Deep learning for protein engineering with Ray - Stanley Bishop DeepChem.io

We will discuss Ray as an active learning orchestrator for protein engineering in the drug/medicine discovery process. In particular, we will look at the deployment of systems that involve active-learning feedback between sequence-to-sequence transformers, Alphafold driven sequence-to-structure prediction, and, more broadly, how these two deep learning methods are revolutionizing the field.

Stanley Bishop is an ML-nerd contributor to the open source-project DeepChem.io, which works to democratize deep learning for science.

Transcript

1. Stanley
   Page 1
   Ray & Protein Engineering
   @ DeepChem
   

   View Slide

2. Stanley
   Page 2
   • Mathematician working as a
   machine learning scientist
   • Developer at deepchem.io, an
   open-source project to
   democratize deep learning for
   science
   ✨come check us out✨
   • Mostly made of proteins
   

   View Slide

3. Stanley
   Page 3
   • The DeepChem project works
   to democratize deep learning
   for science
   • The DeepChem project aims to
   create high quality, open
   source tools for drug
   discovery, materials
   science, quantum chemistry,
   and biology.
   • DeepChem projects are
   managed by a group of open
   source contributors.
   

   View Slide

4. Stanley
   Page 4
   • Proteins are complex
   molecules that perform
   critical functions in
   our bodies
   • Proteins are
   synthesized through the
   processes of
   transcription and
   translation
   • Protein synthesis is
   the bioinformatic
   equivalent to a
   compiler process
   

   View Slide

5. Stanley
   Page 4
   • Proteins are complex
   molecules that perform
   critical functions in
   our bodies
   • Proteins are
   synthesized through the
   processes of
   transcription and
   translation
   • Protein synthesis is
   the bioinformatic
   equivalent to a
   compiler process
   

   View Slide

6. Stanley
   Page 5
   • Proteins are made of amino
   acids
   • Amino acid compositions are
   deterministic and
   combinatorial
   • Imagine Lego bricks
   

   View Slide

7. Stanley
   Page 6
   • Protein structure is the
   result of a three-part
   process
   • Until recently, these
   structures were largely
   observed experimentally
   • AlphaFold has changed the
   game
   

   View Slide

8. View Slide

9. Stanley
   Page 8
   • Protein complexes can
   perform predictable actions
   in response to stimulus
   • Molecules activate a protein
   by ‘docking’ which changes
   its electrochemical dynamics
   • The dynamics of this docking
   are of crucial importance to
   medicine discovery and
   disease treatment
   

   View Slide

10. Stanley
    Page 9
    • In the 1950s there
    were tens of
    thousands birth
    defects caused by
    Thalidomide
    prescriptions to
    expecting mothers
    • Thalidomide-TBX5
    docking complex
    yields reactive
    oxygen species
    • With predictive AI
    technologies, these
    reactions can be
    found prior to human
    testing
    

    View Slide

11. • The space of possible ligand
    molecules can contain hundreds
    of millions to billions of
    potential compounds
    • Per ligand computations are
    relatively expensive
    • Protein ligand docking presents
    a challenging distributed
    computing problem
    • That has been traditionally
    solved with on-prem hardware
    Stanley
    Page 10
    

    View Slide

12. Stanley
    Page 11
    Docking is a DAG!
    

    View Slide

13. Stanley
    Page 12
    ● Task 1: generate 3D structures
    

    View Slide

14. Stanley
    Page 13
    ● Task 2: generate features
    

    View Slide

15. Stanley
    Page 14
    ● Task 3: compute dock complex
    

    View Slide

16. Stanley
    Page 15
    

    View Slide

17. Stanley
    Page 16
    ● Active learning has
    the potential to
    accelerate molecular
    simulation
    ● Teams at Stanford
    and Harvard are
    using Ray for this
    purpose
    ● Early results
    indicate a 10x
    improvement in
    computational
    efficiency
    

    View Slide

18. • The DeepChem is always
    looking for contributors.
    Check us out at DeepChem.io
    • Bioinformatics is likely to
    soon lead the machine
    learning charge in terms of
    the data scale of models… so
    there will be a lot to build
    • Deep gratitude to the Ray
    community for building such
    important infostructure for
    the machine learning
    revolution
    

    View Slide