A Handbook of Artificial Intelligence in Drug Delivery

A Handbook of Artificial Intelligence in Drug Delivery explores the use of Artificial Intelligence (AI) in drug delivery strategies. The book covers pharmaceutical AI and drug discovery challenges, Artificial Intelligence tools for drug research, AI enabled intelligent drug delivery systems and next generation novel therapeutics, broad utility of AI for designing novel micro/nanosystems for drug delivery, AI driven personalized medicine and Gene therapy, 3D Organ printing and tissue engineering, Advanced nanosystems based on AI principles (nanorobots, nanomachines), opportunities and challenges using artificial intelligence in ADME/Tox in drug development, commercialization and regulatory perspectives, ethics in AI, and more.

This book will be useful to academic and industrial researchers interested in drug delivery, chemical biology, computational chemistry, medicinal chemistry and bioinformatics. The massive time and costs investments in drug research and development necessitate application of more innovative techniques and smart strategies.

• Focuses on the use of Artificial Intelligence in drug delivery strategies and future impacts
• Provides insights into how artificial intelligence can be effectively used for the development of advanced drug delivery systems
• Written by experts in the field of advanced drug delivery systems and digital health

• Language: English
• Publisher: Academic Press
• Release date: Mar 27, 2023
• ISBN: 9780323903738

Book preview

Chapter 1: An overview of artificial intelligence in drug development

Anil K. Philipa; Md. Faiyazuddinb,c    a School of Pharmacy, University of Nizwa, Nizwa, Oman

b School of Pharmacy, Al-Karim University, Katihar, Bihar, India

c Nano Drug Delivery® (A product development partnership company), Raleigh-Durham, NC, United States

Abstract

Artificial intelligence (AI) has become a very popular field of study in recent years, with the hope of making computers more intelligent by mimicking human behavior. However, AI technology has already been applied to various other areas such as drug development. Pharmaceutical companies have partnered with software companies specializing in AI to take advantage of AI in drug development. AI can also assist in the reduction of compounds being considered for drug development, as well as the removal of drugs that may cause negative side effects. Additionally, nanobots could be used to deliver drugs to specific parts of the body in the future, which would improve efficacy and minimize side effects. Finally, combination drug development could benefit from AI-based optimization techniques to explore different combinations to maximize efficacy, minimize side effects, and improve patient compliance.

Keywords

Artificial intelligence; Drug development; Nanobots; Repurposing; Algorithms

1.1: Introduction

Artificial Intelligence (AI) became popular in the 1940s, and it was studied to see if computers could process information and make decisions faster than humans [1]. AI is a term that was coined in 1956 [2]. One of the most exciting subfields of computer science is AI, which promises to make computers more intelligent by mimicking human behavior [3]. In the early 1970s, researchers discovered how AI systems could be applied to various areas of the life sciences [4]. AI in drug development was made possible by the large amounts of chemical and biological data accumulated over decades and technological automation was made possible by the use of high-performance processors [5,6]. There have been some breakthroughs in AI technologies. AI technologies have the potential to outperform humans in many clinical areas [7].

Drug development has long been expensive and time-consuming. However, with the help of AI, new drugs can be produced faster and more efficiently [8]. The pharmaceutical industry first used AI in drug development before other medical fields [9]. As a result, it is expected to improve the precision and efficiency of drug developers [10]. Several pharmaceutical companies have partnered with software companies specializing in AI to take advantage of AI in drug development [11]. Smalley reported that AI-based algorithms can assist in the reduction of compounds being considered for drug development, as well as the removal of drugs that may cause negative side effects [12]. AI in drug development can help speed up the drug manufacturing process. Translational research typically takes 14 years, but with the in-silico approach of AI, it has become easier to conduct tests in vitro and in vivo [13,14]. Nanobots could be used to deliver drugs to specific parts of the body in the future. Integrating AI into the nanobot system can allow more specific drugs to reach the exact areas needed for maximum effect [9]. Combination drug development could benefit from AI-based optimization techniques to explore different combinations to maximize efficacy, minimize side effects, and improve patient adherence [15].

Virtual reality (VR) technology has gained popularity. In the pharmaceutical industry, VR can replace or supplement pharmacotherapy. Pharmacist education is currently using VR to help students learn more immersive. Drug development teams are using VR to test new drugs and get feedback from users earlier in the process. VR is also being used to help discover new drugs and their side effects. Scientists are using VR to study different animals behavior and see how they respond to new drugs. Patients are being counseled about their medications through VR simulations. Interest in VR research grows despite financial constraints [16]. Drug development by using AI technology is progressing at a rapid pace. Many pharmaceutical companies are partnering with AI startups and academic institutions to launch their internal research and development projects. This shows that the industry is beginning to realize the immense potential of this technology to revolutionize drug discovery and development [17]. Zhavoronkov describes in detail the drug development process and how Deep Learning can be used in this process [18]. Ekert et al. also examined how artificial intelligence technology is used in drug discovery and development. In his opinion, machine learning, computational modeling, AI, and in-vitro modeling will proliferate in the future [19]. A report discusses how artificial intelligence can be used for rational drug design [20].

Turea reports that most healthcare decision-makers see the benefits of AI and understand its possibilities. However, they also fear that it could be responsible for a fatal error. Despite these concerns, AI has proven its value in many cases and continues to improve [21].

1.2: Impact of AI on drug development

The drug development process is complex, and to find promising compounds, developers must process a large amount of information. However, AI applications have been introduced to make this process more efficient. By using these applications, developers can narrow down the search for potential drugs and save time and resources [22]. Pharmaceutical companies and startups are increasingly using AI to research and develop new drugs [23]. The next decade of drug development will have the AI approaches under intense scrutiny. These will help improve the flow of work and provide insights that researchers can observe, analyze, and understand [24]. Many companies are developing AI capabilities for applications beyond drug discovery. Schuhmacher et al. reported AI capability development was still in its early stages, which means that counting the number of AI patents filed by a company is unlikely to impact the development of new drug candidates or the sourcing of external candidates during the observation period [25]. The AI model can be updated after studying cells or organoids, to develop a molecular optimization plan. A high-throughput bioassay and AI design can be used to automate a drug development cycle based on the biological effects of the drug. This will significantly accelerate the production of new drugs [26]. A drug's pharmacokinetic properties are affected by physicochemical properties such as ionization, solubility, and permeability. In the development of new drugs, these factors are of critical importance. As a result, the probability of discovering a new drug is higher when using AI-based techniques [27]. The growth of data-driven and algorithm-based research and development has created the need for a new way of thinking about how data mining and AI technologies can be used to discover and develop new medicines. Some things that are important to the success of this approach include classifying diseases into endotypes and integrating artificial intelligence and machine learning into drug development [8].

There is considerable interest in developing therapeutics that alleviate the symptoms of Alzheimer's disease without adverse side effects. This could be achieved by using AI technology to repurpose well-known drugs to treat Alzheimer's disease [28,29]. AI models have been used to determine how the disease spreads and stop it [30]. With the help of artificial intelligence, new eye diseases may be diagnosed [31]. AI algorithms will help identify new biomarkers for diseases. This is possible because they can search for specific features themselves, rather than just recognizing clinical features [32]. AI programs are helping to shape the future of disease prediction and diagnosis. They offer innovative ways to manage diseases and make better treatment decisions [33]. AI solutions to medical problems make it possible to understand how diseases are related to different signs and symptoms. This enables doctors to create better treatment plans and diagnoses for their patients [34]. The near future will show how these AI-based digital technologies can offer new targets, improve clinical trial design, and have a broader impact on the pharmaceutical business [35]. FDA approval of AI applications paves the way for regulatory development to enable faster integration of AI-enabled technologies into healthcare [36]. The use of AI in eHealth means that mental health clinicians can engage with their patients more efficiently and effectively. This represents a shift in how mental healthcare is delivered and how patients interact with it [37,38]. GPCR pharmacology is concerned with how drugs interact with G-protein–coupled receptors (GPCRs). This has led to a high success rate in developing drugs that target these proteins, with 78% of drugs succeeding in Phase I clinical trials. The use of several AI applications in drug discovery has increased success rates further [22]. A good dataset is necessary to build a useful predictive model. Without accurate data, even a complex model will not produce valuable results [8].

1.3: AI in drug repurposing

Repurposing involves using a drug already approved for a particular disease to treat another condition. Watson Drug Discovery uses AI to increase the efficiency of drug repurposing [39]. Among the best-known examples of drug repurposing are antiinflammatory drugs used as anticancer agents. Chloroquine, an antimalarial, and azithromycin, an antibacterial currently being developed as an antiviral, are two such drugs being developed to combat COVID-19 [40]. AI helps traditional technologies find and validate new drugs. AI can also aid in extracting valuable data for drug reuse faster [41]. AI can enable faster and more effective decision-making to help select and validate new targets [42]. In the drug development process, clinical trials are considered a bottleneck, and researchers believe AI technology can help conduct and design clinical trials. AI can predict whether a drug can be reclassified using transcriptomics, molecular structure data, and clinical databases [43]. The potential of using AI for repurposing drugs to treat neurodegenerative diseases is explored by Paranjpe et al. They emphasize the importance of integrating different types of data while using AI tools to avoid bias and increase accuracy. This will help ensure that the most effective drugs are found and used to treat these diseases [44]. A methodical interplay between drug discovery and drug-target interactions, which AI-based technologies can support, could help repurpose prescription drugs [45]. For machine learning and deep learning methods to be effective for specific tasks, AI must be integrated into human workflows, such as drug repurposing and clinical trials. Designing AI solutions for molecular generation is complicated, and standard practices for data exchange need to be developed and strengthened [46]. AI could help researchers connect different biological networks to find new uses for already developed drugs. Zeng et al. have developed a new AI-powered method for drug repurposing known as DeepDTnet. This method uses data from multiple biological entities to predict new drug-target interactions more accurately than previous methods [47]. AI has been used to predict drug-target interactions. This information can be helpful to reuse old medications or avoid taking multiple medications at the same time. A drug that is repurposed automatically qualifies for phase II clinical trials. Examination of patent applications published in 2011–14 for drug repurposing has yielded surprising results. The small number of patents for parasitic or tropical diseases contrasts drastically with this area's extensive research in peer-reviewed publications [48]. The ability to stratify complex diseases into several distinct forms based on large patient populations with well-characterized data is critical. The result is that patients can be classified into distinct subgroups based on the causes and influences of their ailments. These subgroups can be analyzed to identify new drug targets or repurposing opportunities [49]. Scientists have demonstrated that geometric deep learning could help predict and create fingerprints for molecular surface interactions [50]. Finding a drug with the opposite effect of another drug on transcriptional data can target diseases. Considering how ineffective the current approach is, new transcription-based methods such as CuGuCtD are revolutionizing our understanding and ability to determine whether a chemical can modulate gene expression in the same way that disease-modifying drugs do [51].

1.4: AI in developing improved policies

AI can help with policy making in several ways, even if it is still in its early stages. For example, expert systems can help decision-makers understand complex problems, and data mining can help identify patterns in data. Adversarial search can help determine the best possible actions to take [52]. AI techniques have been shown to help reduce the spread of disease in many different countries [53]. To build trust in AI, it is necessary to extend trust on the types of AI that ensure society benefits and reduced application risks. The development of AI faces many of the same challenges as other types of technology. Therefore, governments and AI-based companies need to develop strategies that address these challenges [54]. One way to make governance more accountable is to create well-functioning and transparent algorithms eliminating the harmful consequences of negative human decision-making. In addition, such algorithms can make the government more efficient by providing an accurate record of past decisions [55]. Many countries have been working on legislation and policies to encourage the adoption of AI technology and attract foreign investment in the technology sector. For example, the Department of Health in Abu Dhabi has developed an AI policy to regulate AI in the health-care sector [56]. The FDA's new policy focuses on excellence for developers of AI-based medical devices rather than the approval of those devices [57]. Updating AI models will not require prior FDA review helping improve the speed and accuracy of updates [58].

1.5: Conclusion

Drug repositioning is a way to develop new drugs by using old drugs that have not yet been approved. This is a very efficient, time-saving, and cost-effective way to increase the success rate of drug therapy. AI has a lot of potential to help improve many different aspects of society. However, in order for it to reach its full potential, we need to develop strategies to address the challenges that come with it. Legislation and policies have been developed by various governments in order to encourage the adoption of AI technology and attract foreign investment in the technology sector. As we continue to explore all that AI has to offer, it is important that we make sure that these technologies are used for the benefit of society as a whole.

References

[1] Akbari M., Do T.N.A. A systematic review of machine learning in logistics and supply chain management: current trends and future directions. Benchmarking Int. J. 2021;28(10):2977–3005. doi:10.1108/bij-10-2020-0514.

[2] Sharma S., Ahmed S., Naseem M., Alnumay W.S., Singh S., Cho G.H. A survey on applications of artificial intelligence for pre-parametric project cost and soil shear-strength estimation in construction and geotechnical engineering. Sensors. 2021;21(2):463. doi:10.3390/s21020463.

[3] Liu Y., Bi S., Shi Z., Hanzo L. When machine learning meets big data: a wireless communication perspective. IEEE Veh. Technol. Mag. 2020;15(1):63–72. doi:10.1109/mvt.2019.2953857.

[4] De A., Sarda A., Gupta S., Das S. Use of artificial intelligence in dermatology. Indian J. Dermatol. 2020;65(5):352. doi:10.4103/ijd.ijd_418_20.

[5] Lipinski C.F., Maltarollo V.G., Oliveira P.R., da Silva A.B.F., Honorio K.M. Advances and perspectives in applying deep learning for drug design and discovery. Front. Robot. AI. 2019;6:doi:10.3389/frobt.2019.00108.

[6] Kimber T.B., Chen Y., Volkamer A. Deep learning in virtual screening: recent applications and developments. Int. J. Mol. Sci. 2021;22(9):4435. doi:10.3390/ijms22094435.

[7] Diaz O., Kushibar K., Osuala R., Linardos A., Garrucho L., Igual L., Radeva P., Prior F., Gkontra P., Lekadir K. Data preparation for artificial intelligence in medical imaging: a comprehensive guide to open-access platforms and tools. Phys. Med. 2021;83:25–37. doi:10.1016/j.ejmp.2021.02.007.

[8] Schneider P., Walters W.P., Plowright A.T., Sieroka N., Listgarten J., Goodnow R.A., Fisher J., Jansen J.M., Duca J.S., Rush T.S., Zentgraf M., Hill J.E., Krutoholow E., Kohler M., Blaney J., Funatsu K., Luebkemann C., Schneider G. Rethinking drug design in the artificial intelligence era. Nat. Rev. Drug Discov. 2019;19(5):353–364. doi:10.1038/s41573-019-0050-3.

[9] Paul D., Sanap G., Shenoy S., Kalyane D., Kalia K., Tekade R.K. Artificial intelligence in drug discovery and development. Drug Discov. Today. 2021;26(1):80–93. doi:10.1016/j.drudis.2020.10.010.

[10] Asai A., Konno M., Taniguchi M., Vecchione A., Ishii H. Computational healthcare: present and future perspectives (review). Exp. Ther. Med. 2021;22(6):doi:10.3892/etm.2021.10786.

[11] Mhlanga D. The role of artificial intelligence and machine learning amid the COVID-19 pandemic: what lessons are we learning on 4IR and the sustainable development goals. Int. J. Environ. Res. Public Health. 2022;19(3):1879. doi:10.3390/ijerph19031879.

[12] Smalley E. AI-powered drug discovery captures pharma interest. Nat. Biotechnol. 2017;35(7):604–605. doi:10.1038/nbt0717-604.

[13] Zhavoronkov A., Vanhaelen Q., Oprea T.I. Will artificial intelligence for drug discovery impact clinical pharmacology?. Clin. Pharmacol. Ther. 2020;107(4):780–785. doi:10.1002/cpt.1795.

[14] Mikkili I., Karlapudi A.P., Venkateswarulu T.C., Kodali V.P., Macamdas D.S.S., Sreerama K. Potential of artificial intelligence to accelerate diagnosis and drug discovery for COVID-19. PeerJ. 2021;9:e12073doi:10.7717/peerj.12073.

[15] Mishra A., Gowrav M.P., Balamuralidhara V., Reddy K.S. Health in digital world: a regulatory overview in United States. J. Pharm. Res. Int. 2021;438–450. doi:10.9734/jpri/2021/v33i43b32573.

[16] Luo X. Analysis of the effect of virtual reality technology on improving drug design. In: Advances in Social Science, Education and Humanities Research. Atlantis Press; 2022:doi:10.2991/assehr.k.220110.124.

[17] Peng Y., Liu E., Peng S., Chen Q., Li D., Lian D. Using artificial intelligence technology to fight COVID-19: a review. Artif. Intell. Rev. 2022;doi:10.1007/s10462-021-10106-z.

[18] Zhavoronkov A. Artificial intelligence for drug discovery, biomarker development, and generation of novel chemistry. Mol. Pharm. 2018;15(10):4311–4313. doi:10.1021/acs.molpharmaceut.8b00930.

[19] Ekert J.E., Deakyne J., Pribul-Allen P., Terry R., Schofield C., Jeong C.G., Storey J., Mohamet L., Francis J., Naidoo A., Amador A., Klein J.-L., Rowan W. Recommended guidelines for developing, qualifying, and implementing complex in vitro models (CIVMs) for drug discovery. SLAS Discov. 2020;25(10):1174–1190. doi:10.1177/2472555220923332.

[20] Duch W., Swaminathan K., Meller J. Artificial intelligence approaches for rational drug design and discovery. Curr. Pharm. Des. 2007;13(14):1497–1508. doi:10.2174/138161207780765954.

[21] Ahmed N.J., Alzahrani A.A., Alonazi R.E., Menshawy M.A. The knowledge and attitudes of the public toward the clinical use of artificial intelligence. Asian J. Pharm. 2021;15:168–171. doi:10.22377/ajp.v15i1.3974.

[22] Brown N., Cambruzzi J., Cox P.J., Davies M., Dunbar J., Plumbley D., Sellwood M.A., Sim A., Williams-Jones B.I., Zwierzyna M., Sheppard D.W. Big data in drug discovery. Prog. Med. Chem. 2018;57(1):277–356. doi:10.1016/bs.pmch.2017.12.003.

[23] Damiati S.A. Digital pharmaceutical sciences. AAPS PharmSciTech. 2020;21(6):doi:10.1208/s12249-020-01747-4.

[24] Luo Y., Peng J., Ma J. Next decade's AI-based drug development features tight integration of data and computation. Health Data Sci. 2022;2022:1–3. doi:10.34133/2022/9816939.

[25] Schuhmacher A., Gatto A., Hinder M., Kuss M., Gassmann O. The upside of being a digital pharma player. Drug Discov. Today. 2020;25(9):1569–1574. doi:10.1016/j.drudis.2020.06.002.

[26] Schneider G. Automating drug discovery. Nat. Rev. Drug Discov. 2017;17(2):97–113. doi:10.1038/nrd.2017.232.

[27] Hessler G., Baringhaus K.-H. Artificial intelligence in drug design. Molecules. 2018;23(10):2520. doi:10.3390/molecules23102520.

[28] Fabrizio C., Termine A., Caltagirone C., Sancesario G. Artificial intelligence for Alzheimer's disease: promise or challenge?. Diagnostics. 2021;11(8):1473. doi:10.3390/diagnostics11081473.

[29] Zhang M., Schmitt-Ulms G., Sato C., Xi Z., Zhang Y., Zhou Y., George-Hyslop P.S., Rogaeva E. Drug repositioning for Alzheimer's disease based on systematic ‘omics’ data mining. PLOS One. 2016;11(12):e0168812doi:10.1371/journal.pone.0168812.

[30] Garg R., Patel A., Hoda W. Emerging role of artificial intelligence in medical sciences we ready!. J. Anaesthesiol. Clin. Pharmacol. 2021;37(1):35. doi:10.4103/joacp.joacp_634_20.

[31] Keel S., Lee P.Y., Scheetz J., Li Z., Kotowicz M.A., MacIsaac R.J., He M. Feasibility and patient acceptability of a novel artificial intelligence-based screening model for diabetic retinopathy at endocrinology out-patient services: a pilot study. Sci. Rep. 2018;8(1):doi:10.1038/s41598-018-22612-2.

[32] Moraru A., Costin D., Moraru R., Branisteanu D. Artificial intelligence and deep learning in ophthalmology—present and future (review). Exp. Ther. Med. 2020;doi:10.3892/etm.2020.9118.

[33] Ramesh A.N., Kambhampati C., Monson J.R.T., Drew P.J. Artificial intelligence in medicine. Ann. R. Coll. Surg. Engl. 2004;86(5):334–338. doi:10.1308/147870804290.

[34] Reddy S. Use of Artificial Intelligence in Healthcare Delivery. 2018.doi:10.5772/intechopen.74714.

[35] Hartl D., de Luca V., Kostikova A., Laramie J., Kennedy S., Ferrero E., Siegel R., Fink M., Ahmed S., Millholland J., Schuhmacher A., Hinder M., Piali L., Roth A. Translational precision medicine: an industry perspective. J. Transl. Med. 2021;19(1):doi:10.1186/s12967-021-02910-6.

[36] Filipiak-Strzecka D., Kasprzak J.D., Szymczyk E., Wejner-Mik P., Lipiec P. Bedside screening with the use of pocket-size imaging device can be useful for ruling out carotid artery stenosis in patients scheduled for cardiac surgery. Echocardiography. 2017;34(5):716–722. doi:10.1111/echo.13507.

[37] Bakker D., Kazantzis N., Rickwood D., Rickard N. Mental health smartphone apps: review and evidence-based recommendations for future developments. JMIR Ment. Health. 2016;3(1):e7doi:10.2196/mental.4984.

[38] Barak A., Hen L., Boniel-Nissim M., Shapira N. A comprehensive review and a meta-analysis of the effectiveness of internet-based psychotherapeutic interventions. J. Technol. Hum. Serv. 2008;26(2–4):109–160. doi:10.1080/15228830802094429.

[39] See H.-Q., Chan J.-N., Ling S.-J., Gan S.-C., Leong C.-O., Mai C.-W. Advancing pharmacy service using big data—are we fully utilising the big data's potential yet?. J. Pharm. Pharm. Sci. 2018;21:217–221. doi:10.18433/jpps29869.

[40] Rudrapal M., Khairnar S.J., Jadhav A.G. Drug Repurposing (DR): An Emerging Approach in Drug Discovery. 2020.doi:10.5772/intechopen.93193.

[41] Shaw R., Kim Y.k., Hua J. Governance, technology and citizen behavior in pandemic: lessons from COVID-19 in East Asia. Prog. Disaster Sci. 2020;6:100090doi:10.1016/j.pdisas.2020.100090.

[42] Son W.S. Drug discovery enhanced by artificial intelligence. Biomed. J. Sci. Tech. Res. 2018;12(1):doi:10.26717/bjstr.2018.12.002189.

[43] Mishra R., Li B. The application of artificial intelligence in the genetic study of Alzheimer's disease. Aging Dis. 2020;11(6):1567. doi:10.14336/ad.2020.0312.

[44] Paranjpe M.D., Taubes A., Sirota M. Insights into computational drug repurposing for neurodegenerative disease. Trends Pharmacol. Sci. 2019;40(8):565–576. doi:10.1016/j.tips.2019.06.003.

[45] Rodríguez-Rodríguez I., Rodríguez J.-V., Shirvanizadeh N., Ortiz A., Pardo-Quiles D.-J. Applications of artificial intelligence, machine learning, big data and the internet of things to the COVID-19 pandemic: a scientometric review using text mining. Int. J. Environ. Res. Public Health. 2021;18(16):8578. doi:10.3390/ijerph18168578.

[46] Levin J.M., Oprea T.I., Davidovich S., Clozel T., Overington J.P., Vanhaelen Q., Cantor C.R., Bischof E., Zhavoronkov A. Artificial intelligence, drug repurposing and peer review. Nat. Biotechnol. 2020;38(10):1127–1131 10.1038/s41587-020-0686-x.

[47] Liu Z., Chen X., Carter W., Moruf A., Komatsu T.E., Pahwa S., Chan-Tack K., Snyder K., Petrick N., Cha K., Lal-Nag M., Hatim Q., Thakkar S., Lin Y., Huang R., Wang D., Patterson T.A., Tong W. AI-powered drug repurposing for developing COVID-19 treatments. Ref. Mod. Biomed. Sci. 2022 https://doi.org/10.1016/b978-0-12-824010-6.00005-8.

[48] Mucke H.A., Mucke E. Sources and targets for drug repurposing: landscaping transitions in therapeutic space. Drug Repurpos. Rescue Reposition. 2015;1(1):22–27 10.1089/drrr.2015.0001.

[49] Gardner S., Das S., Taylor K. AI enabled precision medicine: patient stratification, drug repurposing and combination therapies. Artificial Intelligence in Oncology Drug Discovery and Development. IntechOpen; 2020 https://doi.org/10.5772/intechopen.92594.

[50] Kumavath R., Paul S., Pavithran H., Paul M.K., Ghosh P., Barh D., Azevedo V. Emergence of cardiac glycosides as potential drugs: current and future scope for cancer therapeutics. Biomol. Ther. 2021;11(9):1275. doi:10.3390/biom11091275.

[51] Richardson P., Griffin I., Tucker C., Smith D., Oechsle O., Phelan A., Rawling M., Savory E., Stebbing J. Baricitinib as potential treatment for 2019-nCoV acute respiratory disease. Lancet. 2020;395(10223):e30–e31. doi:10.1016/s0140-6736(20)30304-4.

[52] Milano M., O’Sullivan B., Gavanelli M. Sustainable policy making: a strategic challenge for artificial intelligence. AI Mag. 2014;35(3):22–35. doi:10.1609/aimag.v35i3.2534.

[53] Majeed A., Hwang S.O. Data-driven analytics leveraging artificial intelligence in the era of COVID-19: an insightful review of recent developments. Symmetry. 2021;14(1):16. doi:10.3390/sym14010016.

[54] Crawford K., Calo R. There is a blind spot in AI research. Nature. 2016;538(7625):311–313. doi:10.1038/538311a.

[55] Selten F., Meijer A. Managing algorithms for public value. Int. J. Public Adm. Digit. Age. 2021;8(1):1–16. doi:10.4018/ijpada.20210101.oa9.

[56] Hanafi M.M., Kshetri N., Sharma R., Kshetri N. Economics of artificial intelligence in the Gulf cooperation council countries. Computer. 2021;54(12):92–98. doi:10.1109/mc.2021.3113094.

[57] Shah P., Kendall F., Khozin S., Goosen R., Hu J., Laramie J., Ringel M., Schork N. Artificial intelligence and machine learning in clinical development: a translational perspective. npj Digit. Med. 2019;2(1):doi:10.1038/s41746-019-0148-3.

[58] Nordling L. A fairer way forward for AI in health care. Nature. 2019;573(7775):103–105. doi:10.1038/d41586-019-02872-2.

Chapter 2: General considerations on artificial intelligence

Abhay Dharamsia; Archana Mohit Navaleb; Sunil S. Jambhekarc    a Department of Pharmaceutics, Parul Institute of Pharmacy, Parul University, Vadodara, India

b Department of Pharmacology, Parul Institute of Pharmacy, Parul University, Vadodara, India

c School of Pharmacy, LECOM Bradenton Campus, Bradenton, FL, United States

Abstract

Machine learning is a subset of the umbrella term artificial intelligence (AI). AI has already crept into several tasks of our day-to-day life, like digital assistants, internet surfing, online shopping, etc. Machine learning (ML), as the name indicates, is a way (algorithm) of self-learning by computer. The development of ML algorithms originated from the quest of computers that learn on their own based on their experiences. The learning takes place with the help of a dataset provided to the computer as training data. It basically helps in decision making or prediction of an outcome when the situation is having manifold factors and when decision making is not straightforward as per human intelligence. Drug discovery and delivery is a complicated process requiring a lot of human aptitudes and decision-making ability. The process is characterized by abundant data handling with multiple variables, thus making it amenable to the application of ML. Opportunities for the application of ML occur at nearly all stages of drug discovery, like target identification and validation, compound screening, lead identification and optimization, preclinical development, clinical trials, and biomarker identification and analysis. However, for the effective application of ML, its basic understanding is inevitable. The knowledge and technology about ML in healthcare are advancing considerably. Various software libraries are available online that can work with a range of hardware, even simple personal computers. Proper understanding and selection of an appropriate machine learning approach may provide accurate predictions. This chapter will provide various ML approaches and their areas of applications with suitable examples. Several ambiguities in the available methods of ML are cropping up as these are being applied to actual situations in the healthcare sector. However, scientists are also coming up with new techniques better suited to the area of healthcare. Deep learning is an approach apt for complex drug discovery data. However, the level of the algorithm to be generated is more complex in this approach. Another challenge in the application of ML to drug discovery is the availability of sufficient, accurate data to be fed for training. Generation of data itself may be a costly affair in certain phases of drug development. Although, there are few bottlenecks still to be resolved before ML can be applied full-fledged in drug discovery. The lack of repeatability and interpretability of the data generated by ML is posing a challenge for its accountability and reliability in different situations like the approval processes and IPR. There are methods that are proposed by scientists to improve the fitting of ML models and improve its output data, but, a great deal of work is required to be done in this area. The purpose of this chapter is to provide a basic understanding of ML concepts. This chapter is expected to bring more clarity to the potential applications of ML in drug design and development. We anticipate providing an overall view of the ML methods, their applications, and limitations so that aspirant researchers can be benefited.

Keywords

Supervised learning; Reinforcement learning; Deep neural network; Deep learning; Artificial neural network

Machine learning is a subset of the umbrella term artificial intelligence (AI). AI has already crept into several tasks of our day-to-day life, like digital assistants, internet surfing, online shopping, etc. Machine learning (ML), as the name indicates, is a way (algorithm) of self-learning by computer. The development of ML algorithms originated from the quest of computers that learn on their own based on their experiences. The learning takes place with the help of a dataset provided to the computer as training data. It basically helps in decision making or prediction of an outcome when the situation is having manifold factors and when decision making is not straightforward as per human intelligence. Drug discovery and delivery is a complicated process requiring a lot of human aptitudes and decision-making ability. The process is characterized by abundant data handling with multiple variables, thus making it amenable to the application of ML. Opportunities for the application of ML occur at nearly all stages of drug discovery, like target identification and validation, compound screening, lead identification and optimization, preclinical development, clinical trials, and biomarker identification and analysis. However, for the effective application of ML, its basic understanding is inevitable. The knowledge and technology about ML in healthcare are advancing considerably. Various software libraries are available online that can work with a range of hardware, even simple personal computers. Proper understanding and selection of an appropriate machine learning approach may provide accurate predictions. This chapter will provide various ML approaches and their areas of applications with suitable examples. Several ambiguities in the available methods of ML are cropping up as these are being applied to actual situations in the healthcare sector. However, scientists are also coming up with new techniques better suited to the area of healthcare. Deep learning is an approach apt for complex drug discovery data. However, the level of the algorithm to be generated is more complex in this approach. Another challenge in the application of ML to drug discovery is the availability of sufficient, accurate data to be fed for training. Generation of data itself may be a costly affair in certain phases of drug development. Although, there are few bottlenecks still to be resolved before ML can be applied full-fledged in drug discovery. The lack of repeatability and interpretability of the data generated by ML is posing a challenge for its accountability and reliability in different situations like the approval processes and IPR. There are methods that are proposed by scientists to improve the fitting of ML models and improve its output data, but, a great deal of work is required to be done in this area. The purpose of this chapter is to provide a basic understanding of ML concepts. This chapter is expected to bring more clarity to the potential applications of ML in drug design and development. We anticipate providing an overall view of the ML methods, their applications, and limitations so that aspirant researchers can be benefited.

2.1: The introduction of AI and its importance in pharmaceutical operations

The word intelligence means an ability to learn and solve problems. Thus, artificial intelligence (AI) is the technique where we generate machines which can learn on their own. The concept of AI is not very new. The term artificial intelligence was coined by John McCarthy in 1956, during a conference. However, till many years, there was no significant progress made in this area due to limited computational technology. In past few decades, development of cloud computing and other hardware and software advancements has resulted in considerable progress in AI.

Machine learning is an area of artificial intelligence, where machines learn (predict) tasks based on previous experience (data). Basically, machine learning algorithms are developed using two types of datasets. Training dataset is used to train the model. The type of data used in this set can vary as per the type of ML approach used as describe below. After training a model, it is validated using another dataset. If the validation parameters are acceptable, it can be utilized for evaluating actual data.

Machine learning approaches fall in one of the three categories:

1.Supervised learning: It is basically a classification task performed by machine. Here, the machine is trained using a labeled data. Based on such labeled training dataset, it learns the features of data. It uses this information to predict the label of data when it is provided with test data with known features. Fig. 2.1 depicts the concept with an example. Support vector machines and deep neural networks are the examples of supervised learning algorithms.

Fig. 2.1 Supervised learning approach in ML. The system is fed with training data with labels, rose and lotus. The algorithm learns the features necessary to identify a lotus or a rose. It uses this learning to predict the type of flower when provided with test data. The input as well as output here is labeled.

2.Unsupervised learning: In this type of ML, a machine performs a clustering task. The training data used here is not labeled. The algorithm is fed with mixed data without any labels. It predicts similar patterns in the data, and clusters data with similar features. Fig. 2.2 depicts the concept of unsupervised learning, where the algorithm is provided with images of roses and lotuses, and divides the images into two clusters with similar features. Dimension reduction method like PCA (principal component analysis) is an example of this type of ML approach.

Fig. 2.2 Unsupervised learning approach in ML. The system is fed with training data without any labels. The algorithm learns the features necessary to differentiate the flowers. It uses this learning to cluster the flowers with similar features. Neither the input nor the output is labeled in this type of approach.

3.Reinforcement/sequential learning: Here, the agent learns to make decisions as it takes up a different action. If it achieves desired output, it gets a reward. Based on reward, it learns which action is appropriate for which type of input. Thus, here there is a reinforcement of whatever is learned by the algorithm as it works through the data. That is why it is also called reinforcement learning [1]. Multiarmed bandit (MAB) algorithms are examples of this type of learning. In MABs, a given set of actions, called arms is available. The agent interacts with the environment by selecting an arm at a time. This interaction generates some observation, which is denoted as a reward. Interaction with different arms produces different levels of reward. The agent identifies the arm which produces an optimum average reward suitable to achieve the goal (Fig. 2.3).

Fig. 2.3 A Multi Armed Bandit (MAB) algorithm, an example of sequential learning. The agent selects any one arm in the environment and gets rewarded. It learns to select appropriate arm for a given environment to achieve best reward. (From C. Réda, E. Kaufmann, A. Delahaye-Duriez, Machine learning applications in drug development, Comput. Struct. Biotechnol. 18 (2020) 241–252, https://doi.org/10.1016/j.csbj.2019.12.006.)

The progress made in AI techniques has created a lot of buzz in past decades. The concept is finding its applications in almost all fields. Sciences like engineering, space science, earth science, biotechnology, etc. are utilizing AI advances. Medical and pharmaceutical science are among the areas where AI applications can bring about revolutionary changes. AI has been applied for disease risk prediction, diagnosis and prognosis estimation [2–5]. Pharmaceutical operations like granulation, mixing, compression, etc. can be optimized using AI algorithms [6]. Park et al. summarized various mechanism based modeling techniques applicable to granulation and compression processes. Various modeling principles were evaluated and their predictive values for the intermediate product quality were derived. Application of an appropriate modeling may improve ongoing manufacturing process within shorter time frame. Integration of data science with Process Analytical Technologies (PAT) can improve the outcome of continuous manufacturing line. Roggo et al. developed a DNN (deep neural network) for a continuous manufacturing line for a solid dosage form. In this study, seven critical process parameters and eight quality attributes were identified. A deep learning technique was applied to the process, where process parameters were changed and their impact on quality attributes was recorded. DNN was developed to reduce noise and improve data interpretation. A calibration error value less than 10% was achieved for Active Pharmaceutical Ingredient (API) content and two other quality parameter estimation [7].

Quality-by-Design (QbD) constitutes a well-defined roadmap of the process to ensure final product quality. The concept of artificial neural network (ANN) can be applied to QBD to achieve desired in vitro and in vivo product property. Simõesa et al. developed a feed-forward neural network (FFNN) for setting process parameter limits and material specifications [8]. The product was then tested for in vitro as well as in vivo drug release. It was evaluated in two clinical studies, where it was found to be bioequivalent to the Reference Listed Drug. FFNN is a basic type of DNN, where input just travels forward. There is no reverse journey of signal in the FFNN. Thus, there is no characteristic of loop formation or memory in this type of neural network. This DNN fails to process sequential data. It processes each input data as an independent query and generates output that is not affected by previous data handling. Many applications of FFNNs have been reported, e.g., for modeling drug release from pellets [9], the impact of process parameters for hot melt extrusion process on vaginal film quality [10], the estimation of polymeric nanoparticles size as a function of polymer properties [11], prediction of biophysical features of monoclonal antibodies based on amino acid composition [12], the impact of excipient selection on ejection force of tablets [13] and establishing an in vitro-in vivo correlation for a dry powder inhaler [14].

Some operations in the pharmaceutical manufacturing are based on visual inspection. Evaluation of tablet coating is one such example, where human intervention is considered indispensable. However, image analysis using various classification techniques has been applied to classify tablets with different categories of coating [15]. Mehle et al. also developed a convolutional neural network for differentiation between groups of primary particles or agglomerates in images of pellets taken during coating process [16]. Knowledge of physicochemical and biopharmaceutical properties of drug is necessary in initial stages of formulation design. Studies which are conducted to know about these properties of API are known as preformulation studies. AI can be utilized for prediction of such properties of drugs to ease formulation development. Several studies have described use of ANNs at preformulation stage. The artificial neural network developed by Ebube et al. was trained, validated, and evaluated for prediction of glass transition temperatures, water uptake, and viscosities of various amorphous polymers and their mixtures [17]. The technique was able to predict the properties with good accuracy (error <8%). Another study also used a boosted tree classifier to understand the correlation between selected API characteristics and tensile strength of the tablets. It was found that particle diameter, moisture content, partition coefficients, and modal diameter of APIs were among the most important factors affecting tensile strength of tablets [18].

2.2: Role of ML in drug design and drug delivery

2.2.1: Artificial intelligence in drug design

Drug molecule design is an important task of drug development. The process starts with the identification of a valid target, based on the literature available. AI can contribute very well in this step. Identification of a druggable target is crucial in the drug discovery process. AI can be applied to review the literature regarding disease pathophysiology. Based on the data gathered, a battery of targets can be identified, which can be further studied. Bakkar et al. used IBM Watson to identify RNA binding proteins that are altered in Amyotrophic lateral sclerosis (ALS). It thus identified five RNA binding proteins namely, RBMS3, Syncrip, hnRNPU, NUPL2, and Caprin-1, which were previously no known linkage to ALS [19]. Anticancer drug discovery probably finds the best application of Machine learning. The disease is highly heterogeneous in nature and several types of cancers remain incurable due to a lack of target information. Tumor neoantigens are attractive targets for the immunotherapy of cancer. However, meticulous discovery and evaluation required a lot of time and resources. A reduction in time and resources was possible due to machine learning approaches to identify mutated peptides exhibiting high-affinity binding with HLA molecules. Ott et al. developed a personalized vaccine against melanoma targeting neoantigens. They also expressed a potential of ML applications for discovery and development of such vaccines to suit the individual patient requirement [20].

The design of a drug candidate that fits the selected target (mostly protein) molecule is the next step in the process. This is accomplished via techniques like computational design and molecular docking studies. The data of such previous trials can act as training data for the development of general models. General models must be able to represent wide molecular space, at the same time, they should be able to adapt to a small and specific molecular space depending on the target of interest. Therefore, general models need to be optimized to adapt to the goal. Optimization of the general model can be done by applying reinforcement learning, where designed compounds are analyzed (in silico) for their desired properties. This match (or mismatch) generates reward, which reinforces learning in the model. Later on, such a model will suggest better-optimized compounds with desired properties as defined earlier.

Peptide drugs are attractive tools for therapeutic target modulation. In spite of ample opportunities for protein drug development, the actual drugs hitting the market are fewer. This is due to challenges associated with their design, formulation, pharmacokinetics, and stability, etc. [21]. The application of AI for solving these problems can be useful. An approach called GAN (Generative Adversarial Networks) has been applied by several researchers for protein drug discovery [22,23]. This approach uses two neural networks (NNs) a generator and a discriminator. Generator neural network is trained to generate instances, e.g., DNA sequences. Discriminator neural network discriminates these instances based on predefined properties of interest. It scores the instances based on the probability of their satisfying the desired characteristics. Fig. 2.4 illustrates the concept of the approach in a graphical manner. Gupta and Zou developed a Feedback GAN which can predict DNA sequences coding for a protein with antimicrobial activity (Property of interest). GAN generated proteins had acceptable biological and physical properties [22,23]. Such an approach can be applied to generate peptide structures with other desirable properties.

Fig. 2.4

Fig. 2.4 GAN (Generative Adversarial Networks) approach for protein drug discovery. Generator Neural network generates instances (e.g., chemical structures). The instances are fed to GNN, which evaluates instances to fit the predecided requirements (e.g., receptor affinity). The score such generated is called goodness score, which is sent back to GNN as feedback. GNN learns from the feedback and later on is able to generate instances which have best desired properties. (From C. Réda, E. Kaufmann, A. Delahaye-Duriez, Machine learning applications in drug development, Comput. Struct. Biotechnol. J. 18 (2020) 241–252, https://doi.org/10.1016/j.csbj.2019.12.006.)

For rapid drug discovery, a large group of chemicals needs to be virtually screened for their desired bioactivity. The source for chemical structures can be physical or virtual chemical libraries. However, such libraries have a finite number of chemicals. An alternative source for chemical structures is automatic chemical design [24]. Automatic chemical design is a framework, which can be used to design chemical compounds with desired properties using machine learning techniques. The approach used for this is known as BO (Bayesian Optimization). The major drawback of the BO approach is that it generates many invalid molecular structures. BO uses an environment that is a latent space and an algorithm called variational autoencoder (VA). VA samples points from latent space and generate chemical structures. Invalid chemical structures are generated when it samples points far away from the data points which were used to train the model. Griffiths et al. propose a method to restrict point sampling from latent space [25,26]. This approach is called constrained Bayesian Optimization. Recurrent neural network (RNN) trained using constrained Bayesian Optimization generated fewer invalid structures. Moreover, the druggability score of such compounds was also higher.

Apart from molecules designed using the above applications of ML, there are various chemical libraries that provide small molecules for screening. A number of free as well as commercial libraries are available. Detailed information on types of library, factors of selection and primary methods to use them is available in the literature [27–30].

Once a dataset of compounds is available either from autoencoder or from libraries, the next step is screening of these compounds to assess their biological activity. Advancement in instrumentation and high throughput screening (HTS) methods have brought about a drastic reduction in the time and expenditure in this step. However, in spite of these advantages, when it comes to screening thousands of compounds, the process is still labor intensive. In such scenario, virtual screening of compounds provides a vital alternative or rather a preceding step, to screen out most of the noneligible compounds, thus reducing the number of compounds to be screened by HTS. The branch of bioinformatics that deals with application of computational technique, in the field of chemistry is known as chemoinformatics. Most of the methods of virtual screening (VS) are based on chemoinformatics. Apart from predicting best drug candidates for selected targets, VS also has other applications, such as, prediction of ATC codes for known drugs [31,32] and finding beneficial drug combinations [33], drug repurposing and side effect prediction based on off target binding profile [34]. VS methods use compounds and targets for prediction of interaction between them. For this, compounds and targets have to be fed into the program using proper descriptors. Based on the type of descriptors used for input data, virtual screening methods can be one of the three types:

1.Structure based screening: uses 3D structures to model interaction between compound and target

2.Ligand based screening: uses compound properties to model interaction with target

3.Proteochemometric (PCM) approach: uses nonstructural descriptors of both protein (target) and chemical (test compound) as input

A detailed account of structure-based and ligand-based screening can be obtained from reviews published by Geppert et al. [35], Lavecchia and Giovanni [36], and Glaab [37]. Structure-based virtual screening methods can be applied only when 3D structures of both target and test compound are known. The 3D structure may be predicted practically (using X-ray crystallography or NMR) or virtually (by homology modeling). The interaction between target and test compound can be predicted by virtual screening which provides binding affinity. The output is in the form of affinity scores and binding confirmations [38,39]. DOCK [40], AutoDOCK [41], GOLD [42], Fred [43], Glide [44], and FlexX [45] are some of the commonly used docking tools. Apart from this, similarity-based docking tools such as HomDock [46], eSimDock [47], and fkcombu [48] are available apart from traditional methods. However, the major limitation for the application of a structure-based approach is the lack of availability of 3D structures of compounds or targets. Libraries and tool kits are important resources needed for virtual screens. These are required for various tasks during virtual screening such as, deriving molecular descriptors, conversion of one format of descriptor representation to another format, calculating similarities between pairs of molecules and applying various statistical methods and machine learning algorithms. Table 2.1 enlists various toolkits and libraries useful for chemoinformatics.

Table 2.1

Modified from A.S. Rifaioglu, H. Atas, M.J. Martin, R. Cetin-Atalay, V. Atalay, T. Doğan, Recent applications of deep learning and machine intelligence on in silico drug discovery: methods, tools and databases, Brief. Bioinform. 20(5) (2019) 1878–1912, https://doi.org/10.1093/bib/bby061.

Various toolkits are used in chemoinformatics which allow following functions:

•Interpret and convert one chemical structure format to another

•Substructure matching

•Matching the structural similarity

•MCS, maximal common structure prediction

•Generating virtual fragments of molecules by splitting them

•Build molecules from fragments or substructures

•Predict output reaction product by applying reactions on input reactant structures

•Generation of molecular fingerprints for indexing in chemical databases

2.2.2: Databases for virtual screening

Virtual screening or rather any ML technique needs a good amount of data as input. There are several such datasets available that provide compounds, substances, target information and interaction. Such open access datasets are described in Table 2.2. Apart from these datasets, there are also paid tools and databases, especially for commercial drug development programs. Information about target which is mostly protein, can be obtained from protein information databases. These databases provide information about structure of protein, as well as their functional properties. Some of these databases include information that is curated manually from literature by expert scientists. Apart from this, most databases provide cross references to third party data also.

Table 2.2

Modified from A.S. Rifaioglu, H. Atas, M.J. Martin, R. Cetin-Atalay, V. Atalay, T. Doğan, Recent applications of deep learning and machine intelligence on in silico drug discovery: methods, tools and databases, Brief. Bioinform. 20(5) (2019) 1878–1912, https://doi.org/10.1093/bib/bby061.

2.2.2.1: Gold-standard datasets

The term gold-standard datasets in ML means a trusted set of information, generally curated manually that can be used for purposes like training and testing of ML models, optimization of parameters of developed ML model, evaluation of the performance of the developed model, provide a benchmark to compare between the performance of various prediction models etc. The selection of a proper dataset for training the developed model is a crucial step in the process. Until 10 years ago, there were no sufficient dataset resources available. The first large dataset was created by Hert et al. [49]. It included 11 activity classes and 300–1200 compounds in each class. However, it was not available in the public domain. Later on, Yamanishi et al. [50] created a truly large and freely available dataset that had curated data across four classes of drug targets, i.e., 664 enzymes, 204 ion channels, 95 GPCRs, and 26 nuclear receptors. The drug-target interactions for each class were 2926, 1476, 635, and 90, respectively. It can be accessed via http://web.kuicr.kyoto-u.ac.jp/supp/yoshi/drugtarget/. An updated version of this database was also created by Yamanishi et al. in 2010 with a greater number of targets and DTIs [51]. This dataset is suitable and is widely used for network or graph-based approaches. However, it is not suitable for ML approaches that require large training datasets. DUD-E (Directory of Useful Decoys) developed by Huang, Irwin, and Shoichet [52], MUV (Maximum Unbiased Validation) [53], Kaggle-Merck activity dataset developed in 2012 [54], Tox21 developed via a community data challenge in 2014 [55], and MoleculeNet developed using a novel approach [56] are some of such gold-standard datasets used for developing and testing ML algorithms.

Proteochemometric (PCM) approach uses physicochemical or molecular features of the compound or target [57]. These features are used to train the model using supervised learning approach [58]. Usually, PCM uses nonlinear ML techniques such as support vector machines (SVM), random forest (RF), or linear methods such as partial least squares (PLS) [59]. The requirements for individual methods may vary. In all methods, compounds or target information is gathered from a database. Based on this information, a feature vector is generated for each compound. The feature vector for compound as well as target is fed into ML algorithm as training dataset to generate a predictive model. Following this, when a query compound is submitted to the algorithm, it can predict its bioactivity against a given target. A method SPACE developed by Liu et al. uses a combination of similarity based and feature based approach to predict ATC code of query compound [32]. This program is available at http://www.bprc.ac.cn/space.

2.2.2.2: Applications of DNNs in VS

Deep Neural Networks are algorithms that employ more than two layers of neurons. The uppermost layer is known as output layer and deeper most layer is input layer. There can be multiple neural layers between these two layers. The earlier networks with ML were unable to handle large and complex data, which has become possible with DNNs. Recently scientists are also able to resolve some drawbacks associated with DNNs like, increasing amount of error as data passes through each layer, inability to explain the logic behind output produced etc. Several types of DNNs are applied by scientists to solve different virtual screening problems. Here we take a few examples of such DNNs.

Feed-forward DNNs

As discussed earlier, these are the basic types of DNNs, where input just travels forward. There is no reverse journey of signal in the DNN. Dahl et al. developed a feed-forward DNNs (FFDNNs), to predict activity in multiple assays. The model was trained and tested in 19 different biochemical and cell based assays. Moreover, a recently developed strategy was applied to avoid overfitting of the model. The model performed best in 14 out of 19 assays, when compared to the results obtained using Random Forest (RF), Gradient Boosted Decision Tree Ensembles (GBM), and Single Task Neural Networks (NNET) approaches [60].

Recurrent neural networks

As discussed earlier, the major drawback of FFDNN was that the program does not consider previous input. To overcome this limitation RNNs were developed. In these algorithms, the output generated is influenced by previous results. Such algorithms are suitable to handle sequential data. Goh et al. developed an RNN to predict properties of compounds like toxicity, activity, solubility, and solvation energy based on SMILEs format of compounds [61]. RBM (Restricted Boltzmann Machine) is a two-layered graphical model. In their study, Wang and Zeng employed RBMs for drug target interaction prediction [62]. They constructed an RBM network for each target. For training the models, compound and target information was obtained from MATADOR and STITCH databases. Multiple RBMs were stacked to generate a deep belief network (DBN). Thus, a DBN had multiple RBMs, one for each target activity prediction. The performance of the model was evaluated against a logistic regression model applied at each layer. The performance was shown to improve with increasing depth of hidden layers. The method mentioned above was based on SMILEs input. There was a lack of a method that can process graphical input data. Such as 3D structures of compounds and proteins. Convolutional neural network (CNN) offered a useful approach to process graphical input data and predict the drug target interaction. Each CNN has sequential layers of convolutional and pooling modules. Convolutional modules extract patterns from the query structure. Pooling modules are used for subsample and feature reduction. Goh et al. [61], Gonczarek et al. [63], and Wallach et al. [64] employed 3D structure of binding sites or binding pockets and 3D/2D structures of compounds to predict their drug target interaction through CNN [65]. Altae-Tran et al. [66] developed a Graph Convolutional Neural Network (GCNN) by employing a convolutional approach to graphical data. They used 2D graphs of compounds for drug target interaction prediction.

2.3: Application of AI in biomedical and tissue engineering

The healthcare field is rapidly evolving, generating great opportunities and needs for applications of AI. With the advent of modern medical services and evolving knowledge about multifaceted disease course, it becomes mandatory to deal with a large amount of patient data. Artificial intelligence is offering multiple facets for making the process of patient management efficient. Biomedical engineering is foremost in utilizing the potential of AI via techniques like the Internet of Things (IoT). IoT is an emerging concept to establish communication between connected devices. Several biomedical scientists have developed systems comprising of sensors, data transmitters, internet servers, and physician's device to continuously monitor patient parameters like body temperature, heartbeats, oxygen levels, CO2 levels, room temperature, humidity, etc. [67]. A complete IoT-based patient monitoring setup described by Wan et al. includes wearable sensors (e.g., sensors for ECG, BP sensors), GPS sensors for positioning and localization of patient, an RFID for patient identification. The second layer includes a mechanism for data transmission based on protocols such as ZigBee, 6LowPAN, NB-IoT, and LoRa. The next layer consists of a data processing layer which may be based on data-driven approach, knowledge-driven approach, or a hybrid approach. The uppermost layer consists of an application layer which allows for different modes of management such as self-management, professionally assisted management, passive assisted management or emergency management [68]. Sensors useful for measuring blood glucose level, ECG, EMG, temperature, muscle activity, rate of respiration, heartbeat, blood pressure, etc. can be invaluable in monitoring patients with different diseases like arrhythmias, epilepsy, hypertension, diabetes mellitus, etc. Biomedical science is evolving continuously with compact, rugged and accurate sensor systems. The data transmission and processing technologies are improving with the better availability of smartphones, data transmission protocols and cloud technologies [69]. The second and third layer of the module is becoming gradually efficient. The final layer for data application may be a personally supervised dataset monitored by physicians, e.g., nursing staff. This may alert the healthcare provider for any emergency situation. In situations, such as diabetes mellitus passive patient assistance may be offered based on a previously decided glucose management plan. In other situations, like ECG, etc., professionally assisted remote management can be offered to the patient. In diseases like hypertension or obesity, patient compliance to lifestyle modification advice may be monitored via this approach