Reliable Molecular Retrieval from Mass Spectra Using Conformal Prediction.

Area of Science:

• Computational chemistry
• Cheminformatics
• Spectrometry analysis

Background:

• Identifying molecular structures from LC-MS/MS data is crucial for chemical analysis.
• Current methods using fingerprint similarity lack spectrum-specific reliability.
• Standard metrics like top-k accuracy do not provide confidence for individual spectra.

Purpose of the Study:

• To apply conformal prediction for spectrum-specific molecular structure identification.
• To construct prediction sets with user-defined probability guarantees.
• To evaluate conformal prediction's performance under various data distribution scenarios.

Main Methods:

• Utilized conformal prediction (marginal and conditional) for candidate-based molecular retrieval.
• Evaluated performance on the MassSpecGym benchmark across in-distribution, partially shifted, and out-of-distribution settings.
• Assessed the impact of data alignment and confidence-based grouping on prediction set size and reliability.

Main Results:

• Conformal prediction achieved target coverage with small prediction sets when data distributions aligned.
• Under distribution shift, prediction sets enlarged, reflecting increased ambiguity.
• Conditional conformal prediction enhanced reliability for spectra of varying difficulty, especially with confidence-based grouping.

Conclusions:

• Conformal prediction transforms molecular rankings into reliable, spectrum-specific candidate sets.
• The method offers an explicit trade-off between reliability and efficiency.
• This approach provides valuable spectrum-specific reliability statements previously unavailable.