TITAN-BBB: Predicting BBB Permeability using Multi-Modal Deep-Learning Models

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Traditional experimental assays for blood-brain barrier (BBB) permeability are resource-intensive and slow, hindering early-stage drug discovery.
• Existing machine learning models show promise but have not fully integrated diverse data types like chemical descriptors and deep learning embeddings.

Purpose of the Study:

• To introduce TITAN-BBB, a novel multi-modal deep learning architecture for predicting BBB permeability.
• To create the largest aggregated BBB permeability dataset for robust model training and evaluation.
• To assess the performance of TITAN-BBB against state-of-the-art methods in both classification and regression.

Main Methods:

• Developed a multi-modal deep learning architecture (TITAN-BBB) integrating tabular, image, and text-based features using attention mechanisms.
• Aggregated data from multiple literature sources to construct the largest publicly available BBB permeability dataset.
• Evaluated the model on both classification and regression tasks for BBB permeability prediction.

Main Results:

• TITAN-BBB achieved 86.5% balanced accuracy in classification, outperforming state-of-the-art by 3.1%.
• TITAN-BBB achieved a mean absolute error of 0.436 in regression, reducing error by 20% compared to existing models.
• The model demonstrated superior performance by effectively combining deep and domain-specific representations.

Conclusions:

• TITAN-BBB represents a significant advancement in computational prediction of BBB permeability.
• The integration of multi-modal data and attention mechanisms enhances predictive accuracy.
• This approach accelerates drug discovery by providing a faster, more efficient alternative to experimental assays.