CASMI 2026 / AN INTERACTIVE SEARCH FOR MOLECULES
Which molecule made these peaks?
Follow the evidence. Find the structure. Explore our search.
The molecule-finding machine
One unknown structure. Many possible answers. Four ways to gather evidence.
Try the live match ↓Synthetic teaching example · schematic structures · mass filtering and spectral scores run live in WASM. Other channels illustrate the full pipeline.
ZOOM IN / MATCH THE FRAGMENTS
Make the peaks line up.
Every peak is a clue.
The shortlist
Choose a row to compare its peaks. Scores are similarities, not probabilities.
Synthetic spectra and schematic structures. WASM computes mass filtering, spectral similarity, and reciprocal rank. Full model channels are illustrated. Methodology ↗
THE RESEARCH ATLAS / FOLLOW A BRANCH
Where we’ve been. Where we go next.
A local win on 63 molecules. Independent validation comes next.
Open the experiment notebook 6 evidence records + caveats
September 21, 2026 snapshot. Lines show research relationships, not measured contributions. Local gains are not leaderboard gains.
THE FINISH LINE / RANK THE RIGHT ANSWER
Finding it is good.
Finding it first is better.
Move the answer along the list. Reward = 1 ÷ rank. Outside the top 25 = zero.
0.328 MRR does not mean 32.8% accuracy. A mean reward is different from the share ranked first.
The small-print dictionary
What is a mass spectrum?
A list of fragment mass-to-charge ratios and their signal intensities. The pattern offers clues about the original molecule; one peak rarely identifies it alone.
What are candidates, coverage, and ranking?
Candidates are possible structures. Coverage asks whether the correct structure is included. Ranking asks where it appears. A ranker cannot retrieve an answer absent from its candidate pool.
What do C1, C2, and C3 mean here?
Our local evaluation regimes: C1 tests direct-library availability with exact-identity holdouts; C2 tests scaffold-held-out retrieval for COCONUT members; C3 tests scaffold-held-out cases beyond that subset. They are research diagnostics, not separate official leaderboard scores.
Why can a promising experiment fail?
A change can overfit the development molecules, exploit training leakage, miss the test instrument’s chemistry, or benefit a different candidate distribution. A confidence interval crossing zero means the measured improvement remains uncertain under that analysis.