Q&A with SandboxAQ vice president of engineering Stefan Leichnauer
SandboxAQ recently published several scientific breakthroughs in using AI for biopharma drug discovery, the development of new materials and battery chemistry. The company has also announced its generative AI-powered objective molecular design solution, IDOLpro.
In this Q&A, SandboxAQ vice president of engineering Stefan Leichnauer discusses these innovations, the limitations of large language models (LLMs) and the introduction of large quantitative models (LQMs).
AI Business: What is the need for LQMs in molecular simulation?
Stefan Leichnauer: The world is a little bit taken right now with LLMs and for the last two years, everything has been looked at as a language problem. These language models are so powerful and capable, people are being fooled into thinking that they can do everything.
But they're not the right tool for every job, because not every problem is a language problem and using them doesn’t mean that you know how to discover the next big blockbuster drug. LLMs don't actually know what they're doing and they have problems with hallucinations.
Quantitative AI differs in the kind of data we’re manipulating and the kinds of problems we are trying to solve. A large quantitative model is an analog of a large language model, but it's built to handle these more quantitative problems. It can simulate and build in a way that is not susceptible to hallucinations
You can build on the first principles of physical laws that govern how chemistry works so that the model knows what it's doing and doesn't make those kinds of mistakes. You simulate at a fundamental physical level in a way that actually represents what happens in real life, making it trustworthy.
You recently presented at the 38th Annual Conference on Neural Information Processing Systems (NeurIPS) on using generative AI to enhance structure-based drug design with the IDOLpro framework. What did you discuss?
The problem with traditional structure-based drug design is you have a protein target that you're trying to attack so you need to find a molecule that will attach to this protein and is synthesizable in the lab. The traditional approach is to mix chemicals together and see what happens, but that’s haphazard and costly.
The next step is to use a computational chemistry calculation to simulate whether the molecule will bind in the desired way. It’s cheaper and faster than the traditional method but could still take several days to go through a long list of drugs. It could take several days.
Now, you can use generative AI to generate a molecule that's going to fit into the protein. But you can run into problems, For example, if you're generating a new molecule to attach to this protein using this generative model, the candidate molecule may be impossible to synthesize or prohibitively expensive.
At NeurIPS, we introduced inverse design of optimal ligands for protein pockets (IDOLpro). Imagine you have trained a big AI model to generate molecules, but you want to get better outputs . That's what the goal is, to get better outputs from the model you already have.
When AI generates a model, it starts from a fuzzy picture of what the molecule might be, and goes through a process of focusing on what the molecule looks like. Maybe there's an atom over here and another atom over there.
We run that process halfway through and pause it, when it’s still fuzzy and you're not quite sure what the outcome is going to be. Instead of just completing the process and getting a single answer, IDOLpro says: “If you did complete the process, I estimate you’ll get something that looks like this. That’s not a perfect answer for a few reasons.”
This means we can go back and make an adjustment at this fuzzy halfway-through point. We don't go back to the beginning and start from scratch again.
The initial focus is on structure-based drug design. What other sectors could benefit from the technology?
You can look at other domains that involve these sorts of molecular interactions. There are analogous problems in next-generation materials. For example, if you want lighter, stronger materials that are producible and inexpensive to make for next-generation vehicles or aircraft.
Similarly for next-generation batteries, you want to maximize the energy storage capabilities, but you don't want it to be toxic or to explode when you charge it.
It would also be used to produce better catalysts to make chemical reactions faster or using less energy. It’s a similar process to finding a drug that binds to a protein; the catalyst comes in and binds to the chemicals that are trying to react.
You May Also Like
