On January 6, the US Food and Drug Administration (FDA) proposed draft guidelines on the use of artificial intelligence (AI) to assess the safety and effectiveness of drugs.
The influential body has said that in the last decade, the number of submissions from drugmakers that include an AI or machine-learning component has seen an exponential rise. There was only one such submission per year in 2016 and 2017 but in the next two years it tripled; in 2021, the FDA reported a remarkable 10-fold increase on the previous year alone with 132 submissions including an AI and/or machine-learning component.
Drug development pitfalls
It takes nearly 10 years and over a billion dollars to develop a drug using conventional (animal-based) processes, which also have a success rate of only 14%. Emerging technologies like AI provide opportunities for us to catalyse and improve the human-relevant drug-development pipeline.
For example, rats can eliminate some drugs from their bodies much faster than humans can, which means that for the same dose level, humans would be exposed to the drug for a longer duration. As a result, the data for a compound obtained by testing with rats will have to be adjusted for this skew.
The responses of humans belonging to different populations around the world to drugs and diseases also vary according to age, sex, preexisting medical conditions, and genetic variabilities, among other factors. It’s often difficult to predict this range of responses from a homogenous, lab-bred animal population.
Inputs to predictivity
Researchers today use AI across the breadth of the drug development cycle.
In the discovery phase, researchers comb through databases with thousands of compounds to select a few hundred promising candidates for a particular use case. Then they test these compounds on animals during preclinical research. The data for compounds that produce encouraging results in animal models are submitted to drug regulators for permission to conduct human clinical trials.
The compounds found to be safe and effective in these clinical trials — conducted in three phases depending on the requirement — are thn released into the market following the Drug Controller-General’s approval. In the post-marketing stage, the drug manufacturer monitors the drug’s effects on the population, under an obligation to report adverse effects.
There are now AI tools that can digest data from a human adult about how their body absorbs, distributes, and eliminates a drug and based on that predict the response of vulnerable populations, such as children, whose participation in clinical trials raises thorny ethical and technical issues.
Another pain point in drug development that AI could surmount is predicting whether a drug could have unintended effects. In December 2024, researchers from the UK reported in the journal Toxicological Sciences a “safety toolbox” comprising a group of computational models that could predict the undesirable side effects of a chemical compound on the entire body or on specific organs the compound isn’t designed to target.
This framework involves integrating multiple types of data, such as the level and manner of exposure to the substance (topical, oral, etc.), its structural properties, and any information about its chemical properties.
Where do AI models fall short?
Despite the potential to overcome the barriers of conventional testing, AI comes with its own challenges. In particular, the reliability of data analysis performed by an AI tool depends on the quality of the data the model is trained with.
Participants at an FDA-sponsored workshop at Duke University in the US in 2022 used the adage “garbage in, garbage out” to describe this problem. The use of biased and/or under-representative data of a target population will also compromise the output.
Another challenge is transparency. The inner workings of most AI models in use are not open to independent scrutiny nor is the data used to train them easily accessible, so the models’ performance can’t be assessed as required.
FDA’s draft guidelines
The FDA has been open to the idea of using AI and its draft guidelines present a stepwise framework to assess models’ credibility. The text emphasises the importance of identifying questions of interest, the context for each question, how a model will help address it. This is because a model developed to identify the risk of one adverse reaction to one drug based on previous clinical trials may not be equally good at identifying the risk of other reactions and/or to other drugs.
The guidelines also stress the importance of assessing the risk AI models may pose. If a model concludes a patient is at low risk for an adverse reaction to a drug, an incorrect prediction could have life-threatening implications. Identifying the level of this risk is another parameter of importance. Axiomatically, improving the quality and quantity of data used to train the AI model and the identification of possible biases will strengthen the model’s integrity and value.
AI models can be self-learning, their outputs can change based on new inputs, and they can constantly adapt without human intervention. In response, the FDA framework recognises a need to continuously monitor and provide detailed maintenance plans across the lifecycle of these models. Given the currently vigorous AI landscape, the draft guidelines encourage the industry to engage with the FDA to discuss and design appropriate ways to assess their AI models.
The guidelines focus on the use of AI in the preclinical stage in particular, where it is critical to understand if a compound of interest is safe enough to be approved for human clinical trials.
Regulators and the pharmaceutical industry have traditionally banked on animal models’ response to the compounds for this assessment. But there is a growing body of work suggesting we need to improve the quality of data available at this stage as well as reduce animal suffering.
From guidelines to adoption
The European Medicines Agency and the International Council for Harmonisation of Technical Requirements for Pharmaceuticals for Human Use (a.k.a. ICH) have released similar documents on the use of AI in drug development processes. But the FDA guidelines are notable because they focus on the use of AI to support decisions regarding the safety and effectiveness of a drug before starting human clinical trials.
In 2023, India passed the New Drugs and Clinical Trials (Amendment) Rules 2023. It allowed data generated by advanced computational models to be used to assess the safety and efficacy of new drugs, freeing researchers from relying on animal trials alone.
This said, guidelines issued by regulators can help harmonise (i) government policy, (ii) manufacturers’ expectations and compliance burden, (iii) researchers’ strategy, and (iv) consumer safety.
In effect, the guidelines serve as a fixed point in the shifting AI space, an anchor where all stakeholders can pause to take stock together, before making the next decision.
Surat Parvatam is senior strategist (research and regulatory science) at Humane Society International India. Arvind Ramanathan is head of research and associate professor, DBT-inStem.
Published – February 13, 2025 05:30 am IST
