De Novo Structure Prediction from Tandem Mass Spectra: Algorithms, Benchmarks, and Limitations.

Area of Science:

• Analytical Chemistry
• Computational Chemistry
• Cheminformatics

Background:

• Molecule identification from analytical data is vital for drug discovery and metabolomics.
• Tandem mass spectrometry generates structural fingerprints, but most spectra lack library references.
• De novo generative models aim to predict molecular structures but face accuracy assessment challenges.

Purpose of the Study:

• To critically analyze the accuracy of state-of-the-art de novo generative models for molecule identification.
• To investigate the impact of data leakage on reported model performance.
• To review the evolution of generative model architectures and propose a roadmap for future development.

Main Methods:

• Analysis of de novo generative models using leakage-controlled benchmarks (MassSpecGym).
• Review of three architectural eras: RNNs, sequence models, and graph-native diffusion.
• Evaluation of formula-conditioned vs. unconstrained generation and scaffold-based approaches.

Main Results:

• State-of-the-art models exhibit only 4.1% top-10 accuracy on rigorously controlled benchmarks, contrary to prior reports.
• Data leakage in naive data splits is identified as a primary cause for inflated performance metrics.
• Explicitly conditioning models on molecular formulas significantly improves exact-match accuracy.

Conclusions:

• Current de novo generative models for molecule identification require significant improvement in accuracy.
• Standardized, leakage-aware benchmarking and transparent reporting are essential for reliable tool development.
• Formula-conditioned generation shows promise for unknown discovery, while scaffold-based methods face bottlenecks with predicted scaffolds.