Generative AI in new drug development

Within months of its release, ChatGPT found use in nearly every industry, including biotechnology, pharmaceuticals, and healthcare. ChatGPT and other major language models like it belong to a family of artificial intelligence (AI) technologies dubbed generative AI. Based on statistical associations collected from large data sets, generative AI models generate content that is similar to the trained data.

This is how ChatGPT generates realistic text, and AI art generators like Midjourney or DALL-E create images in the style of Picasso or van Gogh. Similarly, the generative chemistry model envisions chemicals that can pave the way for new drugs, materials or agrochemicals. For example, one study reported on an AI model that was trained on amino acid sequences new proteins are generated.

Labiotech spoke with Petrina Kamya, head of AI platforms at Insilico Medicine, about how companies are using generative AI to design new drugs. The company uses generative chemistry and big language models in its drug development program. The drug candidate primarily targets idiopathic pulmonary fibrosis, a rare lung disease.

How does your generative chemistry model work?

Generative AI algorithms are designed to generate new data samples that resemble certain data sets. What we’ve done is we’ve taken lots of different generative AI algorithms like diffusion models or variational automated encoders and adapted them to generate molecular structures like new drugs. They are trained under FDA-approved drugs or molecules that are destined to become drugs and they design drug-like molecules.

Are these drugs derivatives or could these also be new?

They really could become a novel. The idea with generative adversarial networks is to design molecules that trick humans into thinking they’ve seen them before, when they haven’t. You can produce molecules that look very similar. You can create existing ones and molecules that look very, very different. So there is a spectrum and you can control it to some extent. You can control the recency level in terms of output.

Hallucinations are a problem with ChatGPT because they can sometimes produce output that sounds reasonable but doesn’t make sense. Is there a similar problem in generative chemistry and how does your platform solve it?

The generative algorithm is pre-trained to generate molecules and then you have a reward function (on top of that) which acts as a guardrail in terms of the type of molecule that is removed from the platform. The reward function can be adjusted so that the generative algorithm creates molecules that meet certain criteria and doesn’t explode and produce bananas or nonsensical molecules.

ChatGPT struggles with languages ​​for which we don’t have much training data. Do generative chemistry models also underperform in cases of diseases, say rare diseases, where we have less data?

No. With generative chemical algorithms, the constraint is the target you are working on. If you generate a molecule and kind of filter it subconsciously without really knowing what it’s targeting, that’s the limit. Once you have a validated target it is relatively easy to produce a molecule that produces the desired result.

What does generative chemistry have to offer for rare disease drug development?

The advantage of using generative AI is that you can optimize molecular properties much faster than using traditional methods. With the reward function and active learning cycles of generative chemistry, you can obtain relatively viable molecules that you can synthesize and screen for potential toxic effects and penetration of the blood-brain barrier. You have less iterative cycles when you design it. I would say that the way generative chemistry helps rare diseases specifically is by producing drugs more quickly.

You have just integrated ChatGPT into your platform. How does that help your work?

There is a lot of textual data you can mine. And we’ve mapped that data for specific diseases in a knowledge graph. We use the ChatGPT functionality to query that knowledge graph and look for answers to questions. For example, if you see scleroderma, you can ask about the latest drugs currently undergoing clinical trials for scleroderma. It will unravel all of that data or come up with what is publicly available in terms of peer-reviewed journals and clinical trials.

What role does generative AI play in the development of drug candidates against idiopathic pulmonary fibrosis?

We use our platform with 20 different machine learning models to identify targets. Then we use generative chemistry to design molecules that act on those targets. We also use transformers and large language models to predict the likely transitions of our program from phase to phase.

This story was made possible with the support of the National Press Foundation. The Foundation does not influence the research or reporting of this article.