3.1. ML and DL For Vaccine Development

In this context, ML and DL play two important supporting roles: the dissemination of vaccine components by observing the viral protein structure, and helping medical researchers to review a large number of important research papers at an incredible rate. Three main types of vaccines are available: vaccines for any pathogen, such as flu or MMR, use deadly or compromised immune system infections; subunit vaccines (e.g., pertussis, shingles) only use part of the virus, such as protein; and, vaccines for nucleic acid inject viral genes into human cells to improve the response of the body ^[7]. The newest is the COVID-19 vaccine, which began trials in the United States this week. AI helps to speed up the growth of subunits and nucleic acids.

In analysing how it functions, understanding protein composition is critical. Researchers can produce drugs that work in various protein shapes once the situation is acknowledged. However, testing any protein structure will take quite a long time before finding its unique 3D structure. The method of evaluating protein structure and its genetic sequence can be simplified by AI systems that are based on DL.

Google DeepMind ^[8][9] introduced AlphaFold in January, an advanced and specialised system that forecasts the formation of 3D proteins while using their genetic sequence. In the beginning of March, the system was tested in COVID-19. In order to asses the scientific community for understanding the virus, DeepMind has published a protein prediction for several untreated proteins that are related to SARS-CoV-2, since it is the main cause for COVID-19. Meanwhile, scientists at the University of Texas at Austin and the National Institutes of Health have applied a standard biological method to establish the first 3D atomic map on the scale of a spike protein component of a virus that binds to human cells. Other Coronaviruses, including SARS-CoV and MERS-CoV, have spent years collaborating with the team that is responsible for this crucial breakthrough. Accurate predictions for this spike structure were provided by one of the predictions released by AlphaFold. Computer simulations for building 3D atom models on the SARS-CoV-2 protein scale neatly related to the results that were found on the UT Austin board were also used by ^[10] at the Institute for Protein Design at the University of Washington. By constructing new proteins to minimize Coronavirus, they are currently continuing on this work. In principle, these proteins will conform to a protein spike that protects healthy cells from being invaded by viral particles.

Researchers merged AI with cloud computing to stop the Spike protein from binding to the ACE2 receptor in human cells and to produce a possible vaccine for COVID-19 ^[11]. Flinders University researchers studied the COVID-19 virus and then applied their data to model a vaccine, called Covax-19, as shown in Figure 6.

Figure 6. Covax-19™ is an Australian-designed COVID-19 vaccine modeled by AI-based technologies ^[12].

In order to determine how the virus was harming human cells, the researchers utilized computer generated models of the S protein and its human receptor, the enzyme-converting angiotensin 2 (ACE2). Subsequently, they attempted to produce a vaccine that could prevent this mechanism. The team used the most innovative AI and advanced cloud computing technology to speed up the production of vaccines ^[9][11][13].

3.2. Drug Development

With the development of a predictive learning model, AI systems that are based on ML and DL are used in field design and they perform a fast-paced test to accurately represent the performance. The AI systems can easily classify drugs that can battle infectious diseases, such as COVID-19, with a drug recovery process. With this breakthrough, which is a medical evidence-based tool, will potentially enhance drug accessibility, preparation, care, and recorded patient outcomes COVID-19. Drug development is a very dangerous, lengthy, and costly phase. While it takes ten to fifteen years to make a new molecular venture, the success rate is only 2.01%, as reported by the Eastern Research Group (ERG) ^[14]. For the treatment of a never-considered medicinal indication, the idea of drug repurposing reuses old medications. It is an experimental strategy for the detection of pre-approved, discontinued, shelved, and investigational drugs for the treatment of other diseases with approved restatement. The development of conventional drugs generally involves five steps ^[15]:

• Drug discovery and development.

• Drug pre-clinical research.

• Drug clinical research.

• FDA drug review.

• FDA drug after-market safety control and development.

But only drug repurposing requires four steps ^[15]:

• Compound identification.

• Compound acquisition.

• Drug clinical research.

• FDA drug after-market safety control and development.

The discovery and development of new molecular structure are slow, time consuming, and expensive. Therefore, the best option is to reconstruct the approved drugs for SARS-COV-2 therapy. In this case, in the treatment of viral infection, Chloroquine (CQ) and its hydroxyl analogue Hydroxychloroquine (HCQ) have been identified. These drugs have antimalarial efficacy and they have also been shown to treat COVID-19 in vitro ^[16].

One AI-based drug development team focuses on the discovery at the molecular level of new drug-like compounds. Beck et al. ^[17] suggested a DL-based drug-target interaction model, called Molecule Transformer-Drug Target Interaction (MT-DTI), to predict potential drugs for COVID-19. SMILES strings and amino-acid sequences were used in the MT-DTI model in order to classify target proteins with 3D crystal structures. The findings showed that atazanavir, an antiretroviral drug that is used to treat and prevent the human immunodeficiency virus (HIV), is the best chemical compound with an inhibitory potency with Kd of 94.94 nM against 3C-like SARS-CoV-2 proteinase, followed by remdesivir (113.13 nM), efavirenz (199.17 nM), ritonavir (204.05 nM), and dolutegravir (336.91 nM). From the databases of NCBI, Drug Target Popular (DTC), and BindingDB, the authors obtained the amino acid sequences of 3C-like proteases and associated antiviral drugs and drug targets. In addition, they utilised a molecular docking and virtual screening method (AutoDock Vina) in order to determine the binding affinity between 3410 drugs and SARS-CoV-2 3CLpro. Six possible medicines, such as Remdesivir, Atazanavir, Efavirenz, Ritonavir, Dolutegravir, Kaletra (lopinavir/ritonavir), produced experimental results. Note that, in a clinical trial, Remdesivir looks promising.

Using effective machine learning techniques, two similarity-based methods, KronRLS ^[18] and SimBoost ^[19], have been proposed. However, there are two downsides to this matrix. First, the representation of features is reduced, which would cause the prediction faulty. Second, it involves the estimation of the matrix of similarity, which, in the training phase, will limit the maximum number of molecules. A deep learning based DTI model, DeepDTA ^[20], was proposed in order to address these limitations. It is a CNN-based end-to-end model that waives the feature engineering requirement. From a raw molecule and protein sequence, the model automatically finds useful features. The performance was seen on two publicly accessible DTI benchmarks, i.e., Sim-Boost and KronRLS. In order to extract features representations from the raw protein sequences and SMILES strings and train them, they used CNN blocks and combined these features representations to input into a deep CNN, naming it DeepDTA. They applied Smith–Waterman (S-W) and Pubchem Similarity algorithms to process the proteins and ligands’ pair-wise similarities, respectively. In tarining the patterns of the data, three alternative combinations applied this knowledge as input to the proposed ad enhanced DeepDTA model. The following are the three alternative combinations for training this model: