PaperlayerView exact revision on arXiv

Source-linked AI summary

In silico generation of novel, drug-like chemical matter using the LSTM neural network

Peter Ertl, Richard Lewis, Eric Martin, Valery Polyakov

arXiv:1712.07449v2cs.LGq-bio.QM

TL;DR

Exhaustive enumeration becomes infeasible for larger drug-like molecules, motivating methods to explore novel chemical space. This study trained an LSTM on ChEMBL molecules to generate new structures. The generated molecules were novel and diverse, and virtual screening indicated comparable activity potential to the ChEMBL set.

  • Problem

    Exhaustive enumeration of molecules around the average drug-molecule size is not currently possible, while exploring novel chemical space is important to drug discovery.

    Source: Introduction · p. 1

  • Method

    The authors trained an LSTM neural network on ChEMBL molecular SMILES and generated new SMILES character by character.

    Source: Results · p. 3

  • Results

    The generated structures were novel and diverse, had properties and synthetic accessibility similar to ChEMBL molecules, and showed comparable activity potential in virtual screening.

    Source: Results · p. 5, Conclusions · p. 6

  • Takeaways & Limitations

    The generated structures may be used in virtual screening and in exploring novel areas of chemical space.

    Source: Conclusions · p. 6

    The selected architecture may not be optimal, although the authors considered it sufficient to produce good results.

    Source: Neural network architecture · p. 3

Abstract

from arXiv · show

The exploration of novel chemical spaces is one of the most important tasks of cheminformatics when supporting the drug discovery process. Properly designed and trained deep neural networks can provide a viable alternative to brute-force de novo approaches or various other machine-learning techniques for generating novel drug-like molecules. In this article we present a method to generate molecules using a long short-term memory (LSTM) neural network and provide an analysis of the results, including a virtual screening test. Using the network one million drug-like molecules were generated in 2 hours. The molecules are novel, diverse (contain numerous novel chemotypes), have good physicochemical properties and have good synthetic accessibility, even though these qualities were not specific constraints. Although novel, their structural features and functional groups remain closely within the drug-like space defined by the bioactive molecules from ChEMBL. Virtual screening using the profile QSAR approach confirms that the potential of these novel molecules to show bioactivity is comparable to the ChEMBL set from which they were derived. The molecule generator written in Python used in this study is available on request.

Introduction

Exploring novel chemical space supports drug discovery, but exhaustive molecule enumeration becomes infeasible as molecule size grows. The introduction presents LSTM networks as a way to learn molecular patterns and generate new structures, within a field already exploring neural molecule generation.

Introduction

  • Enumerating molecules of average drug-molecule size (25–30 atoms) is not currently possible, motivating algorithms that sample representative chemical space.The passage contrasts this limit with GDB-17’s 166 billion molecules.Source: Introduction · p. 1
  • LSTM networks can learn characteristics of training examples and generate similar new objects, making them candidates for generating novel chemical matter.Source: Introduction · pp. 1–2
  • Prior work used ZINC-trained LSTMs to generate molecules with retrosynthetic routes identified for most samples [7], while varied training SMILES representations improved results [8].Source: Introduction · p. 2

Methodology

The study selected an LSTM architecture through experimentation and trained it to predict the next character in filtered ChEMBL SMILES sequences. The described design combines recurrent layers, dropout, and a character-probability output.

Neural network architecture

  • Architecture selection required trial and error because no general rules specify optimal layer, activation, optimization, encoding, learning-rate, and dropout settings.Source: Neural network architecture · p. 2
  • The selected network used double LSTM layers, dropout, and a 23-neuron softmax output to predict the next SMILES character.The architecture was selected after experiments; Figure 1 depicts the chosen design.Source: Neural network architecture · p. 2, Figure 1 · p. 2
  • The authors describe the architecture as potentially non-optimal, while sufficient to produce good results.Source: Neural network architecture · p. 3

Training data

  • 509,000 bioactive ChEMBL molecules trained the network after filtering out charged molecules and stereochemistry and limiting ring closures to five.The corpus contained 23,664,668 characters, and training used 40-character sequences to predict the following character.Source: Training data · p. 3

Results

The generated molecules had novel scaffolds while retaining properties and structural features similar to ChEMBL molecules. Profile-QSAR screening found their predicted activity distributions broadly comparable to those of actual ChEMBL compounds.

Results

  • 627,000 unique scaffolds appeared among one million generated structures, with 18,000 scaffolds shared with the 172,000-scaffold ChEMBL training set.Only 2,774 generated molecules were identical to training-set molecules.Source: Results · p. 3
  • Generated and ChEMBL molecules had very similar molecular-weight, logP, and TPSA distributions, and nearly identical synthetic-accessibility-score distributions.The authors interpret the latter as indicating similar structural complexity.Source: Results · pp. 3, 5, Figure 3 · p. 4
  • LSTM-generated substructure features resembled ChEMBL, whereas the naive SMILES baseline produced mostly macrocycles and practically no fused aromatic systems.The baseline did not reproduce the chemical space of ChEMBL molecules.Source: Results · pp. 3–4, Table 1 · p. 4
  • The generated and ChEMBL sets had very similar distributions of common functional groups.Figure 4 reports the ChEMBL percentage first and the generated-molecule percentage second.Source: Results · p. 5, Figure 4 · p. 5

Virtual screening

  • Predicted pIC50 cumulative-frequency curves for actual and generated compounds nearly overlapped across the six profile-QSAR models.The models had R2 > 0.75 on held-out test sets of highly novel compounds.Source: Figure 5 · p. 5, Virtual screening · p. 6
  • For 4 of 6 assays, the KS test allowed the same-distribution null with 95% confidence; assays 809247 and 809170 rejected it.Even for assay 809170, the largest vertical distance between distributions was <10%.Source: Virtual screening · p. 6, Table 2 · p. 6

Conclusions

The authors conclude that an LSTM trained on ChEMBL drug-like molecules generated novel, diverse structures with good properties and synthetic accessibility. The generated molecules occupy the same chemical-space area as known bioactive molecules, and virtual screening indicates comparable activity probability across targets.

Conclusions

  • LSTM-generated structures are novel and diverse, with good properties and synthetic accessibility.The training set contained 550 thousand drug-like ChEMBL molecules.Source: Conclusions · p. 6
  • The generated molecules occupy the same chemical-space area as known bioactive molecules, despite being novel.Source: Conclusions · p. 6
  • Virtual screening indicates comparable probability of activity on various targets for generated molecules and known compounds.Source: Conclusions · p. 6
  • The algorithm is sufficient to generate hundreds of millions or even billions of new drug-like molecules.The authors identify virtual screening and exploration of novel chemical-universe areas as possible uses.Source: Conclusions · p. 6

Author’s contributions

PE designed the network and generated and analyzed the molecules; RL wrote the naive SMILES generator. EM and VP conducted and analyzed virtual screening, and all authors wrote the article.

Author’s contributions

Loading 1712.07449v2…