Chapter14 : Artificial Intelligence and Quantum Computing as the Next Pharma Disruptors

reading notes of《Artificial Intelligence in Drug Design》

---

1.Introduction

• Clinical trials (phase I, II, and III) are conducted in a highly regulated environment, under the good clinical practice standards. Phase I essentially addresses drug safety, while Phase II focuses on effectivity, and Phase III on confirming the findings on a larger population cohort.
  

2.Methods

2.1.The Basics of AI

• In recent times, due to the availability of large data sets and the increasing computational power, deep neural networks originated a great amount of attention from the scientific community.

2.1.1.Neural Networks

• A list of all activation functions and their different computational and learning properties can be found in Refs. 22, 23.
  

2.1.2.Deep Learning

• One of the first applications of ML in pharmaceutical sciences consisted in ANNs in quantitative structure–activity relationship (QSAR) studies.

2.2.AI in Drug Discovery and Development

• Here, we will mention the general approach to using deep learning in Virtual Screening (VS), one of the most widely used computational methods in drug development, and examples of the successful applications of deep learning in this field.

2.2.1.Virtual Screening

• In general, there are two different approaches to perform a VS, that is, ligand- based (LBVS) and structure-based (SBVS). The first allows the identification of compounds with similar molecular and chemical properties to those displayed by known ligands of the target of interest, and it is used when the 3D structure information of this target is not available. In contrast, SBVS uses the target’s structural information and its active sites to identify the compounds that bind in these regions of the target with the best binding potential.
  
• To solve the imbalance situation and the possible bias caused by the majority class, studies in VS have used strategies such as over- sampling of the minority class and subsampling guided by the distances between the compounds to retain the most representative samples and ensure structural diversity. More strategies to solve the imbalance issue can be found in ref. 34.
• For the case of LBVS, different databases such as PubChem, ChEMBL, DrugBank, and ZINC, among others, are used to create the datasets that will include the known ligands of the desired targets together with their experimental activity.
• Similarly, for the SBSV, databases such as BindingDB and Directory of Useful Decoys Enhanced (DUD-E) are used to create the datasets that will be composed of ligand-target complexes and their corresponding docking binding potential.
• For virtual screening processes, it is common to use molecular descriptors that can be calculated using toolkits such as RDKit, Chemistry Development Toolkit, and Dragon, among others.
  Once the deep learning model has been created and validated, it can predict the output from new compounds. Usually, VS processes analyze large open chemical libraries, such as ChemBridge, ChemDiv, and ZINC, among others, to look for candidate compounds that could be successful in the experimental tests.
• Due to the computational time that would be necessary to vectorize the enormous number of compounds present in chemical libraries, it is usual to apply previous filters based on drug-like properties to reduce the size of the chemical library before using the deep learning model. In this sense, deep learning models have been used to predict desired properties such as aqueous solubility, melting point, and toxicity, among others.

2.2.Application

• In the study of Neves et al., the use of deep learning models guided them in the identification of potential antimalarial agents from a chemical library of 486,115 compounds, finding two with important antiplasmodial activity and low cytotoxicity in mammalian cells.
• Deep learning has also been applied to look for new drug candidates in the current race against SARS-Cov-2. Zhang et al., used a previously designed deep learning model for reverse search drug targets, using the structural information of the 2019- nCov 3C-like protease, which was predicted by homology modeling. The VS was performed in 4 chemical compound databases and a database of tripeptides. As a result, eight potential compounds and six tripeptides were identified and were suggested as hits for further experimental tests.
• Similarly, in the study of Joshi et al., a deep learning model to predict the potential IC50 value against SARS coronavirus-I 3C-like proteinase was developed and later used to screen a database of 1611 natural compounds. Their workflow also included further computational analysis, such as molecular docking, drug- like filters, and molecular dynamics. Their final results suggest two natural compounds that are likely to have good experimental values.

2.3.How can Quantum Computing Accelerate Drug Discovery

• The significance of computational modelling for designing drugs and delivery systems, and the importance of quantum computers in the development of more effective therapeutic agents or drug carriers, have been recently highlighted in Ref. 69.

2.3.1.AI and Quantum Computing

• The main goal of ML is to increase accuracy, maximizing machine performance by learning from the experience, and providing self-learning algorithms and knowledge. In turn, QC allows improving ML algorithms and accelerating AI.
• For more Quantum algorithms details see Ref 59.
• Two categories of quantum algorithms that are relevant to drug discovery, are quantum algorithms for simulating molecular electronic structure and quantum-enhanced machine learning. These algorithms and the classical quantum chemistry methods, including ab initio and semiempirical approaches, and Density Functional Theory, are duly described in refs.
• Quantum machine learning allows translating classical ML algorithms into a quantum circuit, to be run more efficiently on quantum computers. These are based on common tools used in the early drug discovery processes including supervised and unsupervised learning. Quantum machine learning techniques on both Noisy Intermediate Scale Quantum (NISQ) and Fault-Tolerant Quantum Computer (FTQC) devices are summarized in Table 1.
  

2.3.2.Quantum Computing as the Next Pharma Disruptor

• Major challenges to implementing such scalable computing system, that is, handling parallel runs of computation on increasingly complex tasks, are to (1) translate chemical problems into quantum computational problems, (2) develop hardware and software that take advantage of the benefits of the technology, (3) iden- tify how such a system will fit to the needs and characteristics of the lab, (4) gather specialized knowledge to operate, in a repeatable and sustainable mode, and (5) provide quantum tools for application by nonspecialists: “You can not just take your backroom data analysts and turn them into quantum computer programmers overnight” (from Christopher Savoie, CEO, Zapata Computing).
• One example of a strong collaborative effort involving quan- tum computing and drug discovery is Biogen, which has teamed up with quantum computing software provider 1QBit and consulting firm Accenture to develop a quantum-based tool to compare molecules.
  
  
  
• Designing new drugs with suitable pharmacological profiles is still very challenging. QC has the potential to open new avenues for the (1) simultaneous mapping of a huge plethora of compounds including proteins and exploring the respective conformations, (2) rapid identification of drug candidates, and appropriate binding sites for proteins, accelerate the process of (3) comparing the effects of different drugs and their interactions, (4) predicting the mechan- isms of chemical reactions, (5) drug repurposing, and (6) creating more effective therapeutics by developing analytic methods for large-scale molecules (Fig. 6).
  
• For a full description of a wealth of applications of QM-based methods in drug discovery and development see Refs. 63, 64, 66, 67, 69.

3.Conclusion

• Drug discovery and development are already benefiting from advances in AI, machine learning and quantum chemistry, being pioneering application areas that are endorsing quantum computing into their pipelines. Quantum computing is prone to leverage the existing classical computing technologies, and pharma companies have to develop a workforce for dealing with this high level of computing. Both AI and QC hold a huge disruptive potential for supporting the pharmaceutical business in developing effective and safer drugs, selecting the best candidate molecules for the clinic, at reduced costs and time to market, facilitating the repurposing of preapproved drugs for new applications, and even developing novel antibody structures and motifs for evaluating undruggable targets.