Our first Kaggle competition: 2nd of 250 teams in scientific data harmonization
We wrapped our first-ever Kaggle competition — and it couldn’t have been more demanding. In “Harmonizing the Data of your Data,” run by the NSF National Synthesis Center for Emergence in the Molecular and Cellular Sciences (NCEMS) at Penn State, we finished 2nd out of 250 teams (from over 5,000 submissions): #1 on the public leaderboard and #2 on the final private leaderboard — the only Latin American team on the podium.
The challenge
The question was simple to state and hard to solve: can a single AI pipeline read any proteomics paper and extract clean, structured experimental metadata? Organism, instrument, modifications, sample conditions, treatments, and dozens of other fields — all in SDRF format.
Why it matters
Reproducibility in proteomics depends on metadata that, far too often, lives buried in the free text of publications, in inconsistent formats. Standardizing that information at scale opens the door to better data discovery, better meta-analyses, and ultimately better science. As Ian Sitarik (ICDS) put it: “those data are much more useful when they come with clear contextual information.”
Our approach
There were no shortcuts. We combined NLP extraction from the manuscripts, public APIs (PRIDE, ProteomeXchange), per-file metadata parsing, multi-LLM consensus, and rigorous cross-validation against 103 harmonized training SDRFs. Several tools working together — not a single model.
The honest part
It was our first Kaggle competition. There were long nights and deep dives into proteomics literature we’d never touched. The placement means a lot, but what we value most is what we learned along the way — pushing ourselves outside our comfort zone, which is exactly where the real learning happens.
Thank you
To NCEMS, the National Science Foundation (NSF), and the organizing team — Wout Bittremieux, Shomir Wilson, Iddo Friedberg, and Ian Sitarik — for a competition this well designed and meaningful. And to the Penn State Institute for Computational and Data Sciences (ICDS) and the Penn State Huck Institutes of the Life Sciences for supporting this work.
This is just the beginning.