chapter ten

10 Generative Models for De Novo Design

 

This chapter covers

  • Challenges of navigating chemical space and limitations of traditional methods.
  • Learning a compressed molecular latent space with autoencoders.
  • Autoencoder architecture: tokenization, embeddings, and encoder-decoder structures.
  • Using Variational Autoencoders (VAEs) to generate novel molecules.
  • Advanced techniques (GRUs, cyclical annealing) for powerful generative models.

The journey of discovering a new drug is often likened to finding a needle in a colossal haystack. It's a process fraught with challenges, immense costs, and high attrition rates. At its heart, drug discovery is a molecular design problem: identifying or creating a molecule with the precise set of properties needed to safely and effectively treat a disease. This chapter delves into computational techniques that aim to make this quest more efficient and targeted, designing novel molecules with desired characteristics.

Our specific goal is to build a generative model that learns a compressed "latent space" – a simplified, continuous map of the chemical universe. By learning the underlying rules of chemistry from a large dataset of molecules, this model allows us to navigate its map to design and optimize entirely new molecules, biasing our search toward its most promising regions rather than sifting through the entire haystack at random. The following sections document the step-by-step construction of this powerful molecular design engine.

10.1 The Quest for Designer Molecules

10.1.1 The Challenge of Chemical Space

10.1.2 Generative Models: A New Paradigm for Molecular Design

10.1.3 Reinforcement Learning for Targeted Generation

10.2 Building the World: Generative Models for Molecules

10.2.1 Essential Properties of a Good Molecular Latent Space

10.2.2 Learning to Compress and Recreate: The Autoencoder

10.2.3 The Autoencoder Architecture

10.2.4 Experiment on the MOSES Benchmark

10.3 Creating a Continuous Chemical Universe: Variational Autoencoders

10.4 Understanding Sequential Molecular Structure: Recurrent Neural Networks

10.4.1 How RNNs Process Sequences

10.4.2 Resolving Vanishing Gradients with Gated Recurrent Units

10.4.3 Seq2Seq Architecture: Encoding and Decoding Molecules

10.4.4 Revisiting Tokenization: Byte-Pair Encoding for Molecules

10.4.5 Putting It All Together: VAE-CYC

10.5 Summary