Artificial Intelligence/Machine Learning (AI/ML) has the potential to significantly lower the costs of drug development and improve treatment outcomes across therapeutic areas in biotechnology and healthcare settings. Computer-aided processes have been a key tool in drug development for over 30 years; however, their utility in developing new medicines has hit a roadblock. In fact, a third of all drug applications in 2022 received complete response letters, the highest percentage in five years (1). This high failure rate discourages investment into the biotechnology sector and diverts valuable resources that would otherwise go to more impactful medicines.
This could all change with the advent and suitable application of advanced artificial intelligence and machine learning. Applications that stand to benefit the most from AI/ML are derived from molecular, biological, preclinical and clinical data sets. For example, trained algorithms can aid in the design of new molecular entities (NMEs) and streamline their manufacture. They can optimize treatment regimens based on clinical outcomes, select and recruit patients for clinical trials and monitor patients post-treatment. Furthermore, AI/ML’s predictive nature in silico eliminates the need for iterative and costly wet lab experiments. The end result is better patient outcomes with fewer resources.
A key consideration in the successful application of AI/ML is sourcing quality data from representative populations and applying that data in an ethical and secure setting. Regardless, there continues to be a high unmet medical need for more safe and effective treatments and AI has an opportunity to improve care on multiple fronts, resulting in better outcomes for patients and investors. Here are three areas that AI/ML will unlock substantial value for drug development in the years to come:
1. AI Speeds Up Drug Discovery
AI/ML models can be used to accelerate discovery of novel targets, design compounds and identify drug combinations for more effective treatments. The first step is to optimize the predictive nature of drug–target interactions. Historically, this process requires iterative experimentation that takes significant time and resources, even with the advent of molecular docking software. Recently developed ML methods use data-driven frameworks to improve predictive power through integration of growth inhibition data, gene expression data, adverse reaction data, chemical structure data, and other drug data (2). One such framework is DrugnomeAI, characterized as an “ensemble machine-learning framework for predicting druggability of candidate drug targets.” It is an adaptation of the Mantis-ML framework classifying genes as ‘druggable’ or ‘not druggable’ based on supervised learning of known gene-disease associations (3).
Once a target has been identified, methods of screening for drug-hit compounds range from structure-based screening using docking scores to ligand-based screening using deep neural virtual screening (DENVIS) (4). The latter model uses graph neural networks of atomic and surface features to predict drug-like compounds with greater speed and similar accuracy to amino acid sequence-based ML models. Likewise, researchers use fragment-based drug design (FBDD), which incorporates ML to identify which molecular fragments are more likely to bind in a pocket. FBDD dramatically improves high-throughput screening (HTS) processing time in the hunt for drug-like molecules.
Generative AI using physics-based approaches to screen small molecules has seen recent deal activity, with pharmaceutical giant Sanofi entering in a $140 million drug discovery collaboration with Aqemia. Although this collaboration is not limited by therapeutic area, there is intense interest in using AI to develop combination approaches, such as with PD1’s in the oncology space. HTS and combinatorial chemistry are used for both de novo design of drugs as well as drug repurposing. As a result, the hope is that there will be new molecular entities (NMEs) used in combination treatments with a higher likelihood of regulatory success as well as repurposed drugs with a shorter time to market via the US FDA’s 505(b)(2) regulatory pathway.
2. AI Enables Personalized Medicine
Picture a world with “single patient segmentation.” A place where AI enables ultimate personalized medicine where treatments are tailored to the genetic make-up of an individual and underlying disease mechanism. AI researchers are actively compiling empirical data sets, modeling genetic heterogeneity, and optimizing complex manufacturing and supply chains to make this happen. With quality genetic databases, AI/ML models are trained to link genetic information to real-world clinical outcomes (5). This is possible through large, integrated clinicogenomic databases, such as the ongoing effort by the National Human Genome Research Institute. Another effort is the Genome Aggregation Database (gnomAD), a compilation of both exome and genome sequencing data. These open-source databases enable researchers worldwide to develop training sets on a variety of disease-specific and population genetic studies to further advance personalized medicine.
Using quality data sets, AI/ML models can develop and identify biomarkers to diagnose and monitor patients, as well as predict patient response to treatment. This data is sourced from a wide range of modalities, including genomics, transcriptomics, and epigenomics, radiomics, and digital pathology (6). One area of research is immune checkpoint inhibitors (ICIs), including PD1s, which made up 43% of all cancer approvals during the last four years. Flow cytometry can be used to develop immunoprofiling training and validation sets for AI/ML models to better predict patient response to various PD1 combinations in clinical trials (7). With regards to genetic variants, deep learning tools such as DeepVariants uses convolutional neural networks with known genotypes to identify single-nucleotide variations, small insertions and deletions, and copy number variants (8). This tool can also aid in the development of tailored gene therapies for Orphan indications.
In addition to the design of NMEs, AI can be deployed to optimize chemistry, manufacturing, and controls. CDMOs use ML models to optimize process parameters and predict synthetic routes to scale-up NMEs and save their clients costs. Similarly, real-time sensor data can be analyzed and automatically adjusted for quality control of continuous batch processes. Of particular focus is the application of AI/ML in the manufacturing and scale up of cell and gene therapy products, as cost-of-goods is a meaningful expense. AI-enabled closed-loop systems allow for optimization of payload design in silico, which can save significant time and resources from inefficiencies at GMP scale from low yields and impurities.
3. AI Improves the Patient Experience
The patient experience encompasses preventative care, disease detection, lifestyle modification and/or pharmacological intervention, and post-treatment monitoring. AI can unlock value at each step, starting with preventative care and detection. AI, for instance, can analyze neuroimaging and other data to detect early signs of neurodegenerative diseases, such as Alzheimer's and Parkinson's. In addition, AI can analyze data from cardiac imaging or from wearable devices to detect abnormalities, predict cardiovascular events, and assist in personalized treatment plans. AI/ML integrated smart watches are being developed to continuously monitor blood pressure and glucose levels to alert physicians of pathologies such as heart disease and diabetes, respectfully. Last, AI enhances medical imaging interpretation, aiding in early cancer detection, tumor characterization, and treatment response.
Clinical trials are the most expensive stage of drug development. Only 10% of IND-cleared compounds reach the market at a cost of almost $3 billion dollars (9). Unacceptable safety and efficacy contribute to this high failure rate. Meanwhile, pricing concerns also affect reimbursement and adoption post-approval. Clinical-stage companies can apply AI to create efficiencies across clinical trials, from Phase 1 to registration trials. There is much interest in AI’s application to recruitment and clinical trial design to improve the patient experience. This includes better patient selection, quicker enrollment, and adaptive clinical trial designs (10). Enrollment in particular is a big challenge for many clinical trials, particularly those for rare diseases, for highly competitive indications, and for those with restrictive patient eligibility criteria. AI/ML models may be used to enhance statistical analysis to better screen patients and provide meaningful information with small or incomplete sample sizes.
Remote monitoring of patients during clinical trials and in post-marketing surveillance allows companies and caregivers to make well-informed treatment decisions. For instance, AI has been used to predict the response to immunotherapy from radiology or histopathology images via surrogate biomarkers (11). Data may also be collected remotely, saving a patient a trip to a clinical trial site and increasing adherence. Last, digital phenotyping may be employed where AI analyzes digital data, such as smartphone usage and social media activity, to provide insights into mental health conditions and potential early warning signs. Overall, it is imperative that collected data is kept confidential and data sets are inclusive to prevent bias, and that algorithms are periodically reviewed by ethics and medical review committees, such as in the context of whether or not to continue treatment.
What is the Value of AI in Drug Development?
Successful application of AI/ML will greatly impact biotech valuations in the years to come. Its various use cases will transform the industry by enhancing research and development, increasing efficiency of drug design and manufacturing, and optimizing clinical trials. All of these factors serve to improve the overall patient experience. What does this look like in valuation terms for a business or product? An immediate impact to net present value (NPV) stems from a combination of:
lower discount rates due to higher chance of regulatory success
higher ROI for R&D and manufacturing spend
lower clinical trial costs due to accelerated approval pathways
faster time to market and higher commercial value from longer post-approval IP runway
lower patient recruitment sample sizes and time between phases
overall lower impact from inflationary events due to lower capital expenditures
higher pricing premiums due to greater healthcare savings and better outcomes
Risk, cost, and development time is expected to decrease along the drug development continuum. Lower upfront capital expenditures will enable more biotech companies to raise private funds as less capital is required to reach value inflection points. This will further democratize drug development, allowing smaller companies to develop highly personalized treatments. It will empower patients to take control over their biometric data and their health, educating them on their predisposition to disease so they can make appropriate lifestyle and medication adjustments in real time. The greatest value will be a population-wide improvement in morbidity and mortality and savings to the healthcare system burdened by aging demographics. The promise of AI/ML is significant and only time will tell how long it takes for incremental improvements to become breakthroughs in a highly regulated industry known for long development timelines and political uncertainty.
Questions for Further Exploration
What applications of AI/ML will have the greatest impact on patients in the near term?
What regulatory guidance should the FDA pursue to speed up drug approvals?
How will the FDA incorporate advances in AI in its review of regulatory submissions, for instance the CMC section of the IND?
What are the biggest ethical concerns regarding the application of AI/ML to human health data?
Where is quality data limited for the application of AI/ML to clinical drug development?
Author's Disclaimer: A human contributed to 90% of this article and AI contributed 10%, similar to the Human:AI ratio in drug development today. I expect the ratio for both writing articles and drug development to reverse in 3-5 years.
References
(1) Pink Sheet FDA Performance Tracker’s Complete Response Letters. Citeline Regulatory.
(2) Wang L et al. Advances of Artificial Intelligence in Anti-Cancer Drug Design: A Review of the Past Decade. Pharmaceuticals (Basel). 2023.
(3) Dimitrios Vitsios, Slavé Petrovski, Mantis-ml: Disease-Agnostic Gene Prioritization from High-Throughput Genomic Screens by Stochastic Semi-supervised Learning, The American Journal of Human Genetics, Volume 106, Issue 5, 2020, Pages 659-678.
(4) Krasoulis et al. DENVIS: Scalable and High-Throughput Virtual Screening Using Graph Neural Networks with Atomic and Surface Protein Pocket Features. J. Chem. Inf. Model. 2022.
(5) Bhandari et al. McKinsey & Company, Life Science Practice “How AI can accelerate R&D for cell and gene therapies”. November 2022.
(6) Prelaj et al. Artificial intelligence for predictive biomarker discovery in immuno-oncology: a systematic review. Ann Oncol. 2024 Jan;35(1):29-65.
(7) Lee et al. Artificial intelligence-based immunoprofiling serves as a potentially predictive biomarker of nivolumab treatment for advanced hepatocellular carcinoma. Frontiers in Medicine. 2022.
(8) Lin et al. Artificial intelligence-based approaches for the detection and prioritization of genomic mutations in congenital surgical diseases. Front. Pediatr., 01 August 2023.
(9) Wouters et al. Estimated Research and Development Investment Needed to Bring a New Medicine to Market, 2009–2018. JAMA. 2020;323:844–853. doi: 10.1001/jama.2020.1166.
(10) Askin et al. Artificial Intelligence Applied to clinical trials: opportunities and challenges. Health Technol (Berl). 2023;13(2):203-213. Epub 2023 Feb 28.
(11) Ghaffari et al. Facts and Hopes on the Use of Artificial Intelligence for Predictive Immunotherapy Biomarkers in Cancer. Clin Cancer Res. 2023 Jan 17;29(2):316-323.
Comments