We develop a machine learning model that tackles both reaction prediction and retrosynthesis by learning from the same dataset. The model is generalizable across chemical space.
Mitochondrial dysfunction has been implicated as an important factor in the development of idiosyncratic organ toxicity. An ability to predict mitochondrial dysfunction early in the drug development process enables the deselection of those drug candidates with potential safety liabilities, allowing resources to be focused on those compounds with the highest chance of success to the market. A database of greater than 2000 compounds was analyzed to identify structural and physicochemical features associated with the uncoupling of oxidative phosphorylation (herein defined as an increase in basal respiration). Many toxicophores associated with potent uncoupling activity were identified, and these could be divided into two main mechanistic classes, protonophores and redox cyclers. For the protonophores, potent uncoupling activity was often promoted by high lipophilicity and apparent stabilization of the anionic charge resulting from deprotonation of the protonophore. The potency of redox cyclers did not appear to be prone to variations in lipophilicity. Only 11 toxicophores were of sufficient predictive performance that they could be incorporated into a structural-alert model. Each alert was associated with one of three confidence levels (high, medium, and low) depending upon the lipophilicity-activity profile of the structural class. The final model identified over 68% of those compounds with potent uncoupling activity and with a value for specificity above 99%. We discuss the advantages and limitations of this approach and conclude that although structural alert methodology is useful for identifying toxicophores associated with mitochondrial dysfunction, they are not a replacement for the mitochondrial dysfunction assays in early screening paradigms.
Throughout the past decade, the expectations from the regulatory agencies for safety, drug-drug interactions (DDIs), pharmacokinetic, and disposition characterization of new chemical entities (NCEs) by pharmaceutical companies seeking registration have increased. DDIs are frequently assessed using in silico, in vitro, and in vivo methodologies. However, a key gap in this screening paradigm is a full structural understanding of time-dependent inhibition (TDI) on the cytochrome P450 systems, particularly P450 3A4. To address this, a number of high-throughput in vitro assays have been developed. This work describes an automated assay for TDI using two concentrations at two time points (2 + 2 assay). Data generated with this assay for over 2000 compounds from multiple therapeutic programs were used to generate in silico Bayesian classification models of P450 3A4-mediated TDI. These in silico models were validated using several external test sets and multiple random group testing (receiver operator curve value >0.847). We identified a number of substructures that were likely to elicit TDI, the majority containing indazole rings. These in vitro and in silico approaches have been implemented as a part of the Pfizer screening paradigm. The Bayesian models are available on the intranet to guide synthetic strategy, predict whether a NCE is likely to cause a TDI via P450 3A4, filter for in vitro testing, and identify substructures important for TDI as well as those that do not cause TDI. This represents an integrated in silico-in vitro strategy for addressing P450 3A4 TDI and improving the efficiency of screening.
scite is a Brooklyn-based organization that helps researchers better discover and understand research articles through Smart Citations–citations that display the context of the citation and describe whether the article provides supporting or contrasting evidence. scite is used by students and researchers from around the world and is funded in part by the National Science Foundation and the National Institute on Drug Abuse of the National Institutes of Health.