Parameter-Free Molecular Classification and Regression with Gzip

Quickstart

Our implementation, MolZip, is available as as a PyPI package, simply install it using pip:

pip install molzip

After that, it is easy to get started with compression-based classification and regression of molecules.

from molzip import ZipClassifier, ZipRegressor

zc = ZipClassifier()
zr = ZipRegressor()

query = ["CNCO", "COCCCC"]

pred = zc.fit_predict(["CNC", "COC", "CCC"], [1, 2, 3], query, k=2)
multilabel_pred = zc.fit_predict(["CNC", "COC", "CCC"], [[1, 4], [2, 5], [3, 6]], query, k=2)

pred = zr.fit_predict(["CNC", "COC", "CCC"], [1.0, 1.5, 2.0], query, k=2)
pred_multitarget = zr.fit_predict(["CNC", "COC", "CCC"], [[1.0, 2.5], [1.5, 1.0], [2.0, 0.25]], query, k=2)

PDBBind Data

1. Download data form http://www.pdbbind.org.cn
2. Specifically, the refind set PDBbind_v2016_refined.tar.gz
3. and the core set CASF-2016.tar.gz
4. From the core set, only extract the required files (the folder coreset)
5. Finally, also download the index file (INDEX_refined_data.2016from PDBbind_2016_plain_text_index.tar.gz)
6. You can then use the scriptpreprocess_pdbbind.py to prepare the data e.g.:

python scripts/preprocess_pdbbind.py data/pdbbind/refined/ data/pdbbind/core/ data/pdbbind/INDEX_structure.2020 data/pdbbind/pdbbind.csv

This script will remove samples found in the core set from the refined set and create a CSV file containing amino-acid sequences and SMILES.

Reproducing results

The configurations to reproduce the results can be found in the *_config_*.py files in the root directory and are loaded on line 20 in main.py.

Abstract

TBD

Introduction

The classification of a molecule on a wide variety of physicochemical and pharmakological properties, such as solubility, efficacy against specific diseases, or toxicity, has become a task of high interest in chemistry and biology. With the rise of deep learning during tha past decade, molecular classification has increasingly be carried out by ever-larger models, with mixed results. The newly published parameter-free text classification approach that makes use of Gzip compression has shown excellent performance compared to deep learning architectures, such as transformers, on benchmark data sets.[^1] As the SMILES string encoding of molecular graphs has been shown to be a well-performing molecular representation for applying NLP methods, such as transformers, to chemical tasks including molecular classification, a comparison with the Gzip-based classification method is also relevant in the context of molecular classification.

Methods

The Gzip-based classifier introduced in this article has been adapted from the implementation presented by Jiang et al. and differs in three points: (1) as, as the authors have noted, the Gzip-based classification method has a relatively high time complexity, multiprocessing has been added; (2) multi-task classification has been added; and (3) a class weighing scheme has been implemented to account for unbalanced data. Furthermore, the capability to preprocess data, in this case the SMILES strings, has been added to the calling program.

Results

The current results are presented in the table below. Data sets with random splits were ran a total of four times.

Using SECFP (ECFP-style circular substructures as SMILES) doesn't increase the classification performance of the weighted kNN.

Implementing a GZip-based regressor (weighted kNN, k=10) shows performance comparable to baseline performance of common ML implementations from MoleculeNet (https://moleculenet.org/full-results). Interestingly there are improvements when the SMILES are tokenised.

The classifier is also able to classify raw reaction SMILES from the Schneider50k data set (no class weighting).