Skip to content

Latest commit

 

History

History
87 lines (77 loc) · 5.84 KB

tldrs.md

File metadata and controls

87 lines (77 loc) · 5.84 KB
Raw

Highly accurate protein structure prediction with AlphaFold

  • TLDR ID: 1
  • John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, Alex Bridgland, Clemens Meyer, Simon A. A. Kohl, Andrew J. Ballard, Andrew Cowie, Bernardino Romera-Paredes, Stanislav Nikolov, Rishub Jain, Jonas Adler, Trevor Back, Stig Petersen, David Reiman, Ellen Clancy, Michal Zielinski, Martin Steinegger, Michalina Pacholska, Tamas Berghammer, Sebastian Bodenstein, David Silver, Oriol Vinyals, Andrew W. Senior, Koray Kavukcuoglu, Pushmeet Kohli & Demis Hassabis
  • Nature (2021)
  • https://www.nature.com/articles/s41586-021-03819-2
  • Implementations:

What?

  • AlfaFold2 predicts 3D coordinates of a protein from its amino acid sequence.
  • Inputs:
    • Query existing genetic databases to build the multiple sequence alignment.
    • Search for similar protein structures in the database for templates.
  • Feed this to the transformer with a twist (Evoformer).
  • Predictions are done in the iterative way.

Why?

3D structure of a protein gives us information about the protein behaviour. Doing this experimentally takes long time and a lot of money.

So What?

  • Top 1 at ICASP14.
  • Tremendous uptake in the community, AF2 is at the technology stage enabling novel research directions and bosting the old ones.
  • DM releases a 200 mil + database of protein structures

Planning chemical syntheses with deep neural networks and symbolic AI

What?

  • Predicts how to synthesize target molecules.
  • Idea:
    • "World model" pre-trained on chemical reactions predicts reverse chemical reactions.
    • Reaction feasibility model estimates whether a reaction is feasible.
    • The world model is a policy, which is then coupled with the feasibility model in an MCTS/RL-based search algorithm, which operates on a hypergraph representation.
    • Search is recursively done until molecule is decomposed in to a set of pre-defined building block molecules

Why?

Chemical Synthesis can be a bottleneck in molecular drug discovery (in particular hit2lead and lead optimization) and drug manufacture, and is usally quite costly. Predicting chemical syntheses is also important for computational workflows which propose novel molecules.

So What?

  • Demonstrated for the first time that purely ML driven methods can provide reasonable synthesis routes.
  • Paved the way for modern synthesis planning to become part of in-silico discovery pipelines.
  • An early demonstration of RL-based planning in a fully learned world model.

High-throughput property-driven generative design of functional organic molecules

What?

  • Biasing a auto-regressive SchNet (Schütt et al., 2017) model, through repeated retraining, to generate molecules that are "viable" (low energy gap, low ioinisation energy, low synthetic complexity) and satisfy many chemical property requirements simultaneously.
  • Idea:
    1. Initialise a database of known/valid drug molecules and train G-SchNet on it.
    2. Generate candidated molecules (~340K in the paper), filter each of them using pretrained models to predict different properties (synthetic complexity, energy gap, etc.). These pretrained models were trained on other datasets to predict specific properties. Think of the ensemble as a mixture of experts making a final decision whether or not a generated molecule is valid.
    3. Out of all valid molecules, merge the new collection of generated molecules with the original database. This is to bias the network into increasingly produce molecules with ideal properties.
    4. Retrain the same G-SchNet on this new, updated dataset. Repeat process (~10 loops from the paper, each taking ~2 days).

In doing so, the generated molecules will look increasingly similar to real-world molecules with ideal properties.

Why?

High-throughput drug molecule screening is often tedious given the time-consuming nature of DFT calculations. By generating high-quality candidate molecules that satisfy many desired properties, the search space is significantly reduced, meaning lesser, more efficient wet-lab testing.

So What?

  • Demonstrates that explicit graph topology and structural descriptors (SMILES strings) are not needed to generate high-fidelity, high-potential molecules
  • Screens many thousands of drug molecules in a matter of days in contrast to traditional screening approaches taking >20,000 days on average

Note: However, it seems like a computationally expensive process with many loops of the retraining process

Drug Synergistic Combinations Predictions via Large-Scale Pre-Training and Graph Structure Learning

What?

  • Predict whether two drugs are synergistic.
  • Idea:
    • Authors used pre-trained models to encode proteins and molecules
    • The models were fine-tuned in a self-supervised manner on biological knowledge graph (PrimeKG)
    • Then, the authors used active learning to fine-tune for the second time and evaluate their models on DrugComb dataset

Why?

Drug combinations may be more effective than single drugs. However, identifying novel drug combinations through wet-lab experiments is resource intensive due to the vast combinatorial search space

So What?

  • SOTA on several datasets (DrugComb, AstraZeneca)
  • Multimodal data helps on homogeneous benchmarks
  • Using pretrained models is much more effective than train from scratch.