This is the code repo for the paper Highly accurate and large-scale collision cross section prediction with graph neural network for compound identification. We developed a method named Structure included graph merging with adduct method for CCS prediction (SigmaCCS), and a dataset including 282 million CCS values for three different ion adducts ([M+H]+, [M+Na]+ and [M-H]-) of 94 million compounds. For each molecule, there are "Pubchem ID", "SMILES", "InChi", "Inchikey", "Molecular Weight", "Exact Mass", "Formula" and predicted CCS values of three adduct ion types.
We recommend to use conda and pip.
By using the requirements/conda/environment.yml, requirements/pip/requirements.txt file, it will install all the required packages.
git clone https://github.com/zmzhang/SigmaCCS.git
cd SigmaCCS
conda env create -f requirements/conda/environment.yml
conda activate sigma
SigmaCCS is a model for predicting CCS based on graph neural networks, so we need to convert SMILES strings to Graph. The related method is shown in sigma/GraphData.py
1. Generate 3D conformations of molecules.
mol = Chem.MolFromSmiles(smiles)
mol = Chem.AddHs(mol)
ps = AllChem.ETKDGv3()
ps.randomSeed = -1
ps.maxAttempts = 1
ps.numThreads = 0
ps.useRandomCoords = True
re = AllChem.EmbedMultipleConfs(mol, numConfs = 1, params = ps)
re = AllChem.MMFFOptimizeMoleculeConfs(mol, numThreads = 0)
- ETKDGv3 Returns an EmbedParameters object for the ETKDG method - version 3 (macrocycles).
- EmbedMultipleConfs, use distance geometry to obtain multiple sets of coordinates for a molecule.
- MMFFOptimizeMoleculeConfs, uses MMFF to optimize all of a molecule’s conformations
2. Save relevant parameters. For details, see sigma/parameter.py.
- adduct set
- atoms set
- Minimum value in atomic coordinates
- Maximum value in atomic coordinates
3. Generate the Graph dataset. Generate the three matrices used to construct the Graph:
(1) node feature matrix, (2) adjacency matrix, (3) edge feature matrix.
adj, features, edge_features = convertToGraph(smiles, Coordinate, All_Atoms)
DataSet = MyDataset(features, adj, edge_features, ccs)
Optionnal args
- All_Atoms : The set of all elements in the dataset
- Coordinate : Array of coordinates of all molecules
- features : Node feature matrix
- adj : Adjacency matrix
- edge_features : Edge feature matrix
Train the model based on your own training dataset with Model_train function.
Model_train(ifile, ParameterPath, ofile, ofileDataPath, EPOCHS, BATCHS, Vis, All_Atoms=[], adduct_SET=[])
Optionnal args
- ifile : File path for storing the data of smiles and adduct.
- ofile : File path where the model is stored.
- ParameterPath : Save path of related data parameters.
- ofileDataPath : File path for storing model parameter data.
The CCS prediction of the molecule is obtained by feeding the Graph and Adduct into the already trained SigmaCCS model with Model_prediction function.
Model_prediction(ifile, ParameterPath, mfileh5, ofile, Isevaluate = 0)
Optionnal args
- ifile : File path for storing the data of smiles and adduct
- ParameterPath : File path for storing model parameter data
- mfileh5 : File path where the model is stored
- ofile : Path to save ccs prediction values
The example codes for usage is included in the test.ipynb
The following files are in the others folder:
- Attribute Analysis.ipynb. analyze the attribute importance
- UMAP.ipynb. visualize the learned representation with UMAP
- UMAPDataset.py. for generating graph datasets.
- theoretical calculation.ipynb. investigate of the relationship between SigmaCCS and theoretical calculation
- Filtering.ipynb. Filtering of target unknown molecules based on the CCS and mz of the molecules
- Performance on the test set with large rotations.ipynb. Performance of SigmaCCS on the test set and the external test set with different coordinates generated by ETKDG and MMFF94
- Completely random rotations.ipynb. Performance on the test set with completely random rotations
- Visualization of the 3D conformers using different seeds of ETKDG.ipynb
- RotationMatrix.py. define rotation matrix for rotating the molecules
- CFM-ID4. the code for generating MS/MS spectra with CFM-ID 4.0.
- GNN-RT:
- model:
- model.h5
- data:
- Attribute importance data
- Attribute importance (data.csv)
- Coordinate data (Store the 3D coordinate data of all molecules in data.csv)
- UMAP data
- data.csv
- data_molvec.npy (Molecular vectors of all molecules)
- data-UMAP-EUC-60.npy
- Coordinate data (Store the 3D coordinate data of all molecules in data.csv)
- theoretical calculation data
- data.csv
- LJ_data.csv (Get the LJ interaction parameters of different elements according to LJ_data.csv)
- Coordinate data (Store the 3D coordinate data of all molecules in data.csv)
- Filtering data
- data.csv
- Attribute importance data
- MultidimensionalFiltering:
- Example.ipynb. Taken the lipid (PubChem CID: 114944) as an example to show the multidimensional filtering assisted by SigmaCCS
- LipidBlastPredCCS.xlsx. the predicted CCS values of the lipids in the LipidBlast
- MListWithPredictedRTs_329.xlsx. retention times of the molecules in the MList of the lipid (PubChem CID: 114944). Please refer to github GNN-RT repository for details of RT prediction
- Mouse_lung_adduct_negative.xlsx. the mouse lung dataset with 761 lipids in negative ion mode after the removal of unpredictable adducts and empty SMILES strings
- UMAP 0.5.1
slurm script for generating CCS of PubChem in HPC cluster.
The following files are in the slurm folder
- mp.py
- multiple_job.sh (Batch generation of slurm script files)
- normal_job.sh (Submit the slurm script for the mp.py file)