Hybrid-CatBoost: an empirical comparison of molecular representations for aqueous-solubility prediction on the TDC benchmark
- 1College of Life Science and Technology, Beijing University of Chemical Technology, North 3rd Ring, Beijing, China
Abstract
Aqueous-solubility prediction depends on both molecular representation and the definition of the target. We compared fingerprint, descriptor, hybrid, and graph-based representations on the TDC Solubility_AqSolDB benchmark using five official scaffold splits and a fixed test set of 1,997 records. Hybrid-CatBoost achieved a five-run test MAE of 0.749631 ± 0.016216 TDC logS units, compared with 0.764558 ± 0.009666 for descriptor-only CatBoost. However, the corresponding validation improvement was small (ΔMAE = −0.004932) and occurred in only two of five matched splits. Errors increased toward both target extremes and showed opposite-signed residual biases. These results indicate that descriptor-based models capture most of the predictive signal in this benchmark, while the added Morgan fingerprint provides only a modest and split-dependent benefit. Interpretation is limited by incomplete assay metadata and the use of a single fixed test set, which does not provide external validation.
Related
- Solubility prediction for ill-defined polymers: A critical comparison of Hansen parameters and COSMO-RS backfitting
- Developing a Quantum Physics-Inspired Machine Learning Model for Partition Coefficient Prediction and Comparing Its Performance with Empirical Descriptors and Molecular Graph-Based Models
- Task-adaptive multimodal molecular representations for structure-sensitive property prediction
- Transformer-based molecular fragment prediction using SMILES and DeepSMILES representations in a fragment-based drug discovery pipeline.
- DeMiR: Density-Matrix-Informed Representations for Molecular Property Prediction
BibTeX
@misc{li2026hybridcatboost,
title = {{Hybrid-CatBoost: an empirical comparison of molecular representations for aqueous-solubility prediction on the TDC benchmark}},
author = {Ruming Li},
howpublished = {ChemRxiv},
year = {2026},
month = oct,
doi = {10.26434/chemrxiv.15009996/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009996/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Hybrid-CatBoost: an empirical comparison of molecular representations for aqueous-solubility prediction on the TDC benchmark
AU - Ruming Li
PY - 2026
DA - 2026/10/06
PB - ChemRxiv
DO - 10.26434/chemrxiv.15009996/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15009996/v1
AB - Aqueous-solubility prediction depends on both molecular representation and the definition of the target. We compared fingerprint, descriptor, hybrid, and graph-based representations on the TDC Solubility_AqSolDB benchmark using five official scaffold splits and a fixed test set of 1,997 records. Hybrid-CatBoost achieved a five-run test MAE of 0.749631 ± 0.016216 TDC logS units, compared with 0.764558 ± 0.009666 for descriptor-only CatBoost. However, the corresponding validation improvement was small (ΔMAE = −0.004932) and occurred in only two of five matched splits. Errors increased toward both target extremes and showed opposite-signed residual biases. These results indicate that descriptor-based models capture most of the predictive signal in this benchmark, while the added Morgan fingerprint provides only a modest and split-dependent benefit. Interpretation is limited by incomplete assay metadata and the use of a single fixed test set, which does not provide external validation.
ER -