Introduction

For the first eight weeks of this ten week series, we covered the drug development process breaking down the steps and regulations for a drug to become FDA approved and released into the market. Last week wrapped up the final stage, Stage 5: Post-Market Surveillance, where the FDA oversees the approved drug in the market reacting to any cases relating to it. The FDA can react to these cases by removing the drug off the market and putting it back through the clinical process to find a solution. Throughout the process, we also identified many areas that could be significantly improved with the assistance of AI and ML, but face problems before they can be implemented to work autonomously. In previous stages such as drug development and preclinical research, AI and ML have been integrated and have shown significant deductions in time and costs while increasing accuracy of the results. This week, we will identify the major problems that keep AI and ML integration as well as reasons the scientific community is divided over automation. Some of these problems are as simple as ethics while others consist of lack of data or technology to preserve samples.

Challenges in AI

Currently computer scientists and engineers face several challenges in the field of AI such as biases, transparency, generalization, functionality, lack of accurate data, regulation, and assistance. In order to prevent bias in data selection, preprocessing and algorithms the code must be near perfect to minimize bias while gathering, sorting, and filing data. AI learns from existing data, therefore if the bias could lead to unfair treatment in critical areas, which is a risk the industry must consider. Transparency or accountability for AI is grouped with a multitude of ethical problems as AI’s decision making progress is robotic without human emotion and based on the data it is given and may or may not be biased if there is an error in the code as mentioned above. To minimize problems like these stakeholders and users are informed about the AI decision making process and how they work. Companies also provide documentation of data sources, training methodologies, and performance metrics to promote transparency. Other ways to combat this issue is training employees or users by demonstrating ethical AI practices, addressing bias, and integrating human decision making in the process. According to nature.com, AI generalization is the ability of AI systems to apply and/or extrapolate their knowledge to new data which might differ from the original training data. This is a major challenge for effective and responsible implementation of human-centric AI applications. A possible solution proposes selective prediction as a solution.

Adopted from Goetz et al. 2024. The graph shows the problem with generalization where ML models trained on biased or non-representative datasets may fail to generalize to a subset of patients. The bottom half of the graph shows potential solutions to the generalization challenge.

Generalization is a problem due to short, expressive ML models overfit when underspecified. Generalization can also be a result of faulty data that leads to algorithmic biases which results in underrepresented groups after generalization. These problems occur mostly in clinical applications as datasets are high dimensional, consist of multiple inputs, and may not be representative of the target population. The current proposed solutions are focused on effective data collection such as using representatives to create unbiased datasets, model-based selection, and sample-based selection. Including AI and ML into the drug development pipeline has increased the efficiency of the process with each new algorithm supporting the previous or improving it. AI and ML requires computing power to function at its maximum capability, which means high-performance GPUs, TPUs, and other pieces that add to the already complicated and new algorithm. This makes companies ponder if the price is worth the addition as not all algorithms guarantee success as integration may take anywhere from days to years. AI and ML are also machines and machines can have errors that could once again take anywhere from hours to years to fix. There are many risks when it comes to implementing AI and ML into the process, but the efficiency after successful integration significantly increases. This also restricts smaller companies and organizations from having access to this new tech as they do not have the budget to buy all the computer parts to harbor the program. Some solutions to these problems are hardware architectural innovations that allow high performance in smaller pieces and data analytics to decide whether that investment in AI/ML is worth it for the company. As mentioned above, AI/ML learns from existing data, but can not filter if the data is bad or good, which leads to various problems such as biased generalization and bad decision making to name a few. Some solutions to verify the data given to AI can be cross-validation or authentication to ensure the data is not false and human assistance. Finally the ethical and regulatory issues that come with integrating AI/ML into companies can range from data leaks to accidental decisions made by the program that result in harming someone. Privacy is very important when it comes to the substantial amounts of data that can be stored or processed by AI/ML algorithms. There are many solutions to decrease the chances of leaks and breaches, but even with encryption methods and other methods of protection, nobody can fully guarantee security. The tricky problem with the implementation of AI/ML are the legal issues from decisions and who will be held accountable. As mentioned above, AI/ML programs and algorithms have the ability to make the decisions without any emotional bias, but risk having a bias due to bad data, which can result in significant damage to someone’s career, even possibly firing them or defaming them. Most firms have a legal team to decide how the liability is split amongst the company and the owners of the code, but the damage done is irreversible in some cases. This is one of the main reasons why the industry is scared to include AI/ML during clinical trial procedures due to the possibility of a mistake.

Due to these issues and no concrete solution, the assistance of human workers are still needed before the process becomes automated letting AI/ML make their own decisions. As the technology increases, AI/ML will take over the general tasks such as data gathering, filtering, and simple decision making based on inputted statistics. Some false assumptions about AI/ML can also cause a fear factor as there have been theories of how AI will take over the world and take over human jobs, but AI/ML simply states it is a program that reads the data it is given and executes the program and learns throughout that process.

Cases of AI’s Challenges in Drug Discovery Pipeline

As mentioned above there are various risks and challenges when it comes to implementing AI/ML into the drug discovery pipeline and in this section we will introduce two cases where there was a problem caused due to AI/ML and one successful case. In Goetz’s article, she uses the example of AI in breast cancer. Men suffer worse health outcomes and are underrepresented in clinical datasets, which does not allow AI to accurately predict for men. “ . . . a recent breast cancer prognostic algorithm, trained only on female data, offers accurate predictions for women but is expected to underperform for biological men due to exclusion from the dataset. (Goetz et al. 2024). This is caused by a lack of data, which could create a bias in the program and decreases the accuracy overall. The solution to this problem would be to create an accurate data set for men so the program can accurately predict based on the wide range of data it has with men and breasts. cancer. Another example would be predictive model testing for drug efficacy during preclinical trials. AI/ML models are used to forecast and predict the effects of new drug compounds during preclinical trials and even early parts of clinical trials. During one case, the model failed to account for the change in external conditions from inside a lab or a real-world environment, which it gave falsely optimistic predictions. The company looked at these predictions and decided to go forward with the new compound and suffered significant financial losses. There are multiple success stories of AI/ML usage in the drug discovery process, but we will use Exscientia as an example. In early 2020, the first AI-designed drug candidate entered clinical trials and this effort was led by Exscientia. While looking for a cure for obsessive-compulsive disorder (OCD), Exscientia used three AI-designed drug candidates to enter clinical trials. Specifically talking about the discovery of DSP-1181, a full serotonin 5-HT1a receptor agonist that was discovered as part of a collaboration between Excientia and Sumitomo Dainippon Pharma. The drug was discovered within 12 months of research using Centaur Chemist, Exscientia’s AI drug discovery program.

With the development of AI and ML, there are various challenges, but the result of successful AI/ML algorithms increase efficiency and decrease time and labor costs significantly. Majority of the challenges are based on the inputs and limited data we can offer to AI, which restricts its ability to fully learn and develop. Currently scientists and engineers are searching for solutions to decrease the risk and hopefully one day get rid of them.

AI’s Role in Drug Discovery and Development

A quick recap from a previous week’s article, AI’s role in drug discovery can be strictly defined as information gathering and filtering with human assistance. AI/ML efforts have reached the beginning stages of drug discovery and have been becoming more dominant through the clinical trials. As long as the algorithm is given a dataset and a detailed task that does not require the program to make the decision itself, using AI/ML is the most efficient method. Insilico, Exscientia, and other companies have used AI/ML in drug discovery and significantly shortened the time required from one to three years to some compounds being discovered within months.

Works Cited:

Simplilearn. (n.d.). Challenges of Artificial Intelligence. Simplilearn. https://www.simplilearn.com/challenges-of-artificial-intelligence-article

Mahmud, R., Ali, N., Komorowski, M., & Clifton, D. A. (2024). Selective prediction in clinical AI: Generalization and reliability in real-world healthcare settings. npj Digital Medicine, 7(1), Article 11. https://doi.org/10.1038/s41746-024-01127-3

Bai, Y. (2023, July 26). Artificial Intelligence and Synthetic Biology Are Not Harbingers of Doom. Stimson Center. https://www.stimson.org/2023/artificial-intelligence-and-synthetic-biology-are-not-harbingers-of-doom/

CAS. (2023, July 18). AI drug discovery: Assessing the first AI-designed drug candidates to go into human clinical trials. CAS. https://www.cas.org/resources/cas-insights/ai-drug-discovery-assessing-the-first-ai-designed-drug-candidates-to-go-into-human-clinical-trials