Moleculenet Property Prediction Eval

moleculenet-property-prediction-eval

Local-Global Multimodal Contrastive Learning for Molecular Property Prediction — Liu et al. (2026) (arXiv:2601.22610, 2026)

What this evaluates

Evaluates the ability of multimodal molecular representation models to predict diverse physicochemical and biological properties from graph, text, and fingerprint inputs. It probes both classification (binary/multi-label activity prediction) and regression (continuous property estimation) capabilities across standardized chemical benchmarks.

Datasets

• MoleculeNet — total ?; splits: train (-1), val (-1), test (-1); repo https://github.com/lhb0189/LGM-CL

Metrics

• ROC-AUC (primary) — range: [0, 1]
  • Area under the Receiver Operating Characteristic curve, measuring the trade-off between true positive rate and false positive rate across classification thresholds.
• RMSE (primary) — range: other
  • Root Mean Squared Error, calculated as the square root of the average of squared differences between predicted and actual continuous values.

Input / output format

Input: Molecular graph structures (nodes/edges), SMILES strings, LLM-enhanced textual descriptions, and molecular fingerprint features.

Output: Predicted class labels (for classification) or continuous property values (for regression).

Scoring recipe

def compute_metric(y_true, y_pred, task_type):
    if task_type == 'classification':
        return roc_auc_score(y_true, y_pred)
    else:
        return np.sqrt(mean_squared_error(y_true, y_pred))

Common pitfalls

• The paper primarily reports results using random splitting (0.6/0.2/0.2); scaffold splitting results are only in the supplementary and should not be conflated with main results.
• Metrics are task-dependent: ROC-AUC is used for all 7 classification datasets, while RMSE is used for all 3 regression datasets. Mixing them up during comparison leads to incorrect conclusions.
• Performance must be averaged over three independent runs with different random seeds, as explicitly stated in the protocol.

Evidence (verbatim from paper)

Regarding data splitting, we follow the random splitting protocol adopted in S-CGIB [lee2025pre], using a split ratio of $0.6/0.2/0.2$ for training, validation, and testing, and report the average performance over three runs with different random seeds. Performance is measured using ROC-AUC for classification tasks and RMSE for regression tasks.

Citation

@misc{liu2026localglobalmultimodal,
  title={Local-Global Multimodal Contrastive Learning for Molecular Property Prediction},
  author={Liu et al. (2026)},
  year={2026},
  note={arXiv:2601.22610}
}

• arXiv: 2601.22610