Problem
For each MS/MS spectrum of an unknown natural product, rank candidate structures so that the right one appears as high as possible in the top 25. The candidates are not given: they have to be retrieved from structure libraries first.
selected case study
Kaggle Competitions 2026: CASMI and Playground S6E10
Two live Kaggle competitions worked end to end: Enveda CASMI 2026, a featured competition that ranks candidate molecular structures for MS/MS spectra, and Playground S6E10, a tabular ROC-AUC competition on airline satisfaction. In both, decisions were made on cross-validation and the public leaderboard was treated as a noisy sample.
Quick technical review
The case study in four points.
Architecture
An own pipeline first: spectral retrieval, analog propagation from library neighbours, a bond-cleavage fragment explanation, biotransformations of spectral analogs as extra candidates, and a LightGBM LambdaRank ranker validated on 600 frozen queries…
Read complete note
An own pipeline first: spectral retrieval, analog propagation from library neighbours, a bond-cleavage fragment explanation, biotransformations of spectral analogs as extra candidates, and a LightGBM LambdaRank ranker validated on 600 frozen queries from libraries it never trained on. It reached a public score of 0.309. The community then built a public pipeline around spectrum simulation (ICEBERG and GLACIER) and a PubChem popularity prior that scores 0.420; the current entry fuses that pipeline with this ranker as an extra rank list and scores 0.422, 8 Oct 2026 snapshot rank 353 of 2,797.
Results
CASMI 2026: public 0.422, 8 Oct 2026 snapshot rank 353 of 2,797
Limitations & scope
Public-leaderboard ranks move daily and are not final. The CASMI score of 0.422 rests on a public community pipeline; the own pipeline alone scores 0.309. The S6E10 stack includes predictions published by other Kagglers, credited in the repository;…
Read complete note
Public-leaderboard ranks move daily and are not final. The CASMI score of 0.422 rests on a public community pipeline; the own pipeline alone scores 0.309. The S6E10 stack includes predictions published by other Kagglers, credited in the repository; the foundation-model members were not trained here.
Summaries reuse the project’s documented evidence. Review the complete case study for evaluation boundaries.
Evidence-led visual review
Explore the actual work.
Concept art, dated captures and repository figures are identified separately.
Portfolio illustration
Original portfolio illustration. Diagrams are conceptual; verified methods and results appear in the linked case study. This is not an application screenshot.
CASMI 2026: public 0.422, 8 Oct 2026 snapshot rank 353 of 2,797
Own CASMI pipeline alone: 0.309
S6E10: public 0.96177, 8 Oct 2026 snapshot rank 34 of 1,126
S6E10 nested stack CV 0.96211
Project proof
Verified external evidence
Last verified 2026-10-08
document proof
CASMI notebook: your ranker learns how you built the pool
document proof
CASMI notebook: a bigger candidate pool lowered my score
document proof
S6E10 notebook: why your OOF blend gain vanishes on test
document proof
S6E10 notebook: will your CV transfer?
document proof
Enveda CASMI 2026 competition
document proof
Playground S6E10 competition
Problem and approach
CASMI 2026: problem
For each MS/MS spectrum of an unknown natural product, rank candidate structures so that the right one appears as high as possible in the top 25. The candidates are not given: they have to be retrieved from structure libraries first.
CASMI 2026: what was built
An own pipeline first: spectral retrieval, analog propagation from library neighbours, a bond-cleavage fragment explanation, biotransformations of spectral analogs as extra candidates, and a LightGBM LambdaRank ranker validated on 600 frozen queries from libraries it never trained on. It reached a public score of 0.309. The community then built a public pipeline around spectrum simulation (ICEBERG and GLACIER) and a PubChem popularity prior that scores 0.420; the current entry fuses that pipeline with this ranker as an extra rank list and scores 0.422, 8 Oct 2026 snapshot rank 353 of 2,797.
S6E10: what was built
29 own models (LightGBM, XGBoost, CatBoost and RealMLP) on features that treat Flight Distance as a route ID, use the original dataset as a teacher model rather than as extra rows, predict every rating from the other columns, and cross each rating with the passenger context; 10-fold versions of the strongest. They are stacked by logistic regression on logits with credited public predictions of tabular foundation models and other public OOF libraries. Nested stack CV 0.96211; public score 0.96177, 8 Oct 2026 snapshot rank 34 of 1,126.
Validation
Every stack score is nested: the combiner never scores rows it was fitted on. A held-out experiment showed why blends of reseeded models win on OOF and lose on the leaderboard (fold-averaged test predictions already contain that gain), and it is published as a notebook. The two final S6E10 submissions were chosen on CV only, one of them without any public OOF library, as a hedge against leaky public predictions.
What is not claimed
Public-leaderboard ranks move daily and are not final. The CASMI score of 0.422 rests on a public community pipeline; the own pipeline alone scores 0.309. The S6E10 stack includes predictions published by other Kagglers, credited in the repository; the foundation-model members were not trained here.
Architecture and analysis
- Treats the public leaderboard as a noisy sample and selects on nested cross-validation.
- Separates its own models from community pipelines and public predictions, and says which is which.
- Publishes the analysis behind each decision as Kaggle notebooks.
Testing and validation
- CASMI: retrieval, analog propagation, fragment explanation and a LambdaRank ranker, fused with the public spectrum-simulation pipeline
- S6E10: feature screen, Kaggle-GPU stack members, nested-CV stacker and the blend-gap experiment
Operational value
- CASMI 2026: public score 0.422, 8 Oct 2026 snapshot rank 353 of 2,797; the own pipeline alone reached 0.309.
- S6E10: public score 0.96177, 8 Oct 2026 snapshot rank 34 of 1,126, from a nested stack CV of 0.96211.
- Showed on held-out data that a seed blend gains about +0.00023 on OOF but about +0.00005 on fold-averaged test predictions.
How to inspect
- Read the blend-gap notebook before the scores: it explains how the S6E10 submissions were chosen.
- Read the CASMI pool notebook: a larger candidate pool made the score worse, and it says why.