molzip 0.9.2

An implementation of compression-based regression and classification for molecular SMILES.

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)

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.

Data Set	Split	AUROC (Valid)	F1 (Valid)	AUROC (Test)	F1 (Test)
bbbp	scaffold	0.891 +/- 0.0	0.902 +/- 0.0	0.679 +/- 0.0	0.686 +/- 0.0
bace_classification	random	0.793 +/- 0.038	0.793 +/- 0.038	0.789 +/- 0.038	0.789 +/- 0.038
clintox	random	0.805 +/- 0.038	0.965 +/- 0.038	0.77 +/- 0.038	0.958 +/- 0.038
tox21	random	0.6 +/- 0.007	0.308 +/- 0.007	0.599 +/- 0.007	0.303 +/- 0.007
sider	random	0.56 +/- 0.007	0.788 +/- 0.007	0.563 +/- 0.007	0.778 +/- 0.007

Implementing a weighted version of the kNN algorithm does not necessary lead to better classification performance on unbalanced data sets.

Data Set	Split	AUROC/RMSE (Valid)	F1/MAE (Valid)	AUROC/RMSE (Test)	F1/MAE (Test)
sider	scaffold	0.551 +/- 0.0	0.707 +/- 0.0	0.577 +/- 0.0	0.666 +/- 0.0
sider	random	0.454 +/- 0.262	0.657 +/- 0.262	0.581 +/- 0.262	0.647 +/- 0.262
bbbp	scaffold	0.931 +/- 0.0	0.931 +/- 0.0	0.639 +/- 0.0	0.627 +/- 0.0
bace_classification	scaffold	0.694 +/- 0.0	0.702 +/- 0.0	0.701 +/- 0.0	0.697 +/- 0.0
bace_classification	random	0.817 +/- 0.005	0.815 +/- 0.005	0.774 +/- 0.005	0.771 +/- 0.005
clintox	scaffold	0.805 +/- 0.0	0.854 +/- 0.0	0.891 +/- 0.0	0.891 +/- 0.0
clintox	random	0.925 +/- 0.032	0.924 +/- 0.032	0.913 +/- 0.032	0.91 +/- 0.032
tox21	scaffold	0.635 +/- 0.0	0.247 +/- 0.0	0.618 +/- 0.0	0.227 +/- 0.0
tox21	random	0.705 +/- 0.006	0.295 +/- 0.006	0.694 +/- 0.006	0.29 +/- 0.006
hiv	scaffold	0.714 +/- 0.0	0.901 +/- 0.0	0.689 +/- 0.0	0.887 +/- 0.0

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

Data Set	Split	AUROC (Valid)	F1 (Valid)	AUROC (Test)	F1 (Test)
bbbp	scaffold	0.83 +/- 0.0	0.819 +/- 0.0	0.632 +/- 0.0	0.627 +/- 0.0
bace_classification	random	0.833 +/- 0.015	0.829 +/- 0.015	0.826 +/- 0.015	0.821 +/- 0.015
clintox	random	0.74 +/- 0.076	0.831 +/- 0.076	0.747 +/- 0.076	0.84 +/- 0.076
tox21	random	0.712 +/- 0.011	0.305 +/- 0.011	0.718 +/- 0.011	0.31 +/- 0.011
sider	random	0.604 +/- 0.022	0.62 +/- 0.022	0.614 +/- 0.022	0.624 +/- 0.022

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.

Data Set	Split	AUROC/RMSE (Valid)	F1/MAE (Valid)	AUROC/RMSE (Test)	F1/MAE (Test)
freesolv	random	0.641 +/- 0.144	0.375 +/- 0.144	0.527 +/- 0.144	0.321 +/- 0.144
delaney	random	1.443 +/- 0.088	1.097 +/- 0.088	1.283 +/- 0.088	0.966 +/- 0.088
lipo	random	0.938 +/- 0.042	0.765 +/- 0.042	0.911 +/- 0.042	0.727 +/- 0.042

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

Data Set	Split	AUROC/RMSE (Valid)	F1/MAE (Valid)	AUROC/RMSE (Test)	F1/MAE (Test)
schneider	random	0.0 +/- 0.0	0.801 +/- 0.0	0.0 +/- 0.0	0.801 +/- 0.0

Discussion

TBD

References

[^1] https://arxiv.org/abs/2212.09410

What is this?

This is an experiment for a small open source manuscript/article that aims to validate and evaluate the performance of compression-based molecular classification using Gzip. If you want to join/help out, leave a message or a pull request that includes your name and, if available, your affiliation.