Unknown Molecule Identification

with Probabilistic Deep Learning and Rotational Spectroscopy

Kelvin Lee,
Michael McCarthy

ISMS 2021—Talk FF02

Complex discharge mixtures

Incredibly rich spectroscopic discovery space of microwave discharge assays

Three components of analysis

Line identification

Line assignment

Molecule identification

**The most time consuming and ill-defined step!**

McCarthy & Lee, *JPCA*, 2020(124) 3002

Molecule identification

Experimentally determined spectroscopic parameters are uninformative

Mapping constants to structures

Databases

Machine learning

Mass spectrometry
Fast—nearest-neighbor lookup
Uniform accuracy
Static

Drug discovery
Slower, but more modeling flexibility
Sampling dependent accuracy
Generative

Deep learning as an attractive solution to generative molecule identification

Deep neural networks

Universal function approximators
Parameterize (non)deterministic mapping between constants and features
Human intuition takes with sufficient information
Variational inference

Dataset generation

Molecules systematically generated using Open Molecule Generator
Up to 2,000 structures per formula, up to $\mathrm{H_{18}C_8O_3N_3}$
$\omega$B97X-D/6-31G(d) singlet equilibrium structures
Rotational constants uncertainty used as training augmentation

Structures encoded as Coulomb matrix eigenspectra

$M_{ij} = \begin{cases} 0.5Z^{2.4}_i & \text{for}~i = j\\ \frac{Z_iZ_j}{\vert \mathbf{R}_i - \mathbf{R}_j} & \text{for}~i \neq j\\ \end{cases}$

Ten largest eigenvalues for structurally similar species

From experimental data to identifying features

No free lunch

Models trained on four subdatasets

Pure hydrocarbons

Oxygen-bearing species

Nitrogen-bearing species

Oxygen/nitrogen-bearing species

Quantitative testing

Use benzene as quantitative testing of model behaviors

Eigenspectrum regression

The most critical step—eigenspectra encodes approximate structure and atom composition

What matters most?

Input gradients indicate $\kappa$ and the dipole moments are most important to the structure

Decoding the eigenspectrum

Extracting identifying information from the Coulomb eigenspectrum encoding

Molecular formula

Rotational constants Formula
Hydrcarbon model predicts correct formula within uncertainty
Remaining models yield roughly the correct mass

Converted formulae are comparable to mass spectrometry!

Multiclass classification for functional group identification

Intuition from formula + functional group

Areas for improvement

Single model approach

Faster training/inference

Constants to molecular graph mapping

End-to-end pipeline for molecule identification

Available now!

Conclusions

Uncertainty aware deep learning model for molecule identification

Experimentally determinable parameters

Fast, functional interface in PySpecTools

AST-1615847

AST-1908576

NASA-NNX13AE59G

NASA-80NSSC18K0396

Smithsonian Institution Hydra Cluster

Acknowledgements

Thank you!

github.com/laserkelvin

@cmmmsubmm

Google scholar

Rich complex mixtures provide a wealth of spectroscopic data.