Fast, lightweight BBB permeability predictions for molecular design
- 1Department of Chemistry, University of Warwick Coventry CV4 7AL West Midlands UK
Abstract
The blood–brain barrier (BBB) is a selectively permeable border that regulates the transport of chemical compounds into and out of the central nervous system (CNS). To discover medicinal drugs that can target the CNS, we must be able to efficiently screen pools of potential candidate CNS-targeting drugs to identify those that can permeate the BBB. However, drugs that readily permeate the BBB are comparatively rare, thus CNS drug discovery typically requires screening very large candidate libraries. In the last two decades, a diverse portfolio of machine learning models has been developed to predict the permeability of the BBB. However, many of the most accurate models reported to date do not scale well to large screenings, because they depend on high-cost descriptor pipelines and/or computationally intensive model architectures. In this work, we used a straightforward, computationally efficient set of descriptors in conjunction with a lightweight random forest model to perform a large-scale screening for BBB permeable drugs. Specifically, we predicted the propensity to cross the BBB for roughly 117 million compounds from PubChem. The screen ran on CPU-only hardware (64 cores) and completed in less than 800 CPU hours; the main resource constraint was memory, with ∼4 TB RAM required for data handling. From these predictions, we prioritised a small set of top-ranked compounds and used well-tempered metadynamics with lipid-bilayer proxies to assess whether their membrane-crossing barriers were consistent with the model-predicted permeability classes. We hope that similar workflows can be adopted in the near term to improve the throughput of BBB-permeability prioritisation in drug discovery.
Related
- CIPER: Interpretable Permeability Predictions Guide Cyclic Peptide Editing
- A decoupled alignment kernel for peptide membrane permeability predictions
- Embedded Morgan Fingerprints for more efficient molecular property predictions with machine learning
- Crystal structure, Hirshfeld surface analysis, computational study, molecular docking predictions and biological evaluation of a new Schiff base (E)-2-((2,4-dichlorobenzylidene)amino)aniline
- Lightweight design and post-impact attitude stability of extra-high grade guardrail based on 700L high-strength steel
BibTeX
@article{jia2026fast,
title = {{Fast, lightweight BBB permeability predictions for molecular design}},
author = {Hengjian Jia and Gabriele C. Sosso},
journal = {Molecular Systems Design \& Engineering},
year = {2026},
month = nov,
doi = {10.1039/d6me00041j},
url = {https://doi.org/10.1039/d6me00041j},
note = {CC BY 3.0},
}
RIS
TY - JOUR
TI - Fast, lightweight BBB permeability predictions for molecular design
AU - Hengjian Jia
AU - Gabriele C. Sosso
PY - 2026
DA - 2026/11/03
JO - Molecular Systems Design & Engineering
DO - 10.1039/d6me00041j
UR - https://doi.org/10.1039/d6me00041j
AB - The blood–brain barrier (BBB) is a selectively permeable border that regulates the transport of chemical compounds into and out of the central nervous system (CNS). To discover medicinal drugs that can target the CNS, we must be able to efficiently screen pools of potential candidate CNS-targeting drugs to identify those that can permeate the BBB. However, drugs that readily permeate the BBB are comparatively rare, thus CNS drug discovery typically requires screening very large candidate libraries. In the last two decades, a diverse portfolio of machine learning models has been developed to predict the permeability of the BBB. However, many of the most accurate models reported to date do not scale well to large screenings, because they depend on high-cost descriptor pipelines and/or computationally intensive model architectures. In this work, we used a straightforward, computationally efficient set of descriptors in conjunction with a lightweight random forest model to perform a large-scale screening for BBB permeable drugs. Specifically, we predicted the propensity to cross the BBB for roughly 117 million compounds from PubChem. The screen ran on CPU-only hardware (64 cores) and completed in less than 800 CPU hours; the main resource constraint was memory, with ∼4 TB RAM required for data handling. From these predictions, we prioritised a small set of top-ranked compounds and used well-tempered metadynamics with lipid-bilayer proxies to assess whether their membrane-crossing barriers were consistent with the model-predicted permeability classes. We hope that similar workflows can be adopted in the near term to improve the throughput of BBB-permeability prioritisation in drug discovery.
SN - 2058-9689
ER -