[Paper Review] Chemical Structure Elucidation from Mass Spectrometry by Matching Substructures

Jing Shan Lim, Joshua M. Wong+5 more|arXiv (Cornell University)|Nov 17, 2018

Computational Drug Discovery Methods10 references6 citations

TL;DR

This paper proposes an automated, data-driven approach to chemical structure elucidation from mass spectrometry by using neural networks to predict substructure presence from mass spectra, then matching these to candidate structures generated from a given molecular formula. The method achieves top-20 ranking for the correct structure in 88% and 71% of test cases for two chemical agent classes, significantly reducing manual effort in identification.

ABSTRACT

Chemical structure elucidation is a serious bottleneck in analytical chemistry today. We address the problem of identifying an unknown chemical threat given its mass spectrum and its chemical formula, a task which might take well trained chemists several days to complete. Given a chemical formula, there could be over a million possible candidate structures. We take a data driven approach to rank these structures by using neural networks to predict the presence of substructures given the mass spectrum, and matching these substructures to the candidate structures. Empirically, we evaluate our approach on a data set of chemical agents built for unknown chemical threat identification. We show that our substructure classifiers can attain over 90% micro F1-score, and we can find the correct structure among the top 20 candidates in 88% and 71% of test cases for two compound classes.

Motivation & Objective

• To address the bottleneck in analytical chemistry where manual chemical structure elucidation (CSE) from mass spectra and formulas can take days.
• To eliminate reliance on expert intervention in substructure selection and filtering by automating the identification workflow.
• To improve the speed and accuracy of identifying unknown chemical threats, particularly chemical agents under the Chemical Weapons Convention.
• To rank candidate structures efficiently using machine learning, enabling downstream computational validation on a shortlist.

Proposed method

• Train deep neural networks to predict the presence of substructures in a compound based solely on its mass spectrum.
• Use a data-driven approach to learn substructure patterns from mass spectral features without requiring manual spectral interpretation.
• Generate all possible candidate structures from the given molecular formula using MOLGEN.
• Match predicted substructures from the neural network to candidate structures to compute a relevance score for ranking.
• Apply multi-representative substructure classification (1 or 2 representatives per compound class) to improve detection of diverse functional groups.
• Rank candidate structures by substructure match confidence and evaluate top-k performance on test sets from the OPCW VGWD database.

Experimental results

Research questions

• RQ1 Can neural networks accurately predict the presence of substructures in unknown compounds using only mass spectral data?
• RQ2 Can substructure prediction from mass spectra effectively rank the correct chemical structure among millions of candidates?
• RQ3 How does the inclusion of multiple representative substructures (e.g., silylated vs. non-silylated) affect model performance and ranking accuracy?
• RQ4 To what extent can this automated workflow reduce the need for expert intervention in chemical structure elucidation?
• RQ5 How does the method compare to traditional approaches that rely on manual filtering and chemist-guided substructure selection?

Key findings

• The substructure classifiers achieved over 90% micro F1-score on the test data, indicating strong predictive performance.
• For the PPTN compound class, the correct structure was ranked in the top 20 candidates in 88% of test cases.
• For the PPN compound class, the correct structure was ranked in the top 20 in 71% of test cases.
• The use of two representatives (e.g., including a silylated variant) significantly improved ranking accuracy for compounds with complex functional groups.
• The method reduced the candidate list to under 20 structures in the majority of cases, enabling efficient downstream verification via simulation or NMR.
• Support vector machines performed worse than neural networks, with a 69.4% F1-score for PPTN (2 representatives) compared to 87.8% for neural networks.