MolGemini

• 

  Example property calculation using MolGemini (distance between centers of mass of 2 chains)

Inspiration

We are scientists who study the structure of proteins and its impact on the function of proteins. Studying the structure and function of proteins offers a chance to understand the fundamental processes happening in living cells, design new drugs, and improve health outcomes for millions of people.

As scientists who study protein structures, we often need to compute and re-compute properties of proteins based on their structure. One bottleneck is that depending on how the structures have been prepared and processed, there are different computational tools to perform analyses on these structures. Also, scientists who want to study proteins but are not familiar with coding have to spend needless hours learning and writing code for calculating properties, something that Gemini can do in a second.

What it does

MolGemini, our AI assistant, asks the user for a protein ID according to the protein data bank (pdb), the property to be computed, and the chain of the protein for which this property needs to be computed. Using this information, MolGemini writes and executes code to directly give you any property that you want for any protein!

Additionally, there is also an option to print the code used by MolGemini to perform the analysis, so you can verify that it is indeed doing the correct thing.

How we built it

We used our expertise in protein structure analysis to engineer prompts that guide Gemini to write code that calculates the desired property. Using a backend API that interfaces with the Protein Data bank, we download the protein structure file and execute the code written by Gemini on this file.

Challenges we ran into

In the current state, AI assistants are not robust enough to write code with guaranteed success. Therefore, we had to implement a recursive strategy to write code that works and gives a reasonable output. We also had challenges in getting the exact information from a variety of user prompts, for which we leveraged Gemini's already fine-tuned interaction capabilities.

Accomplishments that we're proud of

We are very proud of how our product has turned out, especially the fact that it can take arbitrary properties (for which it was never trained or fine-tuned) and compute it for any protein structure. Using smart prompt engineering and trying out a variety of test cases, we have been able to set the foundation for a product that could be very useful for scientists across the world.

What we learned

This journey has been very fruitful for us on many levels. Firstly, we obtained experience in ideating on a product idea that could be useful for real-life users. Secondly, we learned the Vertex AI API which helped us implement what we wanted to. And finally, we learned engineering prompts which massage Gemini's "thinking" in a certain direction.

What's next for MolGemini

We would like to make the product more robust, both in terms of checking the output of the calculations against some benchmark values as well as in terms of better interpreting exact user prompts. In the future, we envision that MolGemini can also read and analyze trajectories of moving proteins from computer simulations.

Built With

• python
• vertex