DeePFAS: Deep Learning-Enabled Rapid Annotation of PFAS: Enhancing Non-Targeted Screening through Spectral Encoding and Latent Space Analysis

This repository provides implementations and code examples for DeePFAS: Deep Learning-Enabled Rapid Annotation of PFAS: Enhancing Non-Targeted Screening through Spectral Encoding and Latent Space Analysis. DeePFAS projects raw MS/MS data into the latent space of chemical structures for PFAS (Per- and Polyfluoroalkyl Substances) identification, facilitating the inference of structurally similar compounds by comparing spectra to multiple candidate molecules within this latent chemical space.

Quickstart

Option1: import module

from DeePFAS.models.inference import inference

ignore_MzRange = 'not ignore' # Whether to ignore the input spectrum mz peaks range (< 1000 Da) limitation ('not ignore', 'ignore')
ignore_CE = 'not ignore' # Whether to ignore the input spectrum collision energy (10 ~ 50 eV) limitation ('not ignore', 'ignore')

# Eval mode (the compounds corresponding to the spectra are known, and Canonical SMILES must be provided in the spectral files for result evaluation)
results = inference(dataset_path='./DeePFAS/dataset/smiles_dataset.tsv',
                    data_id_path='./DeePFAS/dataset/smiles_dataset.id',
                    data_file='./DeePFAS/example/testdata.mgf',
                    mode='eval',
                    topk=20,
                    out_dir='./results',
                    ignore_MzRange=ignore_MzRange,
                    ignore_CE=ignore_CE)

# Inference mode (unknown compounds, Canonical SMILES is not necessary)
results = inference(dataset_path='./DeePFAS/dataset/smiles_dataset.tsv',
                    data_id_path='./DeePFAS/dataset/smiles_dataset.id',
                    data_file='./DeePFAS/example/testdata.mgf',
                    mode='inference',
                    topk=20,
                    out_dir='./results',
                    ignore_MzRange=ignore_MzRange,
                    ignore_CE=ignore_CE)
print(results)

# statistic.pkl store results (python map)
# statistic.json store json format results
import pickle
with open('./results/statistic.pkl', 'rb') as f:
    statistic = pickle.load(f)
print(statistic)

Option2: run script

python3 ./script.py \
    --dataset_path './dataset/smiles_dataset.tsv' \
    --dataset_id_path './dataset/smiles_dataset.id' \
    --topk 20 \
    --data_file './example/testdata.mgf' \
    --mode inference \
    --out_dir './results' \
    --ignore_MzRange 'not ignore' \
    --ignore_CE 'not ignore'

Example of generateing candidates dataset

The example shows how to generate pickle file used in obtaining position of each SMILES in customized candidate dataset. See files in directory dataset.

def pickle_smiles(in_path, out_path):
    offsets = [0]
    with open(in_path, 'r', encoding='utf-8') as fp:
        while fp.readline() != '':
            offsets.append(fp.tell())
    offsets.pop()
    with open(out_path, 'wb') as f:
        pickle.dump(offsets, f)

pickle_smiles('./smiles_dataset.tsv', './smiles_dataset.id')

Example of statistic.json

This is an example of top3 candidates. Title in spectra data is key of .json file.

{
    "0": [
        {
            "loss": 11.045734405517578, 
            "smiles": "O=S(=O)([O-])C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)F",
            "mw": 348.93979885991,
            "mw_diff": 5.946375631272986
        },
        {
            "loss": 9.646349906921387,
            "smiles": "O=S(=O)(O)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)F",
            "mw": 349.94707531200004,
            "mw_diff": 6.953652083363011
        },
        {
            "loss": 9.632779121398926,
            "smiles": "O=C(O)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)F",
            "mw": 363.97689613200004,
            "mw_diff": 20.98347290336301
        },
    ]
}

Construct input spectra data (.mgf)

Example:

from pyteomics import mgf
import numpy as np
data = []

intensity = [0.1, 1.0, 0.3, 0.4]
m_z = [11.1, 23.23, 111.44, 55.2]
spectra = {
    'params': {
        # identifier of spectra in .mgf file (necessary)
        'title': 0,
        # ms level (necessary)
        'mslevel': 2,
        # precursor m/z (necessary)
        'pepmass': 562.957580566406,
        # adduct type (necessary)
        'precursor_type': '[M-H]-',
        # In eval mode, canonicalsmiless is necessary (unnecessary)
        'canonicalsmiles': 'O=C(O)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)C(F)(F)F',
        # collision energy (necessary)
        # absolute collision energy (ACE) format: 'collision_energy': 12
        # normalized collision energy (NCE) format: 'collision_energy': 'NCE=37.5%'
        'collision_energy': 'NCE=37.5%'
    },
    # m/z array (necessary)
    'm/z array': np.array(intensity), 
    # intensity array (necessary)
    'intensity array': np.array(m_z)
}

data.append(spectra)
mgf.write(data, 'spectra.mgf', file_mode='w', write_charges=False)