Advances in the prediction of mouse liver microsomal studies: From machine learning to deep learning

Abstract

In the drug development process, mouse liver microsomal (MLM) studies are an invaluable biological assay used to assess the metabolic stability of novel drug candidates prior to human studies. However, determining MLM stability, in addition to other absorption, distribution, metabolism, and excretion (ADME) properties, can be a time-intensive and expensive process if it were tested in many compounds, thus leading to the need to create computational models capable of predicting properties of novel compounds. Additionally, building accurate computational models for the prediction of MLM stability can greatly accelerate the screening process for the selection of an appropriate drug candidate and further reduce the failure rate of the compounds in later trial stages. Our study outlined within this paper will discuss the history of computational models and their ability to predict MLM stability using traditional machine learning methods, as well as discuss a novel deep learning architecture, graph convolutional neural networks, capable of stronger predictive capabilities when compared to traditional methods. With future advances in hardware and research, deep learning methods applied to the prediction of ADME properties including but not limited to microsomal stability prediction represent an invaluable tool for future drug discovery efforts in both industry and academic settings.

Ensemble modeling with machine learning and deep learning to provide interpretable generalized rules for classifying CNS drugs with high prediction power

“…We further adopted a GCN architecture [ 27 ] to build a CNS drug classifier. The same datasets used in the ML models were applied in this study, with a main dataset (split by fingerprints) and an external testing dataset.…”

Enhancing Multi-species Liver Microsomal Stability Prediction through Artificial Intelligence

“…41 This characterization included eight atomic features, such as the atomic symbol, atomicity, formal charge, number of free radical electrons, hybrid type, aromatics, total hydrogen, and chirality. Additionally, four edge features were incorporated, including the bond type, conjugation, ring membership, and stereoscopic information; (2) MOE2d descriptors: these descriptors provide information on partial charges, atomic and bond numbers, subdivided surface area, and other physical properties; 39 (3) Molecular Access System (MACCS) structure fragments: these fragments offer specialized substructure information; 42 (4) Chemical Advanced Template Searches (CATS) descriptors: these descriptors are pharmacophoric representations used to calculate distances between pairs of atoms; 43 (5) Electrical Topological State Indices (ESTATE) descriptors: these descriptors indicate the perturbed electronic states of an atom influenced by the electronic states of all other atoms in the molecule; 44 (6) RDKit descriptors (RDKit-d): these descriptors are conformationally independent and obtained through molecular symbolic representation, including experimental and theoretical descriptors; 45 (7) Extended Connectivity Fingerprints with a bond diameter of 4 (ECFP4): ECFP4 is a class of 1024 bit circular fingerprints based on the lengths of the two bonds representing the circular atomic neighborhoods; 46 (8) functional class fingerprints with a bond diameter of 4 (FCFP4): FCFP4 is a class of 1024 bit pharmacophoric initial atom identifiers analogous to catalyst pharmacophoric identifiers; 46 (9) RDKit fingerprints (RDKit-f): RDKit-f is a class of 1024 bit hashed substructure or path fingerprints. 45 Notably, the representation of molecular graphs was performed using the Python package DGLlife.…”

“…These enzymes are responsible for the metabolism of a majority (70−80%) of clinically approved drugs, covering most of the phase I metabolism. 8,9 Since drug metabolism affects pharmacokinetic parameters and drug bioavailability, understanding the liver microsomal stability of NCEs is crucial for successful drug discovery and development. 6,10−12 Although the latest in vitro cellular models, such as the liver microsomal stability model, are recommended as initial screening systems for a large number of compounds, significant time and financial costs associated with them could not be ignored.…”

“…These models achieved an accuracy of 0.771 on the test set. In 2020, Renn et al 8 developed a classic Bayesian model using a data set of approximately 1500 molecules with mouse liver microsomal stability information from other literature. The best performing model achieved an accuracy of 0.830 on the test set.…”

“…Moreover, the focus of recent studies has primarily been on comparative analysis of model results, often overlooking the interpretability of graphbased algorithms themselves. 1,[6][7][8]31 In this study, we aimed to enhance multispecies models for liver microsomal stability in humans, rats, and mice while providing comprehensive and practical explanations based on these models and data. To accomplish this objective, we curated a vast and publicly accessible data set by incorporating data from several representative databases.…”