Artificial Intelligence: How can it be useful in drug discovery

It's fundamental for every potential drug candidate to know its adverse effects and to evaluate its benefit-risk balance, that means knowing if a chemical molecule with pharmacological activity has more therapeutic effect or it's more dangerous because of adverse reactions that can arise after its administration, all for eliminating unwanted toxic effects (Debleena et al., 2021). For scientists it's important to discover high-quality chemical molecules that must be safe for humans and the environment, for this reason after finding a new chemical molecule trials and evaluations are performed for predicting pharmacokinetic properties (ADMET): according to many reports adverse reactions are also one of the main causes for the failure of drug development, so the optimization of the ADMET properties is a key factor for the success of drug development (Tripathi et al; 2021). Nowadays in vivo toxicological experiments that control side effects employ too much time and costs to be performed, so computational methods can lead to an improvement of this situation by decreasing the costs of drug manufacturing by the 50% through the use of ML models: AI approaches such as Tox21 Data Challenge was organized by the National Institutes of Health and Us Food and Drug Administration (FDA) for predicting and classifying 12 specific toxicities of 12000 chemical molecules and drugs through Virtual Screening (VS), analyzing the similarities between compounds and upon this basis predicting the toxicities or also by discovering the adverse effects thanks to the input data that regard the features of these molecules from Neural Networks of the ML models (Huang et al; 2016; Debleena et al., 2021; Tripathiet al; 2021).

In particular the Tox21 Data Challenge is able to compare the functions of different computational methods for analyzing the biological activity and obviously in a more detailed way those related to the toxicity and for doing this the techniques use the best predictive ML models, such as OCHEM, that is a web-based database that uses Neural Networks for collecting data about the adverse effects of compounds in a more specific way (Huang et al; 2016). The Random Forest method of ML used for the distribution of features of chemical molecules has been used for predicting the activities of potential drug candidates against p53 mutations and aromatase for example, very important entities to eradicate for the arising of cancer diseases, because the p53 is a protein and it's also a tumor suppressor gene, that means a gene that has the function to suppress tumors because it can induce the so called programmed cell death of the cell that brings the tumor death together with the cell and so it's not a specific process (Robbins and Cotran, 2005; Huang et al; 2016). Mutations in p53 are those that can lead to the formation of tumors instead (like esophageal cancers or pancreatic carcinoma, respectively in the esophagus or pancreas), so for this reason it's important to target this protein otherwise the p53 protein without mutations on the contrary is safe and protects the genome (Huang et al; 2016; Robbins and Cotran, 2005). Aromatase is another protein that instead is required for the formation of estrogens, endocrine women hormones involved in menopause and its overexpression (that means too much biological activity) leads to the formation of endometrial cancer that is the inner layer of the women's uterine cavity, so this is why drugs' features have to be identified and analyzed through AI techniques: an example of the features of chemical molecules found due to ML models and molecular docking through Virtual Screening is Anastrozole, an important aromatase inhibitor that can treat also breast cancer (Brunton et al; 2011; Tortora, 201; Huang et al; 2016; Ishfaq et al; 2022). The problem with the aromatase inhibitors is related to their adverse effects such as arthralgia that means pain in joints, myalgia that is pain in muscles, sweat, loss of sex drive and vaginal dryness and so it's important to find new efficient aromatase inhibitors: in this sense ML models collect and visualize data for making the virtual screening of the features and the biological activities of more than 5000 chemical molecules from the PubChem database, in which eventually 5 of them have been chosen for performing the similarity analysis with other molecules for discovering their characteristics (Tortora, 2014; Ishfaq et al; 2022). Another ML approach called DeepTox that is part of the Tox21 Data Challenge was used for describing other specific features of the chemical molecules such as the molecular weight (M.W.), that is the weight of the molecule and the Van der Waals interactions (weak interactions involving apolar molecules), also for predicting the adverse effects (Silberberg, 2013; Debleena et al., 2021). DeepTox learns the adverse effects of chemical molecules by putting them into the Neural Networks for their analysis and then predicts the toxicities and it was seen to be the most efficient computational method for the screening of adverse effects, exceeding also the RF technique typical of Deep Learning , moreover the aim of the Tox21 Data Challenge, whose DeepTox is part even if it can be autonomously used, is the production of more competent and less-time consuming methods for predicting how chemical molecules affect human health, even if computational models have also disadvantages to overcome such as insufficient accuracy and unreliability like the biological experiments do, in fact the Tox21 Data Challenge with the its computational methods were used for reducing in vitro experiments and animal testing (Mayr et al; 2016). The toxic effects analyzed in the Tox21 Data Challenge computational methods are the stress response effects (SR), that is the response that the body produces when stress occurs and affects all systems of the body including the cardiovascular, respiratory, endocrine, gastrointestinal, nervous, muscular and reproductive systems: the stress symptoms that affect the cardiovascular system for example are the increase of the heart rate, increased heart muscle contractions and its dilation and since the airways of the respiratory system are constricted the volume of the oxygen decreases by forcing the human to increase the frequency of breathing (the number of times a person breaths) for counterbalancing this loss (Chu et al; 2022).

The stress symptoms have different physiological effects, such as the suppression of the immuno-inflammatory mechanism that consists in the recruitment of lymphocytes, molecules of the immune system that protect the body from the entry of pathogens such as bacteria and cause the inflammatory process that is needed for destroying these entities and for reconstructing a damaged tissue (Abbas et al; 2016; Robbins and Cotran, 2005). Cytokines are other molecules of the immune system used against the foreign entities that come from the external environment, such as TNF (Tumor-Necrosis Factor) and IL-1 (Interleukin-1) (Chu et al; 2022). The stress response effects analyzed by the Tox21 Data Challenge are the heat shock response, that is a mechanism that the cell uses as a protection against hypothermia through the rapid activation of proteins called heat shock proteins that in fact as the name suggests overcome the heat and there are also the nuclear receptor effects that arise due to its binding with ligands that can be steroid hormones (such as estrogens), thyroid hormones (such as thyroxine) and retinoids (like isotretinoin) that are chemical molecules used for treating acne and are important for human vision (Wong et al; 2009; Patrick, 2017; Tortora, 2014; Brunton et al; 2011). These adverse effects when activated lead to the desruption of the endocrine system, that is the system related to the production of hormones and the activation of the stress response in turn can lead in fact to liver damage or to cancer and for this reason they were analyzed by the Virtual Screening in the ML models and through the similarity-based approaches by comparing the chemical molecules and by seeing the features (Tortora, 2014; Mayr et al; 2016). Deep Learning techniques represent the chemical features of compounds in the form of images inside the Neural Networks, in other words the features as input data are put inside the ANNs and these characteristics can indicate the adverse effects or the functional groups (a group of atoms with specific properties inside a molecule), a specific endowment of ML approaches called multi-task learning, the capacity of learning simultaneously and autonomously from different topics and this increases the performance of choosing the appropriate drug candidates (McMurry, 2011; Mayr et al; 2016). DeepTox has the main function of cleaning and ensuring the quality control of the data that regard the chemical description of the chemical compounds, putting the data features as input into the Neural Networks and collecting all the chemical compounds in general, but since all of these entities are mixtures of molecules bound together through weak interactions (such as Van der Waals ones) the DeepTox pipeline works through a fragmentation step performed by Artificial Neural Networks (ANNs) in the sense that the different structures are deconstructed into 'compound fragments': examples are Na+ and Cl- ions and these fragments can appear many times while DeepTox labels those that belong to specific adverse effects, trying to remove the ones that are contradictory, meaning that there must be fragments that fit with specific adverse effects (there must be no errors) (Mayr et al; 2016).