CCSBase2

A Collision Cross Section prediction model using a XGBoostRegressor model.

Documentation

• Fingerprint explorer — hosted on Streamlit: https://ccsbase2-fingerprint-explorer.streamlit.app/

Dependencies

1. Create a Conda environment (Python 3.12)

conda create -n ccsbase2 python=3.12 -y
conda activate ccsbase2

2. Install

pip install numpy pandas scikit-learn rdkit xgboost joblib requests matplotlib streamlit shap

Building Database

Build Main Database

CCSbase2 aggregates across the 5 different datasets listed below. AllCCS was obtained by downloading chemicals manually using their online database.

• CCSBase
• PNNL
• ALLCCS
• METLIN-CCS
• Dataset found from ACS publication

Download Model Weights

Please note that building the database from scratch takes a very long time as thousands of API calls need to be made. Instructions have been given below.

Run the code below in another file. We advise you to use Google Colab due to PubChem rate limiting IP addresses which results in a "ServerBusy" error. You may need to just call ccsml.find_smiles() multiple times while commenting out the rest of the method calls.

After determining as many SMILES, call ccsml.find_inchikey() at the end which calls PubChem API to retrieve InChiKey when given SMILES string. This is to obtain InChiKeys for CCSbase datapoints since CCSbase stores SMILES, but not InChiKeys.

ClassyFire also does rate limiting, so call ccsml.find_inchikey() multiple times while commenting out all other method calls. NOTE: APIs will not find SMILES string and/or subclass for all chemicals.

from data import CCSDataIntegration

ccsml = CCSDataIntegration("CCSMLDatabase.db")
ccsml.add_acs()
ccsml.add_ccsbase()
ccsml.add_allccs()
ccsml.add_metlin()
ccsml.add_pnnl()
ccsml.find_smiles()
ccsml.find_inchikey()
ccsml.find_classes()
ccsml.clean()
ccsml.build_ood_dataset("ood_testset.csv")

Classify Unknown Subclasses

ClassyFire will not identify subclasses for all chemicals, so a XGBClassifier was developed to classify the unknown subclasses. All subclasses that are known are grouped and groups less than less than 30 chemicals are not included during training and act as a proxy out of distribution (OOD) dataset.

from classifier import SubclassClassifier

classifier = SubclassClassifier(
    "CCSMLDatabase.db", seed=26, fp_vocab_file="ccsbase2_fp_vocab.joblib",
    min_subclass_count=30, novelty_threshold=0.7,
    n_estimators=[5000, 5500, 6000],
    max_depth=[8, 9],
    learning_rate=[0.03, 0.05],
    subsample=[0.9],
    colsample_bytree=[0.9],
    reg_lambda=[30],
    gamma=[1],
    min_child_weight=[5],
)

classifier.fit()
classifier.eval() # evaluates on proxy OOD set
classifier.predict() # performes inference on compounds with unknown subclasses

Training Model

Run the following code in another file. Set use_metlin=False to train without METLIN dataset.

from train import CCSBase2

ccs_model = CCSBase2("CCSMLDatabase.db",
                    n_estimators=[6000, 7000, 8000, 10000, 12000, 14000, 20000],
                    max_depth=[8, 10, 12, 13, 15],
                    learning_rate=[0.01, 0.02, 0.03],
                    subsample=[0.9],
                    colsample_bytree=[0.5, 0.9],
                    reg_lambda=[30],
                    min_child_weight=[1],
                    gamma=[1],
                    seed=26,
                    use_metlin=True,
                    fp_vocab_file="ccsbase2_fp_vocab.joblib"
                )
ccs_model.fit()
ccs_model.eval() # evaluates on test_data.csv

Inference

To perform inference, create a .csv file with column names smi,adduct and use the CCSBase2 predict() method to pass in your .csv file. Optionally, if you're looking to evaluate the model on you're own dataset, pass the column name in your .csv that holds the ground truth CCS as a second parameter to predict().

ccs_model = CCSBase2("CCSMLDatabase.db",
                    n_estimators=[6000, 7000, 8000, 10000, 12000, 14000, 20000],
                    max_depth=[8, 10, 12, 13, 15],
                    learning_rate=[0.01, 0.02, 0.03],
                    subsample=[0.9],
                    colsample_bytree=[0.5, 0.9],
                    reg_lambda=[30],
                    min_child_weight=[1],
                    gamma=[1],
                    seed=26,
                    use_metlin=True,
                    fp_vocab_file="ccsbase2_fp_vocab.joblib"
                )

ccs_model.predict("./datasets/ood_testset.csv", "ccs")