Artificial intelligence (AI) is the ability of machines or systems to perform tasks that normally require human intelligence, such as reasoning, learning, decision making, and problem solving. AI can be applied in various fields, such as education, entertainment, finance, security, and health care. In particular, AI has the potential to revolutionize the field of drug development, which is the process of discovering, testing, and bringing new drugs to the market.
Drug development is a complex, costly, and time-consuming process that involves many steps, such as target identification, lead optimization, preclinical testing, clinical trials, and regulatory approval. According to a study by the Tufts Center for the Study of Drug Development, the average cost of developing a new drug is about $2.6 billion, and the average time from discovery to approval is about 10 years ¹. Moreover, the success rate of drug development is very low, as only about 12% of the drugs that enter clinical trials are eventually approved by the Food and Drug Administration (FDA) .
AI can help overcome these challenges by offering improved efficiency, accuracy, and speed in the drug development process. AI can leverage large amounts of data, such as genomic, proteomic, chemical, and clinical data, and use advanced algorithms, such as deep learning and machine learning, to analyze, model, and predict various aspects of drug development. For example, AI can help identify new targets, design novel compounds, optimize drug properties, predict drug efficacy and safety, and reduce the need for animal testing. AI can also help streamline the drug development pipeline and enhance the success rate of clinical trials, ultimately resulting in the emergence of more effective and safer drugs.
Understanding Artificial Intelligence
AI is a broad term that encompasses different types and levels of intelligence. One way to classify AI is based on the degree of human-like intelligence that the system exhibits. According to this classification, there are three types of AI: narrow AI, general AI, and super AI .
- Narrow AI, also known as weak AI, is the type of AI that can perform specific tasks that require a limited amount of intelligence, such as playing chess, recognizing faces, or translating languages. Narrow AI is the most common and developed type of AI today, and it is used in various applications, such as search engines, voice assistants, and self-driving cars.
- General AI, also known as strong AI, is the type of AI that can perform any task that a human can do, such as reasoning, learning, planning, and creativity. General AI is the ultimate goal of AI research, but it is still far from being achieved, as it requires the system to have a comprehensive understanding of the world and human cognition.
- Super AI, also known as artificial superintelligence, is the type of AI that can surpass human intelligence in all aspects, such as knowledge, wisdom, and creativity. Super AI is a hypothetical and controversial concept, as it raises many ethical and existential questions, such as whether super AI would be benevolent or malevolent, and whether humans would be able to control or coexist with super AI.
Another way to classify AI is based on the approach or technique that the system uses to achieve intelligence. According to this classification, there are two main types of AI: symbolic AI and sub-symbolic AI.
- Symbolic AI, also known as classical AI or rule-based AI, is the type of AI that uses symbols, such as words, numbers, or logic, to represent and manipulate knowledge. Symbolic AI relies on predefined rules and algorithms to perform tasks, such as inference, deduction, and search. Symbolic AI is suitable for tasks that have clear and structured problems and solutions, such as mathematics, logic, and chess.
- Sub-symbolic AI, also known as connectionist AI or neural network-based AI, is the type of AI that uses sub-symbols, such as neurons, synapses, or activation functions, to represent and process knowledge. Sub-symbolic AI relies on learning from data and experience to perform tasks, such as classification, recognition, and prediction. Sub-symbolic AI is suitable for tasks that have complex and unstructured problems and solutions, such as vision, speech, and natural language.
One of the most popular and powerful techniques of sub-symbolic AI is deep learning, which is a subset of machine learning. Machine learning is the branch of AI that enables the system to learn from data and improve its performance without explicit programming. Deep learning is a type of machine learning that uses multiple layers of artificial neural networks, which are computational models inspired by the structure and function of biological neurons, to learn from data and extract features and patterns. Deep learning can handle large and high-dimensional data, such as images, videos, and texts, and achieve state-of-the-art results in various tasks, such as object detection, face recognition, natural language processing, and speech recognition.
Artificial Intelligence in Drug Discovery
AI can be applied in various stages of drug discovery, which is the process of identifying and optimizing new compounds that can modulate a biological target, such as a protein, gene, or pathway, that is involved in a disease. The main stages of drug discovery are target identification, target validation, hit identification, hit-to-lead optimization, lead optimization, and preclinical testing .
- Target identification is the stage of finding and selecting a potential target that is relevant to the disease of interest. AI can help in this stage by analyzing large and diverse data sources, such as genomic, transcriptomic, proteomic, metabolomic, and phenotypic data, and identifying novel targets, biomarkers, and pathways that are associated with the disease. For example, AI can use network analysis, causal inference, or knowledge graphs to infer the relationships and interactions among different biological entities and discover new targets .
- Target validation is the stage of confirming and characterizing the role and function of the selected target in the disease. AI can help in this stage by predicting and simulating the effects of modulating the target on the disease phenotype, using techniques such as molecular docking, molecular dynamics, or pharmacophore modeling. For example, AI can use molecular docking to estimate the binding affinity and mode of a ligand to a target, or use molecular dynamics to simulate the conformational changes and interactions of a target-ligand complex.
- Hit identification is the stage of finding and screening initial compounds that can bind to and modulate the target. AI can help in this stage by generating and selecting novel compounds, using techniques such as de novo design, virtual screening, or generative adversarial networks. For example, AI can use de novo design to create new compounds from scratch, based on the desired properties and constraints, or use virtual screening to filter and rank a large library of compounds, based on their predicted activity and selectivity.
- Hit-to-lead optimization is the stage of improving and refining the initial compounds, using techniques such as structure-activity relationship (SAR) analysis, to obtain a smaller set of compounds that have higher potency, selectivity, and stability. AI can help in this stage by optimizing and modifying the compounds, using techniques such as multi-objective optimization, reinforcement learning, or genetic algorithms. For example, AI can use multi-objective optimization to balance and maximize multiple criteria, such as activity, solubility, and toxicity, or use reinforcement learning to learn from feedback and iteratively improve the compounds.
- Lead optimization is the stage of further enhancing and evaluating the compounds, using techniques such as quantitative structure-activity relationship (QSAR) analysis, to obtain a final set of compounds that have optimal pharmacokinetic and pharmacodynamic properties. AI can help in this stage by predicting and testing the compounds, using techniques such as deep neural networks, convolutional neural networks, or recurrent neural networks. For example, AI can use deep neural networks to predict the absorption, distribution, metabolism, excretion, and toxicity (ADMET) of the compounds, or use convolutional neural networks to recognize and classify the compounds based on their images .
Preclinical testing is the stage of validating and verifying the safety and efficacy of the compounds in animal models, before proceeding to clinical trials in humans. AI can help in this stage by reducing and replacing the need for animal testing, using techniques such as organ-on-a-chip, microphysiological systems, or digital twins. For example, AI can use organ-on-a-chip to mimic the structure and function of human organs, such as the heart, liver, lung, or kidney, on a microfluidic device that contains living cells and tissues . AI can also use microphysiological systems to integrate multiple organ-on-a-chip models and simulate the interactions and responses of different organs to the compounds. AI can also use digital twins to create virtual replicas of individual patients or populations and predict their outcomes based on their genetic, environmental, and behavioral data .
Deep Learning and Machine Learning in Drug Development
AI can also be applied in various stages of drug development, which is the process of testing and evaluating the compounds that have been discovered and optimized in the previous stage, in order to obtain regulatory approval and market authorization. The main stages of drug development are clinical trials, regulatory review, and post-marketing surveillance .
- Clinical trials are the stage of testing the safety and efficacy of the compounds in human subjects, under controlled and monitored conditions. AI can help in this stage by designing and optimizing the clinical trials, using techniques such as adaptive design, Bayesian statistics, or artificial neural networks. For example, AI can use adaptive design to modify the trial parameters, such as sample size, dose, or endpoint, based on interim data and feedback, to improve the efficiency and accuracy of the trial . AI can also use Bayesian statistics to incorporate prior knowledge and evidence, such as from preclinical studies or historical data, to update the probability and confidence of the trial outcomes. AI can also use artificial neural networks to model and predict the clinical outcomes, such as response, survival, or adverse events, based on the patient characteristics and biomarkers .
- Regulatory review is the stage of submitting and evaluating the data and evidence from the clinical trials, in order to obtain approval and authorization from the regulatory agencies, such as the FDA or the European Medicines Agency (EMA). AI can help in this stage by facilitating and accelerating the regulatory review, using techniques such as natural language processing, data mining, or decision support systems. For example, AI can use natural language processing to extract and analyze the information and documents from the clinical trials, such as protocols, reports, or labels, and generate summaries and insights for the reviewers ¹². AI can also use data mining to identify and compare the patterns and trends from the clinical trials, such as efficacy, safety, or quality, and detect any anomalies or discrepancies for the reviewers ¹³. AI can also use decision support systems to assist and advise the reviewers in making the approval decisions, based on the data and evidence from the clinical trials, as well as the regulatory guidelines and criteria .
- Post-marketing surveillance is the stage of monitoring and evaluating the safety and effectiveness of the compounds after they have been approved and marketed, in order to detect and prevent any adverse events or quality issues that may arise in the real-world setting. AI can help in this stage by enhancing and improving the post-marketing surveillance, using techniques such as pharmacovigilance, signal detection, or pharmacogenomics. For example, AI can use pharmacovigilance to collect and analyze the data and reports from the post-marketing sources, such as electronic health records, social media, or spontaneous reporting systems, and identify and assess any adverse events or quality issues that may be associated with the compounds . AI can also use signal detection to monitor and evaluate the signals of adverse events or quality issues from the post-marketing data and reports, and alert and notify the stakeholders, such as the manufacturers, regulators, or health professionals, for further investigation or action . AI can also use pharmacogenomics to study and understand the genetic variations and factors that may influence the response and reaction of the patients to the compounds, and enable personalized and precision medicine .
Challenges and Solutions
AI has many advantages and benefits in the field of drug development, but it also faces many challenges and limitations that need to be addressed and overcome. Some of the major challenges and solutions are:
- Data quality and availability: AI relies on large and diverse data sets to perform its tasks, but the data may not be always available, reliable, or consistent. For example, the data may be incomplete, inaccurate, or outdated, due to errors, biases, or changes in the sources or methods of collection and analysis. The data may also be fragmented, heterogeneous, or incompatible, due to differences in the formats, standards, or protocols of the data. The data may also be confidential, sensitive, or protected, due to ethical, legal, or regulatory restrictions or requirements. To address these challenges, the data quality and availability need to be improved and ensured, by using techniques such as data cleaning, validation, integration, standardization, or anonymization. The data also need to be shared and exchanged, by using platforms such as data repositories, databases, or networks, that can facilitate and enable the access and use of the data by the stakeholders, while respecting and protecting the data rights and privacy .
- Algorithm complexity and interpretability: AI uses complex and sophisticated algorithms to perform its tasks, but the algorithms may not be always transparent, explainable, or understandable. For example, the algorithms may be black-box, opaque, or nonlinear, due to the use of multiple layers, parameters, or functions that are difficult to trace or comprehend. The algorithms may also be unpredictable, uncertain, or inconsistent, due to the use of probabilistic, stochastic, or heuristic methods that are based on assumptions or approximations. The algorithms may also be biased, unfair, or unethical, due to the use of data or models that are not representative or inclusive of the population or context. To address these challenges, the algorithm complexity and interpretability need to be reduced and enhanced, by using techniques such as algorithm simplification, visualization, or verification. The algorithm also need to be explained and justified, by using techniques such as feature selection, attribution, or counterfactuals, that can provide and communicate the rationale and evidence of the algorithm outcomes and decisions .
- Human–machine interaction and collaboration: AI interacts and collaborates with humans to perform its tasks, but the interaction and collaboration may not be always effective, efficient, or satisfactory. For example, the interaction and collaboration may be difficult, inconvenient, or uncomfortable, due to the lack of common language, understanding, or trust between the humans and the machines. The interaction and collaboration may also be challenging, demanding, or stressful, due to the need of constant supervision, guidance, or feedback from the humans to the machines. The interaction and collaboration may also be disruptive, harmful, or detrimental, due to the potential conflicts, errors, or risks that may arise from the humans or the machines. To address these challenges, the human–machine interaction and collaboration need to be improved and optimized, by using techniques such as natural language processing, speech recognition, or gesture recognition, that can enable and facilitate the communication and exchange of information and instructions between the humans and the machines. The interaction and collaboration also need to be supported and assisted, by using techniques such as user interface, user experience, or user feedback, that can provide and enhance the usability, functionality, or satisfaction of the humans and the machines. The interaction and collaboration also need to be regulated and governed, by using techniques such as rules, policies, or standards, that can define and ensure the roles, responsibilities, or rights of the humans and the machines .
Practical Applications
AI has many applications and examples in the field of drug development, that demonstrate and illustrate its capabilities and potentials. Some of the practical applications are:
- Drug repurposing: AI can help in finding new uses or indications for existing or approved drugs, by using techniques such as similarity search, network analysis, or machine learning, to identify and predict the drug–target, drug–disease, or drug–drug interactions and associations. For example, AI was used to repurpose the antiviral drug remdesivir, which was originally developed for Ebola virus, to treat COVID-19, by using similarity search to find drugs that have similar chemical structures or molecular properties to the COVID-19 inhibitors . AI was also used to repurpose the antimalarial drug chloroquine, which was originally developed for malaria, to treat COVID-19, by using network analysis to find drugs that have similar biological pathways or molecular mechanisms to the COVID-19 modulators . AI was also used to repurpose the antidepressant drug fluoxetine, which was originally developed for depression, to treat COVID-19, by using machine learning to find drugs that have similar gene expression or cell response profiles to the COVID-19 suppressors .
Drug combination: AI can help in finding new combinations or synergies of drugs, by using techniques such as optimization, game theory, or deep learning, to identify and optimize the drug–drug interactions and effects. For example, AI was used to find new combinations of drugs for cancer, by using optimization to find the optimal doses and schedules of the drugs that can maximize the efficacy and minimize the toxicity . AI was also used to find new combinations of drugs for tuberculosis, by using game theory to find the best strategies and outcomes of the drugs that can overcome the drug resistance and heterogeneity . AI was also used to find new combinations of drugs for COVID-19, by using deep learning to find the most effective and safe combinations of drugs from a large pool of candidates .
- Drug synthesis: AI can help in finding new ways or methods of synthesizing or manufacturing the drugs, by using techniques such as retrosynthesis, reaction prediction, or generative models, to identify and optimize the chemical reactions and pathways that can produce the drugs. For example, AI was used to find new ways of synthesizing the antimalarial drug artemisinin, by using retrosynthesis to find the optimal sequence of reactions that can convert a simple starting material into the complex drug molecule . AI was also used to find new ways of synthesizing the antibiotic drug penicillin, by using reaction prediction to find the best conditions and catalysts that can facilitate the formation of the drug molecule . AI was also used to find new ways of synthesizing the anticancer drug paclitaxel, by using generative models to create novel and efficient synthetic routes that can reduce the cost and time of the drug production .
- Drug delivery: AI can help in finding new modes or systems of delivering or administering the drugs, by using techniques such as nanotechnology, biotechnology, or robotics, to design and optimize the drug carriers and devices that can enhance the drug delivery and absorption. For example, AI was used to find new modes of delivering the insulin drug for diabetes, by using nanotechnology to create smart nanoparticles that can sense the blood glucose level and release the drug accordingly . AI was also used to find new modes of delivering the gene therapy drug for cystic fibrosis, by using biotechnology to create engineered viruses that can deliver the correct gene to the defective cells . AI was also used to find new modes of delivering the pain relief drug for chronic pain, by using robotics to create implantable devices that can stimulate the spinal cord and block the pain signals .
The Future
AI has a bright and promising future in the field of drug development, as it can offer new opportunities and possibilities that can transform and improve the drug development process. Some of the future trends and directions are:
- AI will become more integrated and collaborative with other disciplines and technologies, such as biotechnology, nanotechnology, or quantum computing, to create and leverage new data sources, methods, and platforms that can enhance and expand the scope and scale of drug development.
- AI will become more personalized and precise with the patients and populations, by using techniques such as pharmacogenomics, precision medicine, or digital health, to tailor and customize the drugs and treatments according to the individual characteristics and needs of the patients and populations.
- AI will become more ethical and responsible with the society and environment, by using techniques such as explainable AI, trustworthy AI, or sustainable AI, to ensure and maintain the transparency, accountability, and sustainability of the AI applications and outcomes in drug development.
Ethics and Law
AI has many ethical and legal implications and challenges in the field of drug development, that need to be considered and addressed by the stakeholders, such as the researchers, developers, regulators, or users. Some of the ethical and legal issues are:
- Data privacy and security: AI uses large and sensitive data sets to perform its tasks, but the data may be vulnerable to breaches, leaks, or misuse, that may compromise the privacy and security of the data owners, such as the patients, researchers, or companies. For example, the data may be accessed, copied, or stolen by unauthorized parties, such as hackers, competitors, or criminals, who may use the data for malicious or fraudulent purposes. The data may also be shared, sold, or transferred by authorized parties, such as the data providers, collectors, or processors, who may use the data for commercial or financial benefits, without the consent or knowledge of the data owners. To address these issues, the data privacy and security need to be protected and respected, by using techniques such as encryption, authentication, or anonymization, that can prevent and deter the unauthorized or inappropriate access and use of the data. The data also need to be regulated and controlled, by using laws and policies, such as the General Data Protection Regulation (GDPR) or the Health Insurance Portability and Accountability Act (HIPAA), that can define and enforce the rights and responsibilities of the data owners and users .
- Algorithm fairness and bias: AI uses complex and sophisticated algorithms to perform its tasks, but the algorithms may be unfair or biased, that may affect the quality and reliability of the AI outcomes and decisions. For example, the algorithms may be unfair or biased, due to the use of data or models that are not representative or inclusive of the population or context, that may result in the discrimination or exclusion of certain groups or individuals, based on their characteristics, such as gender, race, or age. The algorithms may also be unfair or biased, due to the use of methods or criteria that are not objective or consistent, that may result in the favoritism or prejudice of certain groups or individuals, based on their interests, values, or goals. To address these issues, the algorithm fairness and bias need to be detected and corrected, by using techniques such as auditing, testing, or debiasing, that can identify and measure the sources and impacts of the algorithm unfairness and bias. The algorithm also need to be monitored and evaluated, by using techniques such as feedback, review, or oversight, that can verify and validate the algorithm outcomes and decisions .
- Human–machine trust and responsibility: AI interacts and collaborates with humans to perform its tasks, but the trust and responsibility between the humans and the machines may be unclear or uncertain, that may affect the safety and effectiveness of the AI applications and outcomes. For example, the trust and responsibility between the humans and the machines may be unclear or uncertain, due to the lack of common language, understanding, or agreement between the humans and the machines, that may result in the miscommunication or misunderstanding of the information and instructions between the humans and the machines. The trust and responsibility between the humans and the machines may also be unclear or uncertain, due to the lack of clear and consistent rules, policies, or standards that govern the roles, rights, and duties of the humans and the machines, that may result in the conflicts, errors, or risks that may arise from the humans or the machines. To address these issues, the human–machine trust and responsibility need to be established and maintained, by using techniques such as education, training, or certification, that can enable and facilitate the communication and exchange of information and instructions between the humans and the machines. The human–machine trust and responsibility also need to be clarified and defined, by using techniques such as ethics, law, or regulation, that can determine and ensure the roles, rights, and duties of the humans and the machines .
Conclusion
AI is a powerful and innovative technology that can revolutionize the field of drug development, by offering improved efficiency, accuracy, and speed in the drug development process. AI can help in various stages of drug development, such as drug discovery, drug development, drug synthesis, drug combination, drug delivery, and more, by using various techniques, such as deep learning, machine learning, optimization, game theory, and more, to analyze, model, and predict various aspects of drug development. AI can also offer new opportunities and possibilities that can transform and improve the drug development process, such as integration, personalization, and ethics. However, AI also faces many challenges and limitations that need to be addressed and overcome, such as data quality and availability, algorithm complexity and interpretability, human–machine interaction and collaboration, and more, by using various techniques, such as data cleaning, validation, integration, standardization, or anonymization, algorithm simplification, visualization, or verification, natural language processing, speech recognition, or gesture recognition, and more. AI also has many ethical and legal implications and challenges that need to be considered and addressed, such as data privacy and security, algorithm fairness and bias, human–machine trust and responsibility, and more, by using various techniques, such as encryption, authentication, or anonymization, auditing, testing, or debiasing, feedback, review, or oversight, and more. AI is a promising and exciting technology that can have a significant and positive impact on the field of drug development, and ultimately, on the health and well-being of the society and the environment.