PREDICTIVE RETROSYNTHESIS IN...HBZ Meeting June 2020 - SciFinder-n Predictive Retrosynthesis Created Date: 7/1/2020 1:03:45 PM ...

PREDICTIVE RETROSYNTHESIS IN

Copyright 2020 American Chemical Society. All rights reserved.

Approach, methods and future directions

Orr [email protected]


Synthesis is a Bottleneck

Complex, iterative process

Expensive and time intensive

Requires resources with deep domain knowledge

CASD – a powerful ideation tool

Main value:

• Leads to strategies and reaction steps I wouldn’t have otherwise considered

• Exposes chemistry I was not familiar with

• Allows me to be more innovative

Task Support:

• Finding synthetic routes for known and novel molecules

• Identifying alternatives to known routes

• First pass / brain storming

But also plenty of room for improvement...


A 7-months perspective

Phase I – Experimental only (June 2019)

Retrosynthesis in SciFindern


LDH inhibitor










CONTROLLING THE SEARCH


A targeted bond will be disconnected in the first step

Protected bonds are never disconnected

Control of rules


Guiding Principles

Holistic user experience embedded in the

SciFindern platform

Automated rule generation

Continuous R&D of methods and knowledge

• Diversity of solutions• Thorough coverage of known

chemistry• Tight linkage between

predictions and experimental evidence

Quality of results

• Flexibility in guiding the search• Options to filter and sort

solutions• A component of a larger

research and discovery toolkit• Ease of use


Rule Generation

Identify the essence of reactions

Molecular Perception

DRIVEN BY ADVANCED CHEMICAL PERCEPTION


RingsAromaticityFunctional groupsStereochemistrySymmetry

Reaction Perception

Rule Components


• Rules capture the essence of chemical reactions.

• The rules are described by substructures along with many properties associated with the bonds and atoms that capture the scope of the reaction

RulePrediction

Generalization as Key for Novelty


Using perception to achieve accuracy during extrapolation

Chemical perception facilitates generalizations that can lead to more diverse and more comprehensive solutions sets

EWG

EWG

EWGEWG

EWG

EWG

Rules and Supporting Examples


Populating the rules with the full content leverages the full body of knowledge creates a coherence between predictions and evidence.

Scheme Number of Examples

500,000

` 500,000

1,000

<10

common

rare

EVIDENCE FOR PREDICTED REACTIONS


FURTHER INSIGHTS


Rule Matching During Search


Amide formation from carboxylic acid and a primary amine:

A rule matches if:1. Its product substructure maps perfectly somewhere in the target molecule

2. All the atom and bond properties match as well

+

Retron

Rule Application


A transform is an inverted reaction rule:

The rule is a template for generating precursors:

++

Rule: Transform:

+

The Search


1. Is the molecule commercially available? If cheap enough, stop!

2. Is the molecule a product of known reactions?

3. Are there rules that match the molecule?

Repeat for all precursors until either:

- Reached cheap commercially available substances

- Reached the synthetic depth

$$$

Solutions Ranking


Scoring criteria:

• Complexity reduction

• Convergence

• synthetic efficiency: number of steps, yield

• atom efficiency (waste)

• starting material coverage

• Cost

• Evidence

Personalized Scoring






Stereoselectivity


Predictive Stereoselectivity


Predictive Stereoselectivity


Comprehensive properties for atoms and bonds defining scope of method

• Regioselectivity

• Synthesizability of molecules

• Smarter search algorithms

• Functional group interference

• Ranking methods

• Speed

• User experience


Other Research Projects

Main principles:

• Comprehensive, rich and accurate

• Leveraging the most current data

• Versatile and easy to use

• A facet of a complete discovery experience

• Continuously updated and improved



PREDICTIVE RETROSYNTHESIS IN...HBZ Meeting June 2020 - SciFinder-n Predictive Retrosynthesis Created Date: 7/1/2020 1:03:45 PM ...

Documents