There is hardly any industry that will not benefit from Generative models. While we are still in the exploration and experimentation stage, the promise of Generative models is enormous.
Life sciences is no exception. There are a few different areas that Generative models can impact. These examples are beyond vanilla automation and foray into helping humans expedite life science innovation.
In this article, we will explore one example in detail. The idea is to highlight how Generative models can make an impact. If you are an experienced life science professional, this example should help you brainstorm additional applications, like new enzyme and protein designs, scientific discovery process assistance, etc.
Example I: Lead Compounds Innovation
A critical component of the new drug discovery process is coming up with new compounds. We will explore this sub-process and the use of Generative models in this sub-process in detail.
Current state process:
Generally performed semi-annually, the process leverages the expertise of expert human chemists. These chemists brainstorm to suggest modifications to the core structure.
The brainstorming process goes like this:- Picture of the current molecular series is projected on the screen in a room full of expert chemists. Chemists review the picture and suggest modifications to the core structure of the molecule.
Subsets of the molecules for which modifications were suggested are then actually synthesized and tested. This process keeps getting repeated until a suitable molecule is found. If a suitable molecule is not found, the program is then dropped.
The process is robust and productive since it draws upon the expert insights of leading chemists. But then this same manual aspect of the process inserts drawbacks as well. Since the process requires the involvement of leading experts, scaling it becomes a challenge. This slows the pipeline of new ideas for lead compounds. The impact on humanity of this unavoidable slowdown is significant.
How Generative AI can help
If we can train a Generative model on appropriate molecular representations, the model may rapidly and at scale suggest new compounds. The scale and scope of such a model will expedite the drug discovery pipeline. At this stage, pure automation of the process may not be possible, but the Generative model can curate possibilities for human experts to review.
Like every other process in every other industry, where there is a potential to semi-automate or automate something that needs human intelligence or expertise with Generative models, real-world nuances must be baked into these models.
An example is that while the model has been trained to develop molecules that follow the rules like the rule of valence (ex: each Carbon atom has four bonds), valency does not always equate to stability. A human expert “knows” from their experience that certain functional groups will not be stable. Building these types of “expert knowledge” through training and design (like filters) will differentiate a model that works in the real world from a model that fails.
References:
Designed Analytics Blog 