4 Ways AI Could Revolutionize The Future of Drug Development…No Chat Needed-The HSB Blog 2/16/23
Our Take:
The drug development process is a time-consuming and expensive endeavor fraught with failures even with proper planning and execution. With an average of $1-2 billion spent per successful drug and a development period of 10-15 years, the high cost and lengthy timeline are barriers to entry for many drug manufacturers. However, the integration of artificial intelligence (AI) into the drug development process has the potential to modernize the industry. AI can help researchers in a variety of ways, such as by analyzing large datasets, predicting biological processes, identifying new drug targets, and assisting in the design of new drug molecules. Furthermore, AI can assist in data mining, generating regulatory documents, and identifying suitable candidates for clinical trials. The implications of these developments are significant, as AI has the potential to improve the speed and efficiency of drug development, ultimately leading to the production of more effective treatments, although care must be taken to ensure its factual accuracy and validity as an increasing number of companies adopt AI solutions.
Key Takeaways:
Most drugs take between 10-15 years to be developed at an average cost of $1-2B before receiving [U.S.] approval for clinical use (Chinese Academy of Medical Sciences and the Chinese Pharmaceutical Association)
It is estimated that 85% of the human proteome is considered undruggable and finding effective pharmaceuticals to target these proteins is considered exceptionally hard, or impossible (The Cambridge Crystallographic Data Centre)
Machine learning methods such as eToxPred correctly predict synthetic accessibility and toxicity of drug compounds with accuracy as high as 72% (BMC Pharmacology and Toxicology)
The use of Machine Learning in drug discovery could save approximately $300-400M per drug (U.S. General Accounting Office)
The Problem:
Drug development is a time-consuming, costly process rife with failures even with strong strategic planning and execution of the process. For example, as noted in a recent article in Acta Pharmaceutica Sinica B (the journal of the Chinese Academy of Medical Sciences and the Chinese Pharmaceutical Association), most drugs take between 10-15 years to be developed, with an average cost of $1-2 billion spent before finally receiving federal approval for clinical use. While in theory during “clinical drug development, a delicate balance needs to be achieved among clinical dose, efficacy, and toxicity to optimize the benefit/risk ratios in patients. [Ideally] a drug candidate would have high potency and specificity to inhibit its molecular target [supplying] high drug exposure in disease-targeted tissues to achieve adequate efficacy at an optimal dose (ideally at low doses), and minimal drug exposure in healthy tissues to avoid toxicity at optimal doses (even at high doses).” However, while this is easy to specify in theory, in practice it becomes difficult to execute. For example, according to the article “analyses of clinical trial data from 2010 to 2017 show four possible reasons attributed to the 90% clinical failures of drug development: lack of clinical efficacy (40%–50%), unmanageable toxicity (30%), poor drug-like properties (10%–15%), and lack of commercial needs and poor strategic planning (10%).
Consequently, each of the five stages of drug development 1) discovery & development, 2) preclinical research, 3) clinical research, 4) FDA review & approval, and 5) FDA post-approval drug safety monitoring, require significant funding and resources yet can have precarious returns. As noted in an article in Nature Reviews Drug Discovery, while the probability of going from phase III to launch has risen from 49% to 62% in the periods from 2010-2012 to 2015 to 2017, the probability of a compound going from phase II trials to phase III trials has remained essentially the same at about 25% during this same period. The process is voluminous and requires analysis of large amounts of varying types of data. As noted in a report from the GAO on the benefits and challenges of machine learning in drug development, there are multiple types of data relevant to drug development, including data from biomedical research to better understand the biology of diseases, the pharmacology of potential drugs, the toxicity of known compounds as well as the various forms of patient data necessary to conduct the trial and analyze efficacy. Asa result, drug developers are faced with the task of analyzing ever increasing amounts of data to produce similar declining returns on their research leading them to seek new ways to search for and analyze potential candidates, such as the application of AI to the drug discovery process.
The Backdrop:
AI has the potential to solve a variety of industry problems and is being used in drug development to rapidly speed up the process of creating and assessing the effects of these novel compounds. While some authors have identified at least 10 ways AI can help in the drug discovery process (see Machine Learning in Drug Discovery: A Review) some of the more common uses that researchers associate with the use of AI in drug development include: 1) helping to find promising new drug candidates in lead and biomarker discovery, 2) data analytics and prediction (ex: classification, clustering, and prediction) of effective candidates for further analysis, 3) using AI (and capabilities like digital twins) to improve the speed and efficacy of preclinical development, 4) the detection and understanding of the potential for adverse effects. For example, by feeding this data to AI tools, which find associations between patients’ genotypes and phenotypes, researchers are able to discover new biomarkers that allow for patient stratification as well as the identification of biochemically active genomic regions that respond best to certain drugs. As noted in the Journal of Signal Transduction and Targeted Therapy, not only can AI transform and interpret this data into potential biological processes that could be utilized in the pharmacodynamics of a certain drug compound more rapidly than human researchers, it can do so far more accurately than human researchers given AI’s ability to discover patterns and relationships.
In terms of designing the drugs themselves, AI tools can be used to save researchers a lot of time. AI that has been trained using advanced biology and chemistry data is assisting in identifying new drug targets and helping to build applicable new drug molecules. A significant problem in the process of drug discovery is the proper identification of genomic regions that could be useful in regard to potential drug targets, and an estimated 80% of the human genome is yet untested or simply undruggable. Understanding and examining large volumes of biological data resulting from the genomics, proteomics and experimental interpretation of a certain drug target is a lofty task to complete with traditional methods, and the complex biological networks are difficult to fully break down and map completely. By analyzing a target’s gene expression, protein-protein interactions, results from clinical trials and disease biology, these AI algorithms can predict if the target is suitable for drug interactions and build molecules with specific properties, activities and toxicities that can help identify suitable candidates as per Research and Markets’ report on AI in drug target discovery and validation.
In the preclinical and clinical spheres, AI is rapidly adapting to the needs of researchers in order to set up and analyze the data from necessary experimental trials needed for a drug to receive approval from the FDA and prove its efficacy so that pharmaceutical companies can create a product that works. The development and testing of new drugs creates terabytes to petabytes of biological data at each stage of development, which is ideally suited to AI tools’ ability to work with large datasets. Pfizer, one of the largest pharmaceutical companies in the world, which has been utilizing AI for data mining purposes, have reported that AI runs much faster and more accurately than any human researchers are capable of and provides the added benefit of helping the company to meet regulatory and quality control requirements such as generating the reams of materials necessary to be submitted during the development process. Moreover, as noted in a recent article in Trends in Pharmacological Sciences, outside of the drug development process itself, AI can be used to identify and access the patient records of those who are most likely to benefit from clinical trials, reducing the time to identify suitable trial candidates and improving success rates. This is extremely important to the success and speed of trials given that approximately 48% of trials miss enrollment targets and 49% of patients drop out of trials before completion (thereby making the identification of suitable candidates key to enrolling sufficient numbers to account for this). Additionally, the use of AI in remote patient monitoring solutions such as wearable devices, virtual outpatient services, and more can help to monitor patients and predict adverse health events thereby making pharmacovigilance more effective and cheaper. According to “Artificial Intelligence in Health Care Benefits and Challenges of Machine Learning” from the U.S. General Accounting Office in Drug Development, the use of Machine Learning in drug discovery could save approximately $300-400M per drug.
Implications:
As noted above, AI has the potential to dramatically speed up the development of drug discovery while simultaneously helping to reduce the cost and improve the efficiency compared to traditional technologies currently being leveraged to find new drug molecules. With increased public interest and popularity concerning AI solutions including but not limited to issues in healthcare, new tools are being developed that are already showing great promise. MIT researchers created a geometric deep-learning model called EquiBind that is an estimated 1,200 times faster than one of the fastest, state-of-the-art computational models. EquiBind outperformed the current state-of-the-art model, QuickVina2-W in successfully simulating the binding of drug molecules to protein-coding genes and saved significant amounts of time that are usually spent in computation using cutting-edge geometric reasoning. This advancement will ultimately allow AI to better understand and apply concepts of molecular physics, leading to better predictions and generalizations fueled by the vast amounts of collected information that is difficult and time-consuming for humans to accurately sift through. EquiBind is only one of the multitude of AI tools being developed for drug research, and as AI continues to improve on previous iterations and synthesize increasingly larger volumes of data, this will translate into far greater efficiency and time savings than can be achieved with current industry standards. In addition, AI will have applications in quality control as machine learning methods are used to evaluate drug candidates for toxicity and side effects. For example, according to an article in BMC Pharmacology and Toxicology, a technology called eToxPred can correctly predict the synthetic accessibility and toxicity of drug compounds with accuracy as high as 72%.
Over time as the adoption of AI accelerates in drug development there is the potential for the development of even more personalized medicines tailored to the specific needs and genome of patients. Given the vast amounts of patient data collected and stored by hospitals, insurers, and others in healthcare, and as the industry increasingly digitizes, there is a significant volume of data that is underutilized and which could be informing better care practices, including drug discovery. However, as with the application of AI in any industry, this must be done with ethical considerations in mind and specific policies and protocols in place in terms of data privacy, algorithmic bias, and transparency. AI can identify patients that are most likely to respond positively to a particular drug which could lead to treating individuals sooner, rather than possibly having to wait to participate in clinical trials (once again under the right safety protocols). The idea of more targeted and personalized healthcare remains intriguing, but it must be done with accountability and transparency in mind, so that clinicians and patients understand how and why the algorithms work the way they do. If so, there is the potential to fundamentally change the way we develop new drugs and get these experimental treatments to those who desperately need them more quickly and cheaper than ever before.
Related Reading:
