Inspiration
We were inspired by the need to advance in silico fragmentation for mass spectrometry, targeting molecules that current spectral databases often miss. While the topic was new to some of us, the idea of exploring the "dark chemical space" and pushing scientific boundaries motivated us to dive in and tackle this challenging problem.
What it does
Our solution predicts mass spectrometry spectra from molecular structures, integrating key parameters like collision energy and polarity mode. By leveraging a BART-based model and adversarial training, we aim to improve the accuracy of spectra prediction and expand knowledge beyond existing spectral libraries.
How we built it
We built the project remotely, which presented its own unique challenges. Using a BART-based model, we trained on over 3 million curated MS/MS samples. Setting up the necessary infrastructure was a major task, and we had to overcome hurdles with data processing, model training, and cloud resources to ensure the project could scale to handle large datasets.
Challenges we ran into
One of the biggest challenges was the remote setting—coordinating among team members across different time zones, all while some of us were still learning the fundamentals of mass spectrometry. Another challenge was getting the infrastructure to work smoothly. Setting up the tools and resources to handle massive datasets was harder than anticipated, and we faced several roadblocks before things finally came together.
Accomplishments that we're proud of
Even though we didn’t deliver everything we initially set out to achieve, we're proud of the progress we made. We managed to build a functional model and gained a deep understanding of in silico fragmentation. More importantly, we grew as a team and built a solid foundation for future work.
What we learned
We learned a tremendous amount, not just about the technical aspects of mass spectrometry and machine learning, but also about collaborating remotely and troubleshooting complex issues. The challenges we faced strengthened our problem-solving skills, and by the end, even those unfamiliar with the subject matter at the start became much more knowledgeable.
What's next for de-MS-tifying dark Mols
While we didn’t achieve every milestone, the project has sparked a lot of interest within the team. Some of us are excited to continue working on this problem, refining the model, and expanding its capabilities to make a real impact on the field. There's a lot of potential for future developments, and we look forward to tackling the next steps.
Log in or sign up for Devpost to join the conversation.