QED), structural similarity, and kinase inhibition targets.

## Datasets

- **Mol-AIR Goal-Directed Generation Tasks** — total ?; splits: test (-1); repo https://github.com/DevSlem/Mol-AIR

## Metrics

- `Best Property Score` **(primary)** — range: other
  - The maximum value of the target property score achieved across 3,000 training iterations. Properties include: pLogP (penalized LogP × 0.1), QED (drug-likeness, [0,1)), Tanimoto similarity to celecoxib ([0,1]), GSK3B inhibition score ([0,1]), JNK3 inhibition score ([0,1]), and GSK3B+JNK3 (arithmetic mean of the two inhibition scores). Scores are computed using RDKit, Guacamol scripts, or TDC oracles.

## Input / output format

**Input**: Initial molecular state or empty sequence; model generates molecules step-by-step using SELFIES representation.

**Output**: SELFIES string representing a complete molecular structure.

## Scoring recipe

```python
best_score = -inf
for step in range(3000):
    molecule = model.generate()
    score = compute_property_score(molecule, target_property)
    if score > best_score:
        best_score = score
return best_score
```

## Common pitfalls

- The reported score is the maximum found over 3,000 iterations, not the average or final step score.
- pLogP scores are normalized by multiplying by 0.1 before comparison.
- GSK3B and JNK3 scores rely on the TDC library oracles, not independent calculations.

## Evidence (verbatim from paper)

> For the calculation of pLogP, QED, and similarity scores, we used scripts provided by existing research, and for GSK3B and JNK3 scores, we utilized the oracle provided by the Therapeutics Data Commons (TDC) library. All experimental results were evaluated based on 3,000 iterations of training. Table 1 shows that the proposed method, Mol-AIR, is superior to the baseline approaches, outperforming them in finding the best scoring molecules for all six tasks.

## Citation

```bibtex
@misc{park2024molair,
  title={Mol-AIR: Molecular Reinforcement Learning with Adaptive Intrinsic Rewards for Goal-directed Molecular Generation},
  author={Park et al. (2024)},
  year={2024},
  note={arXiv:2403.20109}
}
```

- arXiv: 2403.20109