Artificial intelligence in Medicinal Chemistry: a real avenue for speeding up neurodrug discovery process

CRAVERO, FIORELLA; DÍAZ, MÓNICA F.; ADRIO, JAVIER; PONZONI, IGNACIO; MARTÍNEZ, MARÍA JIMENA; ROCA, CARLOS; ARRAYÁS, RAMÓN GÓMEZ; A. MARTINEZ; SEBASTIÁN-PÉREZ, VÍCTOR; REQUENA-TRIGUERO, CARLOS; PÁEZ, JUAN A.; C. GIL; CAMPILLO, NURIA E.

Ljubljana

Simposio; EFMC-ISMC 2018 XXV EFMC International Symposium on Medicinal Chemistry; 2018

Section for Medicinal Chemistry of the Slovenian Pharmaceutical Society (SFD), on behalf of the European Federation for Medicinal Chemistry (EFMC)

Artificial Intelligence (AI) has recently become an essential part of the technology industry, solving many challenging problems in computer sciences. Also, the biopharmaceutical industry is looking toward AI to speed updrug discovery, cut R&D costs, decrease failure rates in drug trials and create better medicines.In this sense, machine learning (ML) approaches have emerged as very powerful tools that can be applied in several steps of the iterative drug discovery process, such as Quantitative Structure Activity Relationship (QSAR) for the prediction of activity of large untested databases, discovery of hit compounds or synthesis prioritization for lead optimization. In order to reduce attrition rate in later stages of drug discovery and avoid compounds with undesirable properties, the development of QSPR models for the prediction of the pharmacokinetic and toxicological (ADMET) profile plays also a key role in lead optimization.For this purpose, in this work we have developed several QSXR models (X: activity, enantiomeric excess and ADME properties) to optimize drug discovery process in neurodrugs.The first model was developed to identify inhibitors of BACE1. The work-set includes compounds with a representative chemical space and a wide variety of drug-like properties available from different databases. Models were obtained by the application of several ML methods, model hybridizing strategies, combinatorial analysis and visual analytics. A performance of 85% for corrected classified compounds and ROC value of 0.88 was obtained.Our approach contributes to achieve a QSAR model that can be a useful virtual screening method for prediction of BACE1 inhibitors with a wide applicability domain.Once the hit is identified, hit optimization process is carried out using chemical synthesis where several ML methods can be developed to predict the outcome of the reaction.6 Also, ADME properties are essential in lead optimization. One of the critical steps in QSXR modeling is the identification of the most informative molecular descriptors. For this purpose, two main general approaches can be used: feature selection and feature learning. Toaddress both issues, a performance comparative study of two state-of-art methods taking into account these two approaches was carried out using different databases at these stages of the drug discovery process. These databases include enantiomeric excess in the chemical synthesis and blood brain barrier or human intestinal absorption in physicochemical properties assessment. Regression and classification models were built for the three datasets using both approaches together with their potential hybridization to analyze which technique achieves a better performance to be further applied.