Tdc Admet Eval

---
name: tdc-admet-eval
description: Evaluates molecular property prediction across 22 ADMET tasks. It probes the model's ability to generalize across diverse chemical properties using standardized benchmark splits for both regression and classification. Use when the user wants to benchmark on TDC ADMET Group, or asks about evaluating this task. Reports Classification AUROC.
metadata:
  skill_kind: dataset_eval
  source_arxiv: 2304.12239
  bibtex_key: gao2023uniqsar
  confidence: high
---

# tdc-admet-eval

> Uni-QSAR: an Auto-ML Tool for Molecular Property Prediction — Gao et al. (2023) (arXiv:2304.12239, 2023)

## What this evaluates

Evaluates molecular property prediction across 22 ADMET tasks. It probes the model's ability to generalize across diverse chemical properties using standardized benchmark splits for both regression and classification.

## Datasets

- **TDC ADMET Group** — total ?; splits: test (-1)

## Metrics

- `Regression MAE` — range: other
  - Mean Absolute Error between predicted and true continuous values. Lower is better.
- `Spearman` — range: [-1, 1]
  - Spearman rank correlation coefficient between predictions and ground truth. Higher is better.
- `Classification AUROC` **(primary)** — range: [0, 1]
  - Area Under the Receiver Operating Characteristic curve. Higher is better.
- `AUPRC` — range: [0, 1]
  - Area Under the Precision-Recall Curve. Higher is better.

## Input / output format

**Input**: Molecular representations (1D SMILES, 2D graphs, or 3D conformers) with associated target property values.

**Output**: Predicted continuous values for regression tasks or class probabilities/labels for classification tasks.

## Scoring recipe

```python
def score_regression(y_pred, y_true):
    mae = mean(abs(y_pred - y_true))
    spearman = rank_correlation(y_pred, y_true)
    return mae, spearman

def score_classification(y_pred, y_true):
    auroc = roc_auc_score(y_true, y_pred)
    auprc = average_precision_score(y_true, y_pred)
    return auroc, auprc
```

## Common pitfalls

- Using non-standard data splits instead of the official TDC benchmark configuration.
- Failing to apply auto-target normalization for highly skewed target distributions, which degrades performance.
- Evaluating on internal splits only without external validation for generalization.

## Evidence (verbatim from paper)

> Table 1, 2 show the experiment results of our framework and competitive baselines. ... Regression MAE (lower is better ↓) ... Spearman (higher is better ↑) ... Classification AUROC(higher is better ↑) ... AUPRC(higher is better ↑).

## Citation

```bibtex
@misc{gao2023uniqsar,
  title={Uni-QSAR: an Auto-ML Tool for Molecular Property Prediction},
  author={Gao et al. (2023)},
  year={2023},
  note={arXiv:2304.12239}
}
```

- arXiv: 2304.12239