Optimizing blood-brain barrier permeability in KRAS inhibitors: A structure-constrained molecular generation approach

Abstract

Kirsten rat sarcoma viral oncogene homolog (KRAS) protein inhibitors are a promising class of therapeutics, but research on molecules that effectively penetrate the blood-brain barrier (BBB) remains limited, which is crucial for treating central nervous system (CNS) malignancies. Although molecular generation models have recently advanced drug discovery, they often overlook the complexity of biological and chemical factors, leaving room for improvement. In this study, we present a structure-constrained molecular generation workflow designed to optimize lead compounds for both drug efficacy and drug absorption properties. Our approach utilizes a variational autoencoder (VAE) generative model integrated with reinforcement learning for multi-objective optimization. This method specifically aims to enhance BBB permeability (BBBp) while maintaining high-affinity substructures of KRAS inhibitors. To support this, we incorporate a specialized KRAS BBB predictor based on active learning and an affinity predictor employing comparative learning models. Additionally, we introduce two novel metrics, the knowledge-integrated reproduction score (KIRS) and the composite diversity score (CDS), to assess structural performance and biological relevance. Retrospective validation with KRAS inhibitors, AMG510 and MRTX849, demonstrates the framework's effectiveness in optimizing BBBp and highlights its potential for real-world drug development applications. This study provides a robust framework for accelerating the structural enhancement of lead compounds, advancing the drug development process across diverse targets.

Keywords: Blood-brain barrier permeability; Deep learning; Drug design; Generation models; KRAS inhibitors; Molecular optimization.

Graphical abstract

Fig. 1

The structure-constrained molecular generation workflow for lead optimization. (A) Pretraining strategy. (B) Overall framework of the generation model. (C) Dataset for pretraining. Lead: lead molecules; Opt: optimized molecules; Con: control molecules; BBBp: blood-brain barrier (BBB) permeability; QED: quantitative estimate of drug-likeness; logD: logarithm of the distribution coefficient.

Fig. 2

Performance of blood-brain barrier (BBB) permeability predictor with active leaning. (A) t-distributed stochastic neighbor embedding (t-SNE) visualization of molecular structure in BBB and Kirsten rat sarcoma viral oncogene homolog (KRAS) dataset. (B) The performance of the active learning predictor on Matthews correlation coefficient (MCC) under three sampling strategies: random sampling, entropy sampling, and margin sampling. (C) The result and entropy of the predictor on 17 KRAS inhibitors with and without active learning.

Fig. 3

Performance of Kirsten rat sarcoma viral oncogene homolog (KRAS) affinity predictor. Scatter plot of fitting on (A) the training set and (B) the test set.

Fig. 4

Performance of baseline models and our model on each evaluation metrics. (A) Comparisons of four baseline models with our model in diversity, novelty, similarity, and validity. (B) Comparisons of four baseline models with our model in three property improvements. (C) Probability density distribution of Tanimoto similarity between the molecules generated by four baseline models and our model compared to the input molecules. MolGPT: molecular generative pre-training model; MOLDR: molecular graph decomposition and reassembling; PGMG: pharmacophore-guided deep learning approach for bioactive molecule generation; BBBp: blood-brain barrier (BBB) permeability; QED: quantitative estimate of drug-likeness.

Fig. 5

Examples of model-generated molecular from (A) AMG510 and (B) MRTX849. Polar surface area (PSA) values are predicted by RDKit. BBBp: blood-brain barrier (BBB) permeability.

Fig. 6

Comparison of docking results between AMG510 and structure 2; MRTX849 and structure 5. (A) The overlay of AMG510 (white) and structure 2 (green) docking in Schrödinger (Protein Data Bank (PDB): 6OIM). (B) The two-dimensional (2D) ligand interaction diagram of AMG510. (C) The 2D ligand interaction diagram of structure 2. (D) The overlay of MRTX849 (white) and structure 5 (green) docking (PDB: 6UT0). (E) The 2D ligand interaction diagram of MRTX849. (F) The 2D ligand interaction diagram of structure 5. Interactions in 3D poses: yellow, hydrogen bonds; sky blue, π−π stacking; lime green, π−cation; and magenta, salt bridges. Interactions in 2D diagrams: purple, hydrogen bonds; green, π−π stacking; red, π−cation; and red to blue gradient, salt bridge.