PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative chemistry: drug discovery with deep learning generative models

Yuemin Bian, Xiang-Qun Xie

arXiv:2008.09000v1q-bio.BMcs.LGq-bio.QM

TL;DR

Drug discovery is increasingly costly, creating a need for efficient molecular-design strategies. This review surveys generative chemistry infrastructure and deep-learning architectures for de novo compound generation, while highlighting reported rapid inhibitor discovery and persistent realism and synthesizability challenges.

  • Problem

    Drug discovery is costly and lengthy, motivating efficient strategies for accelerating the process.

  • Method

    The paper reviews chemical databases, molecular representations, machine-learning infrastructure, and RNN, VAE, AAE, and GAN architectures for generative chemistry.

  • Results

    Generative chemistry has identified DDR1 kinase inhibitors within 21 days, while the review finds deep-learning generative models promising for de novo molecular design.

  • Takeaways & Limitations

    Generative modeling offers a reviewed route toward more efficient molecular design within drug discovery.

  • Takeaways & Limitations

    Generative models can propose unrealistic molecules, and synthesizability remains a major hindrance to connecting them with medicinal chemistry.

Abstract

from arXiv · show

The de novo design of molecular structures using deep learning generative models introduces an encouraging solution to drug discovery in the face of the continuously increased cost of new drug development. From the generation of original texts, images, and videos, to the scratching of novel molecular structures, the incredible creativity of deep learning generative models surprised us about the height machine intelligence can achieve. The purpose of this paper is to review the latest advances in generative chemistry which relies on generative modeling to expedite the drug discovery process. This review starts with a brief history of artificial intelligence in drug discovery to outline this emerging paradigm. Commonly used chemical databases, molecular representations, and tools in cheminformatics and machine learning are covered as the infrastructure for the generative chemistry. The detailed discussions on utilizing cutting-edge generative architectures, including recurrent neural network, variational autoencoder, adversarial autoencoder, and generative adversarial network for compound generation are focused. Challenges and future perspectives follow.

1. INTRODUCTION

Drug discovery remains costly and lengthy, motivating deep learning generative models as a route to design molecules with desired properties and accelerate the process. This review surveys the field’s foundations, architectures, and applications in de novo drug design.

  • The average cost of developing a new drug reaches 2.6 billion USD, while the overall discovery process takes over 12 years.
  • Conventional structure-based and ligand-based approaches, virtual screening, and machine-learning pipelines support lead identification and compound modification.
  • Deep learning generative models offer fresh opportunities to design molecular structures with desired properties more efficiently.
  • Generative chemistry has reported DDR1 kinase inhibitor identification within 21 days using deep learning generative models.
  • The review covers AI evolution, cheminformatics and machine-learning infrastructures, and generative architectures for de novo drug design.

2. ARTIFICIAL INTELLIGENCE IN DRUG DISCOVERY

The paper traces drug-discovery AI from symbolic rules through machine learning and deep learning. It explains how these paradigms support prediction, screening, representation learning, and generative molecular design.

  • Symbolic AI uses explicit rules but has limited capability for blurry knowledge and exhaustive target-specific compound selection.
  • Machine learning derives rules from data and supports supervised, unsupervised, and reinforcement-learning approaches in drug discovery.
  • Supervised learning predicts labels or numeric activities, while unsupervised clustering can improve structural diversity when selecting compounds.
  • Deep learning uses successive neural-network layers to learn increasingly abstract representations from raw data.
  • Generative chemistry draws on deep-learning architectures, with LSTM text generation inspiring SMILES-based design and GAN image generation motivating molecular structure generation.

3. DATA SOURCES AND MACHINE LEARNING INFRASTRUCTURES

Generative chemistry depends on curated chemical and biological data together with machine-readable molecular representations. The section surveys databases, representations, standardization, and their practical limitations.

  • Deep-learning generative chemistry requires high-quality cheminformatics and bioinformatics data describing molecules and biological systems.
  • Drug-discovery databases provide chemical structures, bioactivity, biological-system, and purchasable-compound data for modeling and validation.
  • ChEMBL, DrugBank, and ZINC serve different needs, spanning bioactive compounds, marketed and discovery-phase drugs, and purchasable compounds.
  • Commercial compounds can expand structural diversity and facilitate experimental validation because they are readily accessible.
  • SMILES represents molecules as ASCII strings and is widely used with sequence-generation models after canonicalization and encoding.
  • Fingerprints support molecular classification but cannot directly reconstruct complete molecular structures.

RCSB PDB47

The reviewed resources include databases for biological structures, chemical information, and biomolecular binding data. These databases support structural and activity-related stages of drug discovery.

  • ChEMBL focuses on bioactive molecules, and PDBbind collects experimentally measured binding affinities for biomolecular complexes.

ChEMBL50

The reviewed infrastructure includes chemical databases and resources supporting molecular information collection, target-specific compound libraries, virtual screening, and drug discovery workflows.

  • ChEMBL-related resources cover patent compounds, experimentally measured binding affinities, drug data, and commercially available compound collections.
  • PDB and BindingDB provide experimentally measured structural or binding data for evaluating receptor–ligand interactions and target-focused compound sets.
  • ZINC and related resources support virtual screening, hit identification, fragment-based discovery, and commercially available compound validation.
  • The infrastructure also includes chemical information spanning NMR spectra, physicochemical properties, and biomolecular interactions.
  • Combinatorial databases and chemical rules expand chemical space while imposing chemical stability and synthetic feasibility constraints.

MACCS Keys60

The paper presents cheminformatics toolkits and machine-learning packages as infrastructure for processing molecular data and building predictive models.

  • After data collection and transformation, machine-learning models are applied for predictive analysis using packages such as Scikit-Learn.
  • RDKit, Open Babel, and CDK support chemical file handling, substructure searches, pattern searches, and molecular representation generation.
  • These cheminformatics toolkits can support virtual screening, structural similarity searches, and structure–activity relationship analysis.
  • Workflow environments connect cheminformatics tools as nodes and edges to automate recurring data-processing tasks.
  • TensorFlow, CNTK, Theano, and PyTorch handle low-level tensor operations and differentiation, whereas Keras provides a higher-level modular API.

4. GENERATIVE CHEMISTRY WITH THE RECURRENT NEURAL NETWORK (RNN)

RNN-based generative chemistry models process molecular sequences to propose structures, with LSTM carry tracks addressing long-term dependency and vanishing-gradient problems.

  • RNNs process sequence elements iteratively with an internal state, but long-term dependencies can cause vanishing gradients and hinder learning.
  • LSTM carry tracks transport information across sequence processing, countering gradual signal loss and enabling later access to earlier information.
  • Transfer learning first trains on large compound datasets and then fine-tunes models on smaller focused datasets for targeted generation.
  • LSTM models generate molecules by predicting successive SMILES characters and sampling until an ending character reconstructs molecular structures.
  • Around 90% of generated molecules were unique from known ligands and from one another in an LSTM transfer-learning study.
  • Four of five top-ranked synthesized compounds showed nanomolar to low micromolar activity in cell-based assays.

5. GENERATIVE CHEMISTRY WITH THE VARIATIONAL AUTOENCODER (VAE)

VAE-based generative chemistry maps molecules into continuous latent representations that can be sampled and decoded into novel structures or focused libraries.

  • Sampling and decoding latent representations can produce novel structures alongside inputs and support focused libraries with desired properties.
  • VAEs map molecules to statistical distributions in a continuous latent space, supporting structured sampling and molecular reconstruction.
  • CNN-based encoders with string encoding achieved superior performance in one automatic chemical design comparison.
  • Property-prediction modules can use encoder representations to optimize desired properties for newly generated molecules.
  • VAE latent spaces preserved chemical similarity principles, yielding molecules similar to known active compounds in a DRD2 study.
  • Molecular graphs pose representation challenges because structures have inconsistent numbers of nodes and edges, complicating their use as model inputs.

6. GENERATIVE CHEMISTRY WITH THE ADVERSARIAL AUTOENCODER (AAE)

AAEs combine an encoder–decoder with a discriminator that regularizes latent representations toward a prior distribution. Applications include fingerprint-based anticancer discovery, property-conditioned JAK3 design, and gene-expression-conditioned molecule generation.

  • Architecture and training: AAEs append a discriminator to an autoencoder, training encoder, decoder, and discriminator so the latent distribution q_φ(z|x) follows a chosen prior p(z).Training alternates reconstruction optimization, discrimination between q_φ(z) and p(z), and encoder minimization of adversarial cost.
  • Applications: AAE applications generated molecules for anticancer fingerprints, JAK3-related properties, and desired gene-expression changes.These studies used MACCS fingerprints, property conditioning, molecular docking and assays, or joint molecule–gene-expression modeling.
  • Applications: Conditional AAEs model shared and modality-specific features to make molecule generation consequential for gene-expression targets.The bidirectional model separates shared features from features exclusive to molecules or expression data before conditional decoding.
  • Related latent-space models: LatentGAN combines an autoencoder with a GAN, allowing adversarial optimization directly in latent space before decoding generated vectors into molecular structures.A heteroencoder maps one-hot SMILES to latent vectors, while a pretrained heteroencoder transfers generated vectors back to structures.

7. GENERATIVE CHEMISTRY WITH THE GENERATIVE ADVERSARIAL NETWORK (GAN)

GANs generate molecular strings, graphs, or fingerprints through adversarial generator–discriminator training, with CNNs often used to process local features. Reviewed applications target drug-like compounds, property optimization, and biological signatures, while instability and limited chemical-space coverage remain challenges.

  • GAN architecture and training: GANs train generators and discriminators iteratively: generators propose molecular representations while discriminators distinguish generated compounds from real data.The two losses assess discriminator classification and generator ability to fool the discriminator.
  • GAN architecture and training: GAN training can converge toward generated molecules that an established discriminator classifies as real, reflecting learned patterns from authentic compounds.The review notes that simultaneous optimization is difficult because instability can favor one network over the other.
  • Challenges: GANs may explore only the limited chemical space defined by real compounds, and unstable joint optimization can produce an overly strong generator or discriminator.These constraints are presented as limitations of adversarial molecular generation.
  • Applications: ORGAN adds domain-specific objectives to discriminator rewards so generated samples follow real-data distributions while optimizing selected chemical heuristics.The framework used drug-like and nondrug-like ZINC molecules, SMILES representations, a CNN discriminator, and an LSTM generator.
  • Applications: Reviewed GAN applications produced structures similar to inputs while improving considered properties, generated active-like molecules from target signatures, and produced drug-like or target-biased compounds.Applications include Mol-CycleGAN derivative design, gene-expression-conditioned generation, and latent-space compound design.

8. CONCLUSION AND FUTURE PERSPECTIVES

The review identifies generative chemistry’s promise for accelerating drug discovery while emphasizing unresolved challenges in representations, synthesizability, macromolecule generation, and closed-loop validation.

  • Synthetic feasibility: Generated molecules must be assessed for synthetic feasibility because high quantitative benchmark scores can still accompany unrealistic structures.Existing synthesizability evaluations rely on complex heuristic definitions, and small structural changes can alter synthetic pathways.
  • Molecular representations: Three-dimensional representations are needed because chirality, conformation, bond geometry, and spatial patterns affect molecular activity and feature extraction.Molecular graphs illustrate benefits of working directly with structures, while bond type, length, and angles improve spatial-pattern extraction.
  • Macromolecule generation: Current generative chemistry mainly targets small molecules, while protein and other macromolecule generation requires better representations and expanding protein-related databases.Protein folding and conformation are especially important for determining functionality.
  • Closed-loop automation: A closed-loop design-synthesis-test cycle can iteratively train generative and target-prediction models using experimentally validated active compounds.The proposed cycle generates candidates, ranks them for synthesis and assays, and appends true new actives to the hit set.
  • Review scope: The review covers generative chemistry infrastructure, deep-learning architectures, successful applications, challenges, and future perspectives for drug discovery.Its scope includes chemical databases, molecular representations, cheminformatics and machine-learning tools, and RNN, VAE, AAE, and GAN approaches.
Loading 2008.09000v1…