Team 15 (solo entry) · Scientific AI
LymeScope
Can we trust AI in biology?
Research software that makes every score about a Lyme disease surface protein traceable to its observations, its experimental structure and a reproducible calculation.
Franco Plathottathil
6,305
residue evaluations reproduced
Across two nodes, each present exactly once, with zero numerical deviation.
8.3%
shorter compute phase
Optimized two-node run against a fresh single-node reference. One pair.
0.62
exploratory AUROC
The model signal did not help, and the result was kept.
Figures are the team's own measurements from the hackathon weekend, as reported in their submission.
The problem
A convincing AI prediction is not a discovery. Researchers have limited lab budget and must decide which computational leads deserve experimental spend. Franco Plathottathil had Lyme disease for three years. The one-sheet puts it plainly: “This is the tool I wanted to exist.”
Two alignment columns can both show 100% agreement, one from 30 of 30 observed sequences and one from 1 of 30. Most tools render both the same way.
What was built
LymeScope analyzes OspC, a protein on the outer surface of Borrelia burgdorferi, the bacterium a tick bite transmits. It takes 30 natural OspC sequences through 6,305 ESM-2 residue evaluations and maps them onto the experimental crystal structure PDB 1GGQ.
Coverage is a field that cannot be dropped when results are rendered. Every claim carries its denominator, its provenance hash and its stated limit.
The finding that did not work
Model predictability and observed sequence agreement showed only a weak descriptive relationship. A controlled exploratory experiment then tested whether a model-disagreement signal improved prediction of experimental OspC antibody contacts. It did not, and the result stayed in. No therapeutic, vaccine target or medical predictor claim is made anywhere in the work.
How the cluster was used
Using Fuzzball across two physical NVIDIA GB10 systems with VAST Data storage, the project reproduced all 6,305 evaluations. A second GPU was not automatically faster. An equal split across two nodes took 68.0 s against 52.1 s on one. After measuring a large difference in effective node throughput and dividing the same work to match, the optimized two-node run took 50.0 s against a fresh single-node reference of 54.6 s, without changing a single scientific answer.
The 8.3% figure comes from one confirmation pair and is described as preliminary. The cause of the node asymmetry is documented but unresolved.
What comes next
- More naturally observed proteins and experimental structures.
- Validation against independent laboratory measurements.
- Larger reproducible workloads, with orchestration that accounts for node throughput.