How to Hallucinate Functional Proteins

By Zak Costello, Héctor García Martín. 2019

34 citations recorded by the source corpus.

Here we present a novel approach to protein design and phenotypic inference using a generative model for protein sequences. BioSeqVAE, a variational autoencoder variant, can hallucinate syntactically valid protein sequences that are likely to fold and function. BioSeqVAE is trained on the entire known protein sequence space and learns to generate valid examples of protein sequences in an unsupervised manner. The model is validated by showing that its latent feature space is useful and that it accurately reconstructs sequences. Its usefulness is demonstrated with a selection of relevant downstream design tasks. This work is intended to serve as a computational first step towards a general purpose structure free protein design tool.

Subfields: Biophysics, Molecular Biology

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Biophysics

1. Kadina E. Johnston
2. Richard Socher
3. Debora Marks
4. Frances H. Arnold
5. Kevin Kaichuang Yang
6. Zachary Wu

In Molecular Biology

1. Kevin Kaichuang Yang
2. Debora Marks
3. Christopher Sander
4. Martin Weigt
5. Andrea Pagnani
6. Zachary Wu
7. Adam J. Riesselman
8. Ali Madani

Citations in the corpus

• ProGen: Language Modeling for Protein Generation (2020, methodology): “Meanwhile sequence-only generative modeling have been attempted recently through residual causal dilated convolutional neural networks (Riesselman et al., 2019) and variational autoencoders (Costello & Martin, 2019).”
• Accelerating Protein Design Using Autoregressive Generative Models (2019, background): “The addition of latent variables will allow controlled generation of desirable sequences (Bowman et al., 2015; Costello & Martin, 2019).”
• Efficient generative modeling of protein sequences using simple autoregressive models (2021, background): “[14], which reported a VAE implementation trained on non-aligned sequences from UniProt, with length 10 < L < 1000.”
• Machine-learning-guided directed evolution for protein engineering (2018, background): “In addition, the protein variants generated by a variational auto-encoder or other generative model can be highly sequence divergent from known sequences but potentially still functional 91 .”
• Protein sequence design with deep generative models (2021, methodology): “For example, in BioSeqVAE, the latent representation was learned from 200,000 sequences between 100 and 1000 amino acids in length obtained from SwissProt [18].”