Skip to content
PaperFren

Can a text-style CNN read chemical formulas to predict toxicity?

Open paper intelligence

Treating a molecule's SMILES text string like a one-dimensional image let a CNN beat standard fingerprint methods on toxicity prediction and point to the substructures it relied on.

Source

Convolutional neural network based on SMILES representation of compounds for detecting chemical motif

Hirohara M, Saito Y, Koda Y, et al. · BMC bioinformatics · 2018

doi.org/10.1186/s12859-018-2523-5Read the full paper ↗109 citationscc by

Study at a glance

Design
Computational / modelling — New 1-D CNN on SMILES feature matrices evaluated by five-fold cross-validation and on the official Tox21 challenge split against fingerprint, graph-convolution and DeepTox baselines.
N
About 8000 compounds (the text says 'approximately') across 12 assay subdatasets of Tox21; no single exact N stated.
Population
Chemical compounds in the Tox21 benchmark with active/inactive labels for binding to 12 proteins
Outcome
ROC-AUC for active vs inactive classification; qualitative chemical-space separation and detected motifs

Structured fields used in claim comparison tables when every cited study has a complete layer.

What they did

The authors encoded each compound's SMILES string as a matrix with 42 features per symbol (atom properties or SMILES punctuation) and ran convolutions only along the string. The 64-number output, called a SMILES convolution fingerprint, fed a small classifier. They tested it on the Tox21 benchmark with five-fold cross-validation and on the official challenge split against the winning DeepTox models, then traced high-activation filters back to substructures.

What they found

In cross-validation the CNN had higher ROC-AUC than logistic regression, random forest and a deep network using ECFP fingerprints, and than graph convolution. On the official challenge split it beat a DeepTox network using ECFP alone but was slightly below DeepTox models that added thousands of hand-crafted toxicophore features. Its 64-dimensional fingerprint separated active from inactive compounds in a low-dimensional projection where 1024-bit ECFP did not, and one filter picked out a steroid-like motif known to matter for androgen-receptor binding.

The limits

What it doesn't show

The extracted text lacks the results tables, so exact ROC-AUC values and variability cannot be checked here. The model did not beat the best challenge entry, and no statistical tests of the differences are described. Motif interpretation is shown with selected examples from one subdataset, and detectable motif size is capped by the filter window. Attempts to handle the strong class imbalance did not help, and results are limited to one toxicity benchmark.

Key terms

SMILES
A text notation that writes a molecule's atoms, bonds, rings and branches as a single string of characters.
ECFP (circular fingerprint)
A fixed, hand-designed binary vector marking which small substructures around each atom a molecule contains, hashed into typically 1024-2048 bits.
ROC-AUC
Area under the receiver operating characteristic curve; the probability a classifier ranks a random active compound above a random inactive one (0.5 = chance, 1 = perfect).
Representation learning
Letting a model learn useful features from raw data during training instead of supplying hand-designed ones.
Global max pooling
Taking the single largest activation of each filter across the whole input, giving one number per filter.

Flashcards

1 / 10

0 of 10 answers reviewed

Research intelligence for this paper

See its role on concept claims, tensions it is part of, placement history, and related discoveries.

Open paper intelligence

Quiz yourself

1 / 5

In which direction do the CNN's convolutions move?

Common questions

Why use SMILES strings instead of molecular graphs?

A string is a regular one-dimensional grid, like a DNA sequence, so standard convolutions apply directly; graphs lack this grid structure and some graph methods cannot distinguish stereoisomers.

Did the CNN beat the Tox21 champion?

No. It beat the champion's model that used only ECFP but was slightly worse than versions using thousands of expert-designed toxicophore features.

How are 'chemical motifs' found?

Each fingerprint dimension comes from one filter; filters with unusually high z-scored activations for a compound are traced back to the part of the SMILES string, and hence the substructure, that triggered them.

More on Convolutional networks