An Artificial Intelligence Framework for Optimal Drug Design
Grace Ramey (1), Santiago Vargas (1), Dinesh De Alwis (2), Anastassia N. Alexandrova (1), Joe Distefano III (1), and Peter Bloomingdale (2,3)
(1) UCLA, (2) Merck & Co., Inc. (previous affiliation), (3) Boehringer Ingelheim (current affiliation)
Objectives: To develop an end-to-end in silico framework for optimizing the exposure, safety, and efficacy of drugs. To provide a proof-of-concept of the platform through the generation of novel compounds optimized for brain exposure.
Methods: An artificial intelligence framework that integrates a de novo molecular design algorithm, quantitative structure activity relationship based machine learning, and a physiologically-based pharmacokinetic model of the brain was developed. Publicly sourced data on the plasma and brain pharmacokinetics of 77 small molecule drugs in rats was used for model development. The minimal brain PBPK model consisted of six ordinary differential equations and 19 parameters, which describes drug exposure in plasma, cerebrospinal fluid (CSF), brain interstitial fluid (ISF), and brain homogenate. The PBPK model parameters of 77 drugs were simultaneously estimated using nonlinear mixed effects modeling in Monolix. Support vector regression and multilayer perceptron models were used for QSAR to predict drug-specific PBPK model parameters from molecular descriptors.The combination of a genetic algorithm and variational autoencoder that uses SELF-referencing embedded strings (SELFIES) was used for de novo generation of small molecules. The genetic algorithm was used to simulate a population of 1000 molecules. Upon each new generation of molecules, the original population of 1000 molecules were mutated, cross-bred, scored, and refined to arrive at molecules that were predicted to have improved brain exposure. Individual molecules were subsequently modified using a VAE and a gradient-based optimization algorithm to further improve predicted brain exposure. Models were all converted to and intergrated in Python to generate 300 new molecules. The predicted brain exposure of newly generated molecules were compared against 300 randomly selected small organic molecules and 300 CNS drugs/candidates obtained from the literature.
Results: We have observed an approximate 30-fold and 120-fold increase on average in predicted brain exposure for AI generated molecules compared to known central nervous system drugs and randomly selected small organic molecules. Five of the seven estimated population mean parameters had percent relative standard errors (RSE) values <50%, while PSB and PST were estimated with percent RSE values of 90.4% and 68.4%, respectively. Percent RSE for random effects were all <35%. For the QSAR model, R2 values ranged from 0.19-0.42.
Conclusions: We successfully developed and integrated de novo design, QSAR-based machine learning, and PBPK modeling methods into a single algorithm. Due to the limited data, model acceptance criteria is poor. However, we believe that with additional data and mechanistic modeling this in silico pipeline could facilitate the discovery of a new wave of optimally designed medicines for the treatment of CNS diseases. Further expansion of this research could include the incorporation of additional mechanistic models (quantitative systems pharmacology/toxicology) for the simultaneous optimization of drug exposure, efficacy, and toxicity.
Please refer to our latest publication for additional details [1].
[1] https://www.biorxiv.org/content/10.1101/2022.10.29.514379v1.full.pdf