ºÚÁÏÕýÄÜÁ¿

Polina Avdiunina (she/her)

Mellon College of Science

PhD in Chemistry

Hometown: Moscow, russia

1. Briefly describe the work you're doing in your program at ºÚÁÏÕýÄÜÁ¿.

   I work on applying AI to drug discovery and lab automation. That means both building better computational predictions for potential drug candidates and making the path from those predictions to real experimental results shorter and more reliable. On the drug discovery side, I'm involved in two projects. The first is a collaboration with UPMC focused on a rare metabolic disorder called pyruvate dehydrogenase complex deficiency, which disrupts the body's energy production and currently has very few treatment options. I led the machine learning screening and docking campaign to find small molecules that could restore the function of the affected enzyme, and the candidates are now being experimentally tested with promising preliminary results, which is very exciting. The second is CACHE5, a global computational drug discovery competition, where we developed a pipeline to screen billions of molecules for a brain receptor involved in mood and metabolism and achieved the highest hit rate among all participating teams. On the lab automation side, I'm part of the team building ºÚÁÏÕýÄÜÁ¿'s AI Science Foundry, a cloud lab where scientific experiments are run by robotic systems. My job is figuring out how to translate what a scientist wants to test into something the instruments can actually execute, and making sure the whole pipeline runs smoothly. The common thread across all of this work is simple: making discovery faster, more reliable, and more efficient, from computation all the way through to the experiment.

2. How did you develop an interest in this area? 

   My interest came from observing, project after project, how remarkably slow the modern pharmaceutical pipeline still is. Finding initial hit molecules, improving their drug-like properties, iterating through rounds of real-world testing, all of that can take years before a candidate even enters clinical trials. When I understood the scale of that bottleneck, it clicked for me that this is exactly where AI and robotics can have enormous impact: automating the cycle of discovery so we can test more candidates, generate better data, and feed that data back into intelligent systems that learn from it. But as I dug deeper, I realized that acceleration only matters if the predictions are actually reliable. Many computational models in drug discovery look impressive on paper but fall apart under rigorous evaluation, and that became the focus of a paper I published in the Journal of Chemical Information and Modeling, where we showed that common practices like how datasets are curated and split can dramatically inflate how well a model appears to perform.

    

   That experience taught me something I keep coming back to: impressive output means nothing without rigorous validation. The same principle applies to my lab automation work. I'm building an LLM tailored for automated laboratory use, one that can help a scientist translate their experimental protocol into instrument-executable steps. Think of the way language models already help people write code. The same idea applies here, but in a lab, the stakes are different. A flawed protocol might silently produce bad data or waste limited reagents. So a big part of what I'm building is the verification layer: guardrails and validation checks that evaluate what the LLM proposes before anything actually runs.

3. What are your academic and/or professional goals? 

   Academically, I want to keep building my technical skills through real-world application work. Laboratory automation sits right at the intersection of science and robotics, and I genuinely believe that the key to doing it well is understanding both the experiment and the software that makes it run. That mindset helped me enormously when I joined ºÚÁÏÕýÄÜÁ¿'s Cloud Lab team, which brings together engineers and scientists with very different backgrounds. Being able to speak both languages made all the difference in how I contributed and how quickly I grew.

    

   Professionally, I want to widen my network of people who are passionate about accelerating scientific discovery through lab automation, in both industry and academia. The field is still so young that different groups and companies approach automation in completely different ways, each customizing it for their own systems. Putting myself out there, learning from other pioneers, hearing how they've solved problems I'm also thinking about, that motivates me like nothing else. It helps me pressure-test my own ideas and opens the door for the kinds of conversations and collaborations that push the whole field forward. Eventually, I want to be someone who helps define best practices in this space, not just follow them.

4. How do you spend your time beyond academic work?

   Being outside is where I recharge. I climb, hike, bike, play pickleball, and basically say yes to anything that gets me moving and out the door. I'm a member of ºÚÁÏÕýÄÜÁ¿'s pickleball club, which keeps things fun and a little competitive. Traveling to national parks for climbing and hiking trips is one of my favorite ways to spend a weekend or vacation. I also started running more seriously about a year ago, and I'm working toward the Pittsburgh marathon this fall, which is both exciting and a little terrifying. Outside of all that, I genuinely love to cook and bake, especially for other people. There's something about gathering people around a table with something you made for them that feels like the most natural way to build community.