Qualitative and Quantitative Prediction of Blood-Brain Barrier Permeability of Chemicals via an SE(3)-Transformer-Based Model with LLM-Assisted Explanation.
TL;DR
A multifeature fusion framework for identifying and quantifying the BBB permeability to neurotoxic chemicals, which leverages an SE(3)-Transformer to encode the three-dimensional molecular geometry and a deep neural network to encode collision cross-section and fingerprint.
Abstract
The blood-brain barrier (BBB) regulates the entry of chemicals into the central nervous system, and contaminants with high permeability can accumulate in the brain, posing neurotoxicity risks. Given the urgent need for high-throughput neurotoxicity assessment, this study proposes a multifeature fusion framework for identifying and quantifying the BBB permeability to neurotoxic chemicals. BBBProfiler leverages an SE(3)-Transformer to encode the three-dimensional molecular geometry and a deep neural network to encode collision cross-section (CCS) and fingerprint, followed by adaptive integration of modality representations through gating weights. Under the scaffold-split evaluation, BBBProfiler achieves high performance with an area under the receiver operating characteristic curve (AUC) of 93.72% and a recall of 89.19% for classification, and a coefficient of determination (R2) of 0.72 for regression. Its predictive performance was validated using a BBB bioassay (11 chemicals) and an external data set (2023 chemicals), in which the model achieved 94.42% AUC. Substructure mask explanation analysis revealed CCS as an influential feature that correlated well with BBB permeability to chemicals (R2 = 0.84). Model predictions were supplemented with natural-language explanations generated from a large language model. BBBProfiler is deployed on a publicly accessible platform (https://www.ai4environ.cn/BBBProfiler) to facilitate the development of new approach methodologies for neurotoxicity evaluation.