PaperlayerView exact revision on arXiv

Source-linked AI summary

Generative Models for Automatic Chemical Design

Daniel Schwalbe-Koda, Rafael Gómez-Bombarelli

arXiv:1907.01632v1cs.LGphysics.chem-phstat.ML

TL;DR

Conventional chemical discovery often relies on costly, incremental optimization, while inverse design seeks structures for chosen properties. This chapter reviews early inverse-design algorithms and generative models using molecular representations, and surveys their reported development alongside unresolved challenges in chemical generation.

  • Problem

    Conventional chemical discovery optimizes properties from known structures, while inverse design must search enormous spaces of discrete compounds subject to chemical constraints.

    Source: §1 · pp. 1–2

  • Method

    The chapter revisits early inverse-design algorithms, categorizes generative models by architecture and molecular representation, and reviews molecular-generation schemes.

    Source: §1.3 · pp. 7–8

  • Results

    The chapter surveys reported molecular-generation schemes and identifies continuing challenges in graph generation and the broader development of generative models for chemical design.

    Source: §2.2 · pp. 13–14, §3 · p. 15

  • Takeaways & Limitations

    Generative models offer ways to generate complex samples, optimize meaningful latent representations, and perform inference with supervised models in chemical design.

    Source: §1.3 · p. 7

    Graph generation remains an open problem, with scalability and graph reconstruction limitations, while chemical generative models still face generalization, real-world inference, and novelty challenges.

    Source: §2.2 · pp. 13–14, §3 · p. 15

Abstract

from arXiv · show

Materials discovery is decisive for tackling urgent challenges related to energy, the environment, health care and many others. In chemistry, conventional methodologies for innovation usually rely on expensive and incremental strategies to optimize properties from molecular structures. On the other hand, inverse approaches map properties to structures, thus expediting the design of novel useful compounds. In this chapter, we examine the way in which current deep generative models are addressing the inverse chemical discovery paradigm. We begin by revisiting early inverse design algorithms. Then, we introduce generative models for molecular systems and categorize them according to their architecture and molecular representation. Using this classification, we review the evolution and performance of important molecular generation schemes reported in the literature. Finally, we conclude highlighting the prospects and challenges of generative models as cutting edge tools in materials discovery.

1 Introduction

The chapter frames inverse chemical design as a response to costly, incremental discovery, then surveys early optimization strategies and deep generative models for molecular design. It focuses on model architectures and interpretable molecular representations, especially SMILES and molecular graphs.

1 Introduction

  • Conventional discovery derives properties from known structures through costly, incremental searches, whereas inverse design starts with target properties and infers structures amid vast, discrete, chemically constrained spaces.The accessible drug-like chemical space can be on the order of 10^60 molecules, and candidates must satisfy hard and soft constraints.Source: §1 · pp. 1–2

1.1 Early inverse design strategies for materials

  • Early Monte Carlo and evolutionary methods sample or refine candidates, while continuous representations support property-gradient optimization of molecular structures.Examples include simulated annealing and genetic algorithms, alongside approaches using continuous atomic or potential representations.Source: §1.1 · pp. 2–4
  • These strategies remain constrained by discrete search over rugged property surfaces, chemical validity and diversity trade-offs, and the gap between optimized atom clouds and stable, synthesizable molecules.The chapter identifies a need to sample valid, acceptable regions efficiently and optimize on the manifold of synthetically accessible, stable compounds.Source: §1.1 · p. 4

1.2 Deep learning and generative models

  • VAEs encode data into a latent distribution and decode samples, but modeled distributions may fail to match real data and common divergence penalties often assume Gaussian priors.Wasserstein Auto-Encoders are described as improving latent representations and reconstruction efficiency.Source: §1.2 · p. 5
  • GANs train a generator against a discriminator, but unstable min-max training and mode collapse can limit sample diversity.The generator maps a prior to candidate data; the discriminator distinguishes generated from real samples.Source: §1.2 · p. 6
  • Auto-regressive models factor data distributions into sequential conditional probabilities, enabling tractable likelihoods but making generation slow.The model generates each element conditioned on preceding elements.Source: §1.2 · p. 6

1.3 Generative models meet chemical design

  • For chemical design, the chapter focuses on interpretable SMILES strings and molecular graphs, since generating non-invertible descriptors such as fingerprints leaves molecule reconstruction as an extra step.The chapter presents generative models as useful for producing complex samples, learning latent representations for optimization, and coupling to supervised inference.Source: §1.3 · pp. 7–8, Figure 2 · p. 8

2 Chemical generative models

This section surveys molecular generation through SMILES sequences and molecular graphs, from recurrent and latent-space models to sequential and tree-decomposed graph generation.

2.1 SMILES representation

  • RNNs generate SMILES character by character, while transfer learning and reinforcement learning steer generation toward property targets or valid, rewarded sequences.Transfer learning first trains on a broad molecule dataset and then fine-tunes toward molecules with desired physicochemical properties.Source: §2.1 · pp. 9–10
  • VAEs map molecules into a continuous, differentiable latent space that supports property optimization and conditional generation without hand-made SMILES constraints.Semi-supervised methods can tune the model with relevant properties.Source: §2.1 · p. 10
  • Molecular GANs have been combined with stochastic policies to increase diversity, optimize properties, and maintain drug-likeness, but instability and mode collapse remain obstacles.Other approaches bias generated-molecule distributions toward goals or condition generation on gene-expression signatures.Source: §2.1 · p. 11
  • SMILES sequence models face difficult decoding, non-unique strings, and weak representation of molecular similarity; grammar constraints and decoder validators are among proposed fixes.Random latent points in one reported VAE had decoding success as low as 4%.Source: §2.1 · pp. 11–12

2.2 Molecular graphs

  • Sequential graph generators add nodes and edges, enabling conditional and scaffold-based molecule generation, but their outcomes depend on node ordering and random permutations.The decision process can target drug-likeness and synthetic accessibility.Source: §2.2 · pp. 12–13
  • JT-VAE achieves 100% validity when generating small molecules and 100% novelty when sampling latent codes from a prior; it also has a meaningful latent space essential for optimization and automatic molecular design.It encodes the full graph and its tree structure into two latent spaces and reconstructs the molecule from both representations.Source: §2.2 · p. 13
  • Beyond JT-VAE, DEFactor enables direct property-conditioned graph optimization without retraining, while other approaches use multi-objective reinforcement learning or sequence-based graph representations.DEFactor reports reconstruction accuracy above 80% for molecules with about 25 heavy atoms; one graph RL approach reports 100% validity by restricting actions to chemically valid ones.Source: §2.2 · pp. 13–14
  • Graph generation remains limited by mode collapse in GANs, open problems in arbitrary graph generation and larger-graph scalability, while grammar rules may improve validity and stereochemistry.Comparing graph isomorphism and measuring graph similarity can also require expensive methods.Source: §2.2 · pp. 14–15

3 Challenges and outlook for generative models

Generative models for chemical design remain early-stage: they must balance novelty, chemical validity, and application-specific usefulness while improving how molecules are represented and evaluated. The chapter sees promise in these models but stresses that addressing their challenges is important for accelerating materials discovery.

Challenges and outlook

  • Complete evaluation is difficult because the learned statistical distribution has no available expression; probing neural-network internals could help turn embeddings into logical rules.Source: §3 · p. 15
  • Chemical models must generalize, infer beyond training data, generate novel molecules, and capture synthetically accessible rules while avoiding mode collapse and naive interpolation.The field is still in its infancy, and emerging benchmarks aim to compare models beyond solubility or drug-likeness prediction.Source: §3 · p. 15
  • Chemical-space navigation alone is insufficient: models must generate valid molecules tailored to applications such as drug design and reflect relevant structure-property relationships.The passage connects this goal with Gaussian processes, Bayesian optimization, and virtual screening.Source: §3 · p. 15
  • Molecular generation depends on representation: graphs convey chemical features useful for learning properties, while 3D equilibrium geometries can bypass costly low-energy configuration sampling.The passage also notes SMILES as a useful molecular representation and calls for considering three-dimensional coordinates during decoding.Source: §3 · pp. 15–16
  • Future work should advance decoding graphs from latent codes and invertible representations that encode sequences, connectivity, and spatial arrangement.Source: §3 · p. 16
  • The outlook includes newer normalizing-flow models such as Boltzmann Generator and GraphNVP, alongside unsupervised studies of coarse-graining, synthesizability, and genetic variation.Source: §3 · p. 16
  • Generative models may optimize or learn from data and bypass human supervision during material generation; addressing their challenges is presented as essential to accelerating discovery.The chapter suggests this may also improve human understanding of nature.Source: §3 · p. 16
Loading 1907.01632v1…