Plasma Protein Binding Prediction under Scaffold Split: Tree Models, ChemBERTa, Graph Networks, and Late Fusion
- 1University of North Carolina at Greensboro (UNC Greensboro
Abstract
From idea generation to testing, drug discovery is a lengthy and expensive process. Rather than chasing all of their molecule ideas, scientists prioritize certain molecules that are more likely to be successful, from synthesis to safety to therapeutic use. Machine learning and AI enable researchers to narrow their candidate choices by predicting important properties such as ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity). This study uses the Plasma Protein Binding (PPB) dataset from DeepChem’s MoleculeNet, with several graph neural networks (GNNs), ChemBERTa (a language model for SMILES strings of molecules), and fusions of these models. The performance of these neural models is compared against two standard tree-based models: Random Forest (RF) and XGBoost. Instead of random splitting of the dataset, this study evaluates the models based on scaffold splitting that separates molecules by the Bemis–Murcko scaffold and puts the entire group into only the training, validation, or test set. This study demonstrates that for the PPB dataset, GNNs and ChemBERTa models perform better than the tree models. On a mean basis, the fusion models perform better than the individual GNNs or ChemBERTa-only models. Random splitting on the RF model exhibits a better R2 score than on a scaffold-split dataset. Because under scaffold splitting the model predicts on unseen cores, the task becomes more difficult than under regular random splitting.
Related
- Guidelines to Evaluate Structure-based Methods for Protein-Ligand Binding Prediction
- Quantitative Protein–Ligand Binding Affinity Prediction from Quantum-Mechanical Scoring
- Excited-State Photophysics Improves Scaffold-Robust Phototoxicity Prediction and Reveals Compound-Specific State Effects
- Microstate-Aware Excited-State Descriptors Improve Scaffold-Robust Prediction of Drug Phototoxicity
- Scaffold-Controlled Evaluation of Molecular Representations for Antimalarial Activity Prediction
BibTeX
@misc{pradhan2026plasma,
title = {{Plasma Protein Binding Prediction under Scaffold Split: Tree Models, ChemBERTa, Graph Networks, and Late Fusion}},
author = {Pradyumna Kumar Pradhan},
howpublished = {ChemRxiv},
year = {2026},
month = sep,
doi = {10.26434/chemrxiv.15008530/v1},
url = {https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008530/v1},
note = {CC BY},
}
RIS
TY - UNPB
TI - Plasma Protein Binding Prediction under Scaffold Split: Tree Models, ChemBERTa, Graph Networks, and Late Fusion
AU - Pradyumna Kumar Pradhan
PY - 2026
DA - 2026/09/09
PB - ChemRxiv
DO - 10.26434/chemrxiv.15008530/v1
UR - https://chemrxiv.org/doi/full/10.26434/chemrxiv.15008530/v1
AB - From idea generation to testing, drug discovery is a lengthy and expensive process. Rather than chasing all of their molecule ideas, scientists prioritize certain molecules that are more likely to be successful, from synthesis to safety to therapeutic use. Machine learning and AI enable researchers to narrow their candidate choices by predicting important properties such as ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity). This study uses the Plasma Protein Binding (PPB) dataset from DeepChem’s MoleculeNet, with several graph neural networks (GNNs), ChemBERTa (a language model for SMILES strings of molecules), and fusions of these models. The performance of these neural models is compared against two standard tree-based models: Random Forest (RF) and XGBoost. Instead of random splitting of the dataset, this study evaluates the models based on scaffold splitting that separates molecules by the Bemis–Murcko scaffold and puts the entire group into only the training, validation, or test set. This study demonstrates that for the PPB dataset, GNNs and ChemBERTa models perform better than the tree models. On a mean basis, the fusion models perform better than the individual GNNs or ChemBERTa-only models. Random splitting on the RF model exhibits a better R2 score than on a scaffold-split dataset. Because under scaffold splitting the model predicts on unseen cores, the task becomes more difficult than under regular random splitting.
ER -