dimenwarper/molxspec

ML models to convert molecules to ESI mass spectra and maybe back again

molxspec

Machine learning models to convert molecules to ESI mass spectra (and maybe back again in a future version) trained on GNPS data. Currently the following models are available:

mol2spec

model	description	hidden unit dim	num layers
mlp	MLP with residual blocks trained on 1024 Morgan fingerprints	1024	6
gcn	Simple GCN with deepchem like node features	1024	3
egnn	Equivariant GNN trained on (RDkit optimized) 3D structures	1024	2
bert	MLP trained on representations from the (smaller) ChemBERTa SMILES model	1024	6

Installation

You need to have torch and torch_geometric installed. I don't provide these as part of the dependencies since torch_geometric installs depends a lot on your CUDA and torch setup. To install torch_geometric from scratch use their documentation; e.g. can do it with pip using their wheels:

pip install torch-scatter torch-sparse torch-cluster torch-spline-conv torch-geometric -f https://data.pyg.org/whl/torch-1.9.0+cpu.html

Once torch and torch_geometric are installed, you can install molxspec:

pip install https://github.com/dimenwarper/molxspec/releases/download/v.0.1.0/molxspec-0.1.0-py3-none-any.whl

Usage

You can predict spectra from the command line:

mol2spec --model [mlp | gcn | egnn | bert] input_smiles.txt output.txt

Where input_smiles.txt is a file containing one molecule SMILES for each line. For the egnn model, molecules will have their 3D structure computed and optimized automatically using RDKit. First time use will download the pretrained models automatically, which can take some time, though it is a one-time thing only.

You can also predict spectra programmatically:

from molxspec import mol2spec
dict_of_smiles_and_spectra = mol2spec.predict(list_of_smiles)

molxspec's Issues

When I run the training_setup.py,I encountered a problem.

When I run training_ setup. py file, the system prompts checkpoint: optional [any] = none Attributeerror: module 'Torch' has no attribute 'model'.
Have you encountered this problem before? If yes, do you know how to solve it?

Fix duplicate adduct types

These are the adduct types being considered by the models:

ADDUCTS = ['[M+H]+', '[M+Na]+', 'M+H', 'M-H', '[M-H2O+H]+', '[M-H]-', '[M+NH4]+', 'M+NH4', 'M+Na']

As you can see, there are repeated adducts when you remove the brackets, due to different notations. We should merge these adduct types as the bracket version have lower samples in the GNPS data and may be artifactually biasing the results. This of course means retraining the models.