Artificial intelligence for synthetic organic and analytical chemistry

Summary

• Focus: Small organic molecules (useful and versatile)

  • Challenge: Chemical space is vast

  • Molecular discovery: complex multi-objective optimization

    • Typically driven by human intuition

    • Tasks: 

      • Predicting chemical properties, including reactivity

      • Ideating new molecular structures

      • Balancing objectives

• Research threads:

  • AI for synthetic organic chemistry, medicinal chemistry, analytical chemistry

  • Foundational capabilities: chemistry-tailored neural nets, data sharing, autonomous chemistry labs

• History of key chemical tasks

  • Computer-aided retrosynthesis: Compute programs that explore recipes from an expert-encoded rules/heuristics/constraints

  • Explain reactivity trends: data-driven analysis of relation between physical conditions and experimental outcomes

  • Predicting spectra (how molecules look to sensors): rule-based analyses

• Synthesis planning: how we access (new) molecules

  • Input: product to synthesize

  • Output: reactants, intermediaries, conditions

  • Typical approach: start with product and try to reverse it until we get to chemicals that we can purchase

  • Use libraries of valid chemical transformation rules

    • Produced by chemical vendors

    • Expert encoded rules in software (https://www.synthiaonline.com/)

    • Generative models that hypothesize possible transformations

      • Mine databases of historical reactions

      • Critical to create a canonical representation of molecules and reactions

        • Strings (e.g. SMILES)

        • Structural “fingerprints”

        • Descriptions of constituent molecules

        • Graphs & graph edits (requires atom mapping)

        • Condensed graph of reaction (requires atom mapping)

  • Synthesis constrains the space of chemicals, transformations and all influences we can access easily

  • Every transformation requires environmental conditions (solvents, additives, concentrations, temperature, reaction time, etc.)

    • Different approaches focus on different levels of detail

  • Approach:

    • Learning transformations rules for reactions

      • From databases of known reactions

      • 
        

    • Representation: graphs of atoms + covalent bonds

    • Graph neural networks: learn about the behavior of each atom based on its connective structure

      • Limitations

        • No 3D structure (not that important for small molecules)

        • Ignore chirality of molecules

        • Some covalent bond details are not represented

        • Ignore interactions beyond covalent bonds

        • Ignore Atropisomerism

    • Database Learning process

      • Find core transformation from database

      • Add related neighbor reactions

      • Use set of reactions to drive a retrosynthesis search

    • Neural net process

      • Train model to convert from products to reactants

        • SMILES->SMILES

        • Graph->Graph

        • Graph->SMILES

      • Apply repeatedly within a retrosynthesis search loop

        • Many algorithms to search very large space

        • Monte Carlo tree search, best-first, etc.

        • RL is usable but challenging because the space of moves is dynamic

        • Search is vulnerable to hallucination where predicted transformations are wrong and lead it down wrong paths

        • Simulations can be used to check these rules but they are not yet ready to be used reliably (hard to set up, computationally expensive, inaccurate)

    • Reaction condition recommendation as fill-in-the-blank

      • Embedding model: learns vector embedding of reagents based on their function. Groups them in ways that mimic structural relationships.

• ASKCOS: https://askcos.mit.edu/

  • Suite of synthesis planning modules

  • Chemoinformatics & ML

  • Tasks: Retrosynthesis, condition recommendation, reaction product prediction, reaction classification, atom mapping, selectivity prediction, solvation prediction

  • 35k users, 15 companies

• Challenges:

  • Complexity of synthetic targets is changing: more complex molecules and synthesis pathways are needed for modern use-cases

  • Data-driven search programs generate many pathway ideas but what do we do experimentally?

    • Need to score more promising options

    • E.g. feasibility, impurity, greenness, yield, flow compatibility, scalability, cost

    • The source of data from which reactions are sourced affects ability to evaluate reactivity

      • Diverse in substrates/reactions but not conditions: papers, patents

      • Diverse in conditions but not in substrates/reactions: high-throughput experimentation

      • Need techniques that cover both

        • Open Reaction Database: encourages data sharing across teams

        • https://open-reaction-database.org/

          • Database

          • Data structure

          • Different organization approach for community

• Summary:

  • AI/ML has broad relevance for chemistry

  • Old goals, ongoing new approaches

  • ASKCOS suite of tools

  • Overall, great opportunity for supervised learning