AI and machine learning are data-driven, but as anyone in the pharmaceutical industry can attest, not all data is created the same way. Andrew Anderson, vice president of innovation and informatics strategy at ACD/Labs, said: pharmaceutical company executive Learn how these algorithms transform your data into a more usable format.
Pharmaceutical executives: What role can AI and ML play in drug development?
Andrew Anderson: These are interesting times with a lot of excitement around AI and machine learning. You can also see success stories. The challenge is to scale up. I've used this analogy several times. Boiling a quart of water is easy, but boiling a tanker full of water requires a separate system or platform. What I've seen, and what our collaborators have seen, is that there is promise when it comes to AI and ML.
We find that when you have a structured training set, you can develop really good models from which you can get clear outputs that are useful and insightful. What I see as the biggest barrier to scale is the amount, variety, and nature of the training data required for scientific applications.
I've seen presentations from emerging AI companies where the data source is a database called KEMBEL, which is a European bioinformatics handpicked database. Many of these training sets come from the literature, and the KEMBEL data is extracted and curated by humans. The challenge is that the source of data is often not the proprietary screening results of large pharmaceutical organizations. More data are needed to accurately predict treatment-relevant attributes on a ubiquitous scale, including information about the target, what modulates its activity, and what biochemical outcomes occur. . This is especially true when predicting off-target effects.
Although cellular activity can be predicted fairly accurately, off-target effects still require a significant amount of wet lab experimentation. In terms of finding something therapeutic, scale and cellular activity is one of the challenges we've heard.
PE: What other areas do you see as opportunities?
Anderson: Another place where we see interesting opportunities in applying AI and ML technologies is in drug development. Suppose that through a discovery process, you identify a molecule or some type of treatment that is proven to be effective. The challenge lies in preparing the process for creating clinical trial materials and how AI tools can be leveraged there.
When creating materials, regulatory authorities require that you follow certain rules. GMP (Good Manufacturing Practice) is one of the standards that must be followed when manufacturing materials that may come into contact with humans during consumption. Developing processes that produce high quality is another thing, but ensuring those processes are reproducible and reliable is a more difficult but desirable goal in pharmaceutical development.
Where can predictive modeling be applied to reduce the amount of experiments that need to be performed? Confirmatory studies still need to be done. When you create a material, the AI ​​will likely guide you through the process of creating that material. When you actually make the material, you can confirm the high reproducibility and high quality.
PE: What is the industry currently focused on?
Anderson: Which laboratories and customers are we focusing on applying AI principles to process development? what are they looking for? A large amount of training data in a format that these systems can use. The data generated during these processes is the same as at the time of discovery, and all of the different assays and processes generate data in different formats. The next time you need to run a discovery campaign, how can you leverage prior knowledge to accelerate the discovery and development process?
For us, it's about data engineering. You've probably heard the statistic that the total cost of implementing an AI and ML initiative is data engineering. It's all about reducing the cost of AI and ML.
Here's the good news. The level of formality is evolving with the evolution of generative AI, especially these large-scale language models. You don't necessarily have to prepare your data in a very strict structure, as long as you can access the data and train large language models to describe the nature of the data. What we're researching is how to add very early additions to the AI ​​portfolio and get the most out of them. It's a very exciting time.
If you and I had had this conversation three months ago, things would have been very different. This is of interest not only to the pharmaceutical industry but also to society in general.