Multi-objective generative AI for designing novel brain-targeting small molecules

TL;DR

This work tackles the challenge of designing brain-targeting small molecules that cross the BBB, act on a CNS receptor, and are safe by introducing a modular multi-objective generative AI pipeline. It trains graph neural predictors for BBB permeability, D2R bioactivity, and ADME-Tox safety on curated datasets and integrates them into a SyntheMol-based Monte Carlo Tree Search over the Enamine REAL space to generate 26,581 candidates. The approach yields high-diversity molecules with strong predicted BBB penetration and D2R affinity, validated in silico by docking affinities near that of risperidone and outperforming unguided baselines in joint objective scoring. This framework is extensible to other brain targets and, with experimental validation, could accelerate CNS drug discovery for challenging neurological disorders. $P(N) = (1/|N_{mols}|) sum_{i=1}^{|N_{mols}|} ( M_{BBB}(N^i_{mols}) + M_{D2R}(N^i_{mols}) + M_{Tox}(N^i_{mols}) ) / 3$.

Abstract

The strict selectivity of the blood-brain barrier (BBB) represents one of the most formidable challenges to successful central nervous system (CNS) drug delivery. Computational methods to generate BBB permeable drugs in silico may be valuable tools in the CNS drug design pipeline. However, in real-world applications, BBB penetration alone is insufficient; rather, after transiting the BBB, molecules must bind to a specific target or receptor in the brain and must also be safe and non-toxic. To discover small molecules that concurrently satisfy these constraints, we use multi-objective generative AI to synthesize drug-like BBB-permeable small molecules. Specifically, we computationally synthesize molecules with predicted binding affinity against dopamine receptor D2, the primary target for many clinically effective antipsychotic drugs. After training several graph neural network-based property predictors, we adapt SyntheMol (Swanson et al., 2024), a recently developed Monte Carlo Tree Search-based algorithm for antibiotic design, to perform a multi-objective guided traversal over an easily synthesizable molecular space. We design a library of 26,581 novel and diverse small molecules containing hits with high predicted BBB permeability and favorable predicted safety and toxicity profiles, and that could readily be synthesized for experimental validation in the wet lab. We also validate top scoring molecules with molecular docking simulation against the D2 receptor and demonstrate predicted binding affinity on par with risperidone, a clinically prescribed D2-targeting antipsychotic. In the future, the SyntheMol-based computational approach described here may enable the discovery of novel neurotherapeutics for currently intractable disorders of the CNS.

Figures (4)

• Figure 1: Results of SyntheMol-based guided Monte Carlo Tree Search for multi-objective small molecule design. Guided by predictors of BBB permeability, D2R binding, and toxicity, MCTS was performed over 29.6 billion compounds to generate a library of 26,581 molecules.
• Figure S1: Transport and barrier functions of the CNS epithelium. Physical, molecular, metabolic, and extravascular barriers of the BBB regulate access to the CNS milieu. Created using Biorender.com.
• Figure S2: Model architecture of property predictors. To predict blood-brain-barrier permeability and D2R bioactivity, two separate property predictors were trained. (A) First, the SMILES representations of each molecule in the training data were converted to molecular graphs; for example, the SMILES structure and molecular graph of moxalactam, a third-generation cephalosporin antibiotic in the B3DB database, are shown here. Next, we developed graph neural networks (GNNs) with three layers of message-passing, followed by a two-layer feed-forward neural network that combined the output of the GCN with a 200-element RDKit chemical descriptor to predict the likelihood of either blood-brain-barrier permeability or D2R bioactivity. The structured chemical descriptors are included based on the observation that properties like (B) molecular weight and (C) number of heavy atoms are correlated with blood-brain-barrier permeability, as measured by correlation with the numerical blood-brain-barrier concentration ratios for 1,058 of the 7,807 molecules in the B3DB dataset for which the data was available.
• Figure S3: Property predictor performance. To evaluate out-of-sample model performance while maximizing usage of the available training data, property predictor models were trained with 10-fold cross-validation with 30 epochs of training per fold and 80%-10%-10% split across training, validation, and test sets, respectively. The average prediction across all 10 folds was taken as the final likelihood score, and model performance on the independent test set was visualized in area under the receiver operating characteristic (AUROC) curves. Here, we highlight the performance of the first cross-validation fold for the (A) blood-brain-barrier permeability predictor and (B) the dopamine receptor D2 bioactivity predictor. We also show the performance across a set of randomly sampled folds for both the (C) BBB and (D) D2R predictors.