All our code is in this git, and our website SPNPS is also deployed from this file. By modifying the fn parameter in the “~\backend\getPred.py”, you can debug the complete process of the website. The following contents will describe the key parts in detail:
The input of the neural network must be SMILES after atomic mapping. The preprocessing needs python environment with dependencies: Rdkit, Rxnmapper
In this git, we provide a script to convert the step in “~\backend\data\prodata.py” by command: python prodata.py where you just need to input the filename in main function. Files are organized as one reaction per line.
The network is trained in two steps:
1 Reaction Center: The following commands need to be execute in directory: “~\backend\data\core_wln_global”. Train command: python nntrain_direct.py --train ../data/train.txt.proc --hidden 300 --depth 3 --save_dir model-fine-tune-ori where “train.txt.proc” is your own training reactions after atom mapping.
Test command: python nntest_direct.py --test ../data/test.txt.proc --hidden 300 --depth 3 --model model-fine-tune-3130 --checkpoint ckpt-3130 --verbose 1 --detailed 1 > model-fine-tune-3130/test.cbond_detailed where “test.txt.proc” is your own test reactions after atom mapping, ckpt-3130 is one of the trained models, you need to test and find the best one before executing the next step. When you decide the best model for this step, you need to generate the reaction center predictions for training samples by: python nntest_direct.py --test ../data/train.txt.proc --hidden 300 --depth 3 --model model-fine-tune-3130 --checkpoint ckpt-3130 --verbose 1 --detailed 1 > model-fine-tune-3130/train.cbond_detailed
2 Candidate Scoring: The following commands need to be execute in directory: “~\backend\data\ rank_diff_wln”.
Train command: python nntrain_direct_useScores.py --train ../data/train.txt.proc --cand ../core_wln_global/model-fine-tune-3130/train.cbond_detailed --hidden 500 --depth 3 --ncand 150 --ncore 16 --save_dir model-fine-tune-ori
Test command: python nntest_direct_useScores.py --test ../data/test.txt.proc --cand ../core_wln_global/model-fine-tune-3130/test.cbond_detailed --hidden 500 --depth 3 --ncand 1500 --ncore 16 --model model-fine-tune-ori --checkpoint ckpt-13800 --verbose 1 > model-fine-tune-ori/test.cbond_detailed_13800 where ckpt-13800 is one of the trained models for this setp, you need to test and find the best one before executing the next step.
After you decided the best model, you can generate the predictions by: python eval_by_smiles.py --gold ../data/test.txt.proc --pred model-fine-tune-ori/ test.cbond_detailed_13800 --bonds_as_doubles true
Then the prediction would generate in “rank_diff_wln” folder. Now, you can evaluate the accuracy of your model on your test set. After fine-tuning, you can get the final prediction model adapted to your data set. The model will be used for subsequent virtual product predictions
CFM-ID is provided as a Docker container for users, so a working Docker installation on Windows or Linux is a prerequisite. To install Docker on Windows, Linux, or Mac, follow the instructions at: https://docs.docker.com/get-docker/. After verifying that you have a running Docker installation, pull the latest CFM-ID container: docker pull wishartlab/cfmid
Run CFM-predict in a Docker Container. Command line:
sudo docker run --rm=true -v $(pwd):/cfmid/public/ -d wishartlab/cfmid:latest sh -c "cd /cfmid/public/; cfm-predict <input_file> <prob_thresh> <param_file> <config_file> <annotate_fragments> <output_file> <suppress_exceptions>"
input_file: A .txt file that contains the SMILES of the virtual compounds generated by the AI structure prediction program. The virtual compounds should be listed as space-separated (id, SMILES) pairs, one per line.
prob_thresh: The probability below which to prune unlikely fragmentations during fragmentation graph generation (default 0.001).
param_file: The filename of the parameters of a trained CFM model. This file is the output of CFM-train. For the specific details regarding CFM-train, please see https://sourceforge.net/p/cfm-id/code/HEAD/tree/supplementary_material/cfm-train_example/.
config_file: The filename of the configuration parameters of the CFM model.
annotate_fragments: The default value of 0 indicates that the fragment information in the output spectra is ignored.
output_file: The filename of the output spectra (in-silico database, ISDB) file.
suppress_exceptions: The default value of 0 indicates that most exceptions are suppressed, allowing the program to return normally even in cases where it fails to produce a result.
Example: sudo docker run --rm=true -v $(pwd):/cfmid/public/ -d wishartlab/cfmid:latest sh -c "cd /cfmid/public/; cfm-predict cfm_input.txt 0.001 param_output-1.log param_config.txt 0 cfmid_output.mgf 0"
Tremolo is a program running on Linux, freely available at: http://proteomics.ucsd.edu/Software/Tremolo/#Downloads.
Command line: ./convert <input_file>.mgf input_file: The filename of the experimental MS/MS data. Convert the experimental MS/MS data into a .pklbin file in the current directory.
./main_execmodule ExecSpectralLibrarySearch ./.params
parameters: The filename of the matching parameters.
The important matching parameters to edit are seen as follows:
EXISTING_LIBRARY_MGF: The filename of ISDB.
SEARCHSPECTRA: The filename of the experimental MS/MS data.
RESULTS_DIR: The filename of the output file (list of predicted compounds).
TOLERANCE: ± tolerance for parent mass search in Da. Set a small tolerance for dereplication using parent ion mass as a prefilter.
SCORE_THRESHOLD: A relatively high threshold can be set to enhance the accuracy of matching results. Typically, >= 0.5.
TOP_K_RESULTS: Defines the maximal number of results returned.