Geometric Multi-color Message Passing Graph Neural Networks for Blood-brain Barrier Permeability Prediction
Trung Nguyen, Md Masud Rana, Farjana Tasnim Mukta, Chang-Guo Zhan, and Duc Duy Nguyen · 2025
Read Geometric Multi-color Message Passing Graph Neural Networks for Blood-brain Barrier Permeability Prediction by Trung Nguyen, Md Masud Rana, Farjana Tasnim Mukta, Chang-Guo Zhan, and Duc Duy Nguyen (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2507-18926.txt.
Introduction
Trung Nguyen and co-authors build in 2025 a graph neural network that forecasts whether a drug can cross the blood-brain barrier. That crossing decides the fate of brain drugs, yet most networks read only the molecule's bond map and miss its 3D shape. Their model adds atomic geometry and long-range links to the message passing.
Subgraphs colored by atom type carry the spatial and chemical context. On three benchmark sets, with strict scaffold splits, the model beat prior leaders at sorting compounds and at regressing permeability values. A knock-out study showed which atom-pair contacts drive the gains, including rare but telling chemical motifs.
Catalog record
| CREATOR | Trung Nguyen, Md Masud Rana, Farjana Tasnim Mukta, Chang-Guo Zhan, and Duc Duy Nguyen |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 43 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2507-18926
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Geometric Multi-color Message Passing Graph Neural Networks for Blood-brain Barrier Permeability Prediction Trung Nguyen, Md Masud Rana, Farjana Tasnim Mukta, Chang-Guo Zhan, and Duc Duy Nguyen Abstract Accurate prediction of blood-brain barrier permeability (BBBP) is essential for central nervous system (CNS) drug development. While graph neural networks (GNNs) have advanced molecular property prediction, they often rely on molecular topology and neglect the three-dimensional geometric information crucial for modeling transport mechanisms. This paper introduces the geometric multi-color message-passing graph neural network (GMC-MPNN), a novel framework that enhances standard message-passing architectures by explicitly incorporating atomic-level geometric features and long-range interactions. Our model constructs weighted colored subgraphs based on atom types to capture the spatial relationships and chemical context that govern BBB permeability. We evaluated GMC-MPNN on three benchmark datasets for both classification and regression tasks, using rigorous scaffold-based splitting to ensure a robust assessment of generalization. The results demonstrate that GMC-MPNN consistently outperforms existing state-of-the-art models, achieving superior performance in both classifying compounds as permeable/non-permeable (AUC-ROC of 0.9704 and 0.9685) and in regressing continuous permeability values (RMSE of 0.4609, Pearson correlation of 0.7759). An ablation study further quantified the impact of specific atom-pair interactions, revealing that the model’s predictive power derives from its ability to learn from both common and rare, but chemically significant, functional motifs. By integrating spatial geometry into the graph representation, GMC-MPNN sets a new performance benchmark and offers a more accurate and generalizable tool for drug discovery pipelines. Keywords: Blood-brain barrier, Drug discovery, Graph neural networks, Message passing neural networks, Geometric deep learning, Molecular property prediction, QSAR, Cheminformatics 1 Introduction The blood-brain barrier (BBB) is a highly selective interface that regulates the exchange of substances between the bloodstream and the brain, ensuring central nervous system (CNS) homeostasis by preventing the entry of harmful compounds while allowing essential nutrients and gases to pass through [ 1 , 2 ] . BBB permeability (BBBP) is tightly controlled by the endothelial cells that form the barrier, whose tight adhesion and adhesion junctions restrict the passage of large or hydrophilic molecules, allowing only small, lipophilic, or actively transported compounds to cross [ 2 ] . Predicting BBB permeability is critical for CNS drug development. Compounds with insufficient permeability of BBB can fail to achieve therapeutic concentrations in the brain [ 3 , 4 , 5 ] . BBBP data can be broadly categorized into numerical and categorical types. Numerical data quantify a compound’s permeability by comparing its concentration in the brain to its concentration in the blood, providing a continuous scale for measuring how effectively a substance crosses the BBB. In contrast, categorical data offer a binary classification, labeling compounds as either BBB + (permeable) or BBB - (non-permeable), which simplifies evaluation by indicating whether a substance can penetrate the barrier. Various in vivo and in vitro studies have been conducted to assess BBB permeability in drug molecules. In vivo, models provide physiologically relevant information but are resource-intensive, time-consuming, and ethically challenging [ 6 ] . In vitro models, including coculture and dynamic systems [ 7 ] , offer controlled environments for the testing of compounds but still involve significant costs and time. To address these limitations, in silico methods have gained prominence as efficient alternatives for BBBP prediction. These models can be broadly categorized into two main types: traditional quantitative structure-activity relationship (QSAR) models and machine learning-based models. QSAR models are among the earliest computational approaches used to predict BBB permeability. These models establish statistical correlations between molecular descriptors (e.g., lipophilicity, molecular weight, and hydrogen bonding capacity) and experimentally determined BBB permeability. Early QSAR models relied on multiple linear regression (MLR) and partial least squares (PLS) to identify molecular features associated with BBB permeability [ 8 , 9 , 10 ] . Some QSAR models use physicochemical rules, such as Lipinski’s Rule of Five or Veber’s Rule [ 11 , 12 ] , to estimate whether a compound is likely to penetrate the BBB. While these QSAR models provide interpretability, they often struggle with complex, nonlinear relationships in molecular data, leading to reduced predictive accuracy for diverse chemical compounds [ 9 ] . With the availability of large molecular datasets [ 13 , 14 ] , machine learning (ML) models have emerged as a more flexible alternative to QSAR-based methods. These models improve predictive performance by capturing non-linear relationships between molecular features and BBB permeability. Support vector machine (SVM), Random forest (RF), and gradient boosting trees (GBT) algorithms are particularly effective in handling high-dimensional molecular data, integrating a vector of molecular descriptors to enhance prediction accuracy [ 15 , 12 , 16 , 17 ] . Deep learning architectures have further advanced this field by automating feature extraction. For instance, convolutional neural networks (CNNs) are applied to 2D molecular images, learning spatial patterns associated with BBB permeability [ 18 ] . Whereas recurrent neural networks (RNNs) and long short-term memory (LSTM) networks process SMILES representations to model molecular properties in a sequential manner [ 19 ] . Additionally, hybrid approaches such as the DeePred-BBB Model [ 16 ] and LightBBB framework [ 20 ] incorporate various descriptor-based features, including physicochemical properties and molecular fingerprints, to enhance classification performance. In recent years, graph neural networks (GNNs) have emerged as a promising alternative by shifting from descriptor-based input representations to direct processing of molecular graphs [ 21 ] . Unlike traditional ML/DL models that rely on precomputed molecular fingerprints or an array of descriptors, GNNs represent molecules as graphs, where atoms are treated as nodes and chemical bonds as edges, enabling the direct extraction of structural and relational features [ 22 ] . Message passing neural networks (MPNNs) [ 23 ] have been widely applied in this context, as these architectures allow information to be exchanged between atomic neighbors through iterative message-passing steps. GSL-MPP [ 24 ] extends standard MPNNs by integrating graph structure learning on a molecular similarity graph, utilizing both intra- and inter-molecular relationships. CoMPT [ 25 ] introduces a node-edge message interaction mechanism within a transformer-based framework to explicitly capture edge information. It also employs a message diffusion strategy to prevent over-smoothing and enhance long-range dependencies. CD-MVGNN [ 26 ] introduces a multi-view GNN that equally considers atoms and bonds by using dual encoders and a cross-dependent message-passing scheme for bidirectional information exchange. While GNN-based models have significantly advanced BBBP prediction, their principal reliance on molecular topology constrains their ability to fully capture the structural and physicochemical properties governing blood-brain barrier permeability. These models often overlook the role of three-dimensional geometric information, which is critical for accurately modeling molecular interactions and transport mechanisms. Although some 3D-aware GNN architectures have been introduced for molecular property prediction, these models do not integrate long-range interactions between different atom types. Moreover, such architectures are highly complex, requiring multiple network layers and intricate integration mechanisms. In this study, we introduce a novel geometric multi-color message-passing graph neural network (GMC-MPGNN) that extends the capabilities of standard MPNNs by incorporating atomic-level geometric features. Unlike conventional GNNs that operate purely on molecular connectivity, our proposed model integrates spatial representations and captures multi-range molecular interactions through the construction of multi-color subgraphs. Beyond classification tasks in BBBP, we extend the application of GMC-MPGNN to regression tasks using the B3DB dataset