Predicting blood–brain barrier permeability of molecules with a large language model and machine learning

Eddie Huang, Jai‐Sing Yang, Ken Ying-Kai Liao, Warren C. W. Tseng, Chien-Yu Lee, Michelle Gill, Colin B. Compas, Simon See, Fuu‐Jen TsaiScientific 2024 · DOI

Abstract

Predicting the blood-brain barrier (BBB) permeability of small-molecule compounds is a critical bottleneck in central nervous system (CNS) drug discovery. This study presents a novel artificial intelligence platform that leverages the transformer-based Large Language Model (LLM) MegaMolBART and an XGBoost classifier to predict permeability from chemical structures. By treating Simplified Molecular Input Line Entry System (SMILES) strings as a chemical language, the model achieved an Area Under the Receiver Operating Characteristic (AUC) curve of 0.88 on a held-out test dataset. The computational results were validated in vitro using 3D human BBB spheroids and Liquid Chromatography-Tandem Mass Spectrometry (LC–MS/MS). The model correctly identified 21 permeable and 5 impermeable compounds, demonstrating that transformer-based architectures can significantly accelerate the identification of brain-penetrating therapeutics.

Introduction

The blood–brain barrier (BBB) is a highly selective semipermeable border that protects the brain from harmful substances while simultaneously blocking approximately 98% of potential small-molecule drugs. Traditional in silico prediction methods often rely on manual feature engineering of physicochemical properties, which can be computationally expensive and may fail to capture complex molecular geometries. This research explores a deep learning approach that bypasses manual descriptors by learning molecular representations directly from chemical "grammar."

Methods

In Silico Pipeline

The researchers utilized MegaMolBART, a transformer-based model pre-trained on millions of molecules from the ZINC-15 database. MegaMolBART serves as a SMILES encoder, generating high-dimensional embeddings that capture long-range dependencies within molecular structures. These embeddings were then used as inputs for an XGBoost classifier. To establish a baseline, the team compared this approach against traditional Morgan (Circular) fingerprints, which are the industry standard for molecular similarity.

In Vitro Validation

To confirm the accuracy of the AI predictions, the study employed 3D human BBB spheroids composed of human brain microvascular endothelial cells, brain vascular pericytes, and astrocytes. Compounds including Temozolomide (TMZ), Ferulic acid, and various natural products from the Natural Products Research Laboratories (NPRL) at China Medical University Hospital (CMUH) were tested. Permeability was verified using LC–MS/MS to detect the presence of compounds within the spheroid cells.

Results and Discussion

The integration of LLM embeddings with machine learning classifiers yielded superior results compared to traditional methods. The MegaMolBART variants (NGC and TWS embeddings) consistently outperformed Morgan fingerprints across multiple metrics, including AUC and Accuracy, particularly when tested on the specialized CMUH-NPRL dataset.

Comparison of predictive performance (AUC and Accuracy) between Morgan Fingerprints and MegaMolBART embedding variants.

Analysis using t-SNE visualization revealed that natural products and general drug databases occupy distinct chemical spaces. The researchers found that training the model on a mixture of these datasets was vital for generalization. While the model showed high predictive power, it exhibited a tendency to overfit on smaller, specialized datasets, suggesting that future work should focus on expanding quantitative LogBB data to improve regression accuracy.

Experimental Validation

The biological validation phase provided strong evidence for the model's reliability. Temozolomide (TMZ), a known brain-penetrating drug, and NPRL309 were predicted as permeable (Pred score = 1) and showed clear peaks in the BBB cell mass spectrometry analysis. Conversely, Ferulic acid was predicted as impermeable and showed no significant penetration in the 3D spheroid model.

LC-MS/MS chromatograms showing BBB permeability validation for Ferulic acid (A), Temozolomide (B), and NPRL309 (C).

In total, the model achieved a 100% success rate in this validation subset, correctly identifying all 21 permeable and 5 impermeable compounds tested in vitro.

Conclusion

This study demonstrates that transformer-based models like MegaMolBART, which leverage the "language" of chemistry, offer a powerful and scalable alternative to traditional molecular modeling. By accurately predicting BBB permeability, these AI tools have the potential to reduce the time and cost associated with developing novel treatments for disorders such as Alzheimer’s, Parkinson’s, and brain tumors.

Summary updated

Abstract

Predicting the blood-brain barrier (BBB) permeability of small-molecule compounds using a novel artificial intelligence platform is necessary for drug discovery. Machine learning and a large language model on artificial intelligence (AI) tools improve the accuracy and shorten the time for new drug development. The primary goal of this research is to develop artificial intelligence (AI) computing models and novel deep learning architectures capable of predicting whether molecules can permeate the human blood-brain barrier (BBB). The in silico (computational) and in vitro (experimental) results were validated by the Natural Products Research Laboratories (NPRL) at China Medical University Hospital (CMUH). The transformer-based MegaMolBART was used as the simplified molecular input line entry system (SMILES) encoder with an XGBoost classifier as an in silico method to check if a molecule could cross through the BBB. We used Morgan or Circular fingerprints to apply the Morgan algorithm to a set of atomic invariants as a baseline encoder also with an XGBoost classifier to compare the results. BBB permeability was assessed in vitro using three-dimensional (3D) human BBB spheroids (human brain microvascular endothelial cells, brain vascular pericytes, and astrocytes). Using multiple BBB databases, the results of the final in silico transformer and XGBoost model achieved an area under the receiver operating characteristic curve of 0.88 on the held-out test dataset. Temozolomide (TMZ) and 21 randomly selected BBB permeable compounds (Pred scores = 1, indicating BBB-permeable) from the NPRL penetrated human BBB spheroid cells. No evidence suggests that ferulic acid or five BBB-impermeable compounds (Pred scores < 1.29423E-05, which designate compounds that pass through the human BBB) can pass through the spheroid cells of the BBB. Our validation of in vitro experiments indicated that the in silico prediction of small-molecule permeation in the BBB model is accurate. Transformer-based models like MegaMolBART, leveraging the SMILES representations of molecules, show great promise for applications in new drug discovery. These models have the potential to accelerate the development of novel targeted treatments for disorders of the central nervous system.
BibTeX
@inproceedings{huang2024predicting,
  author = {Eddie Huang and Jai‐Sing Yang and Ken Ying-Kai Liao and Warren C. W. Tseng and Chien-Yu Lee and Michelle Gill and Colin B. Compas and Simon See and Fuu‐Jen Tsai},
  title = {{Predicting blood–brain barrier permeability of molecules with a large language model and machine learning}},
  booktitle = {Scientific},
  year = {2024},
  doi = {10.1038/s41598-024-66897-y},
}
Figure