GC-MS Deconvolution Project

Exploring chemometrics with synthetic data and machine learning

Disclaimer: I'm learning as I go here — I have no formal background in analytical chemistry or chemometrics. This is very much a "figure it out as you build it" project, and nothing here should be taken as state of the art. If you spot something wrong or know a better way, I'd love to hear about it!

• Part 0: Why Deconvolute?

  What happens when molecules overlap, and how separating them fixes identification.

• Part 1: Building a Synthetic GC-MS Data Generator

  From raw CDF files to a realistic data generator using real elution profiles and mass spectra.

• Part 2: Counting Components with SVD

  Using singular value decomposition and a random forest to estimate overlapping molecule count — 98.5% accuracy.

• Part 3: Recovering Elution Profiles

  Upcoming

• Part 4: Deconvoluting Real Peaks

  Upcoming — hopefully that works! 😱😂

• Part 5: Benchmarking Against Existing Tools

  Upcoming

By Jonas Berdoz · Data sources: Copenhagen Soft Camel Cheese GC-MS dataset, MassBank mass spectral library