What compound to synthesize next? - Wie Machine Learning und KI das Wirkstoffdesign beeinflussen - PharmaForum
←
→
Page content transcription
If your browser does not render page correctly, please read the page content below
What compound to synthesize next? Wie Machine Learning und KI das Wirkstoffdesign beeinflussen Daniel Kuhn Computational Chemistry & Biologics Merck Healthcare KGaA 2021-04-14
Efficiency in drug discovery is declining over years
Scannell et al., Nat Rev Drug Disc 2012
Daniel Kuhn | 2021-04-14 | PharmaForum 2021
https://www.nature.com/articles/nrd3681
Drug discovery development is costly and timely
Duration: 12-14 years
Graphic created by John Chodera
https://www.choderalab.org/s/2021-03-31OxfordCentreforMedicinesDiscovery-compressed.pdf
Daniel Kuhn | 2021-04-14 | PharmaForum 2021
Drug development
A multi-parameter optimization problem
Efficacy
PGP permeability
hERG
Microsomal stability
Solubility
HepG2 cytotoxicity CACO2 permeability
Protein kinase selectivity
TEPMETKO (Tepotinib) bound to c-Met
Daniel Kuhn | 2021-04-14 | PharmaForum 2021
Which compounds to make next?
Challenge: Chemical space is huge – which compounds to make next?
Chemical space is huge
ML/AI
Foote et al., J. Med. Chem. 2013, 56, 2125–2138
Daniel Kuhn | 2021-04-14 | PharmaForum 2021
AI creatively inventing novel stuff
Neural network DALL·E is a 12-billion parameter version
of GPT-3 language model trained to generate images from
text descriptions
„Draw a picture of an
armchair in the shape of an DALL·E NN
avocado “
http://openai.com - DALL·E: Creating images from text
Daniel Kuhn | 2021-04-14 | PharmaForum 2021
„Design a cellular potent c-Met inhibitor with good microsomal
stability and high CACO2 permeability“
AI-NN
Daniel Kuhn | 2021-04-14 | PharmaForum 2021
Predictive models in MOCCA: Endpoints & performance
Very
PGP permeability good
Good
hERG
Microsomal stability
Solubility
HepG2 cytotoxicity CACO2 permeability
Protein kinase selectivity
Daniel Kuhn | 2021-04-14 | PharmaForum 2021
Generally global models are preferrable due to greater in-house modelling experience and larger AD, but we are happy to support projects with local models if needed.
e.g.
Combination is key for impact in compound optimization
MOCCA: A MASSIV: Enumeration of synthetically
2 Application of ML/DNN predictive models a ffo
ld
1 accessible chemical space
Sc
Generative design
A
ld
a ffo
MASSIV & Generative Design
Sc
Virtual Screening as 1st filter
MOCCA FEP: Binding constant
FEP calculations
3 prediction
12 Daniel Kuhn | 2021-04-14 | PharmaForum 2021Discovery of new chemical starting points with FEP+ML
Use case 3: From fragment to hit
Enamine RealSpace
903 ideas
SPR KDss = 300 µM Top 1 in FEP
LE = 0.25 IC50 = 1.2 µM
3D ROCS overlay ITC KD = 1 µM
LE = 0.41
750 ideas
Docking + MMGBSA
400 ideas
IC50 = 24 µM
Synthesis at Enamine ML model: CLint
- 4 weeks 250 ideas
- < 100 EUR per compound
5 out of 8 molecules
have IC50 < 100 µM IC50 = 47 µM
FEP
8 ideas
Daniel Kuhn | 2021-04-14 | PharmaForum 2021
Christina SchindlerThe next ten years
Medicinal Chemistry
Predictive modelling for
& AIthe Futuredrug design
driving
Merck celebrated it’s 352th birthday on August 26th 2020
X
Ten
14Acknowledgements
IT-Consultants
Christina Schindler Johannes Schimunek Jan Fiedler
Lukas Friedrich Kristina Preuer Christian Röder
Friedrich Rippmann Günter Klambauer Samo Turk
Vanita Sood Sepp Hochreiter Andrew Dalke
Mireille Krier
Tim Knehans
Cornelius Kohl
Gianna Pohl
Michael Krug
Quentin Perron
Jakub Gunera Yann Gaston-Mathe Franca Klingler
Christian Lemmen
Theresa Johnson
Hans-Peter Buchstaller Paul Czodrowski
Marcel Baltruschat
Daniel Kuhn | 2021-04-14 | PharmaForum 2021You can also read
