Workshop Chemical Robotics ChemAI 231116.pptx

From Words to Wonders:
Language Models for Life
Sciences
Room L1.02
Robots Unleashed: The Rise of AI-
Driven Chemical Discovery
Room L1.01
16 November 2023

The Rise of AI-driven Chemical Discovery
- Language Models, Robotics, and Digitizing
Operations
ChemAI 2023
Amsterdam
16 November 2023
Dr. Amol Thakkar
�tha@zurich.ibm.com
Research Scientist
AI for Scientific Discovery
IBM Research

The research
lab is the
central
element in
scientific
discovery
2
IBM Research| ©2023 IBM Corporation

Up to 70% of experimentation is not
reproducible because of flawed
experimental data or metadata1
Only one-third to one-half of original
findings are also observed in replication
studies2
Studies can last years instead of months
due to small differences in protocols3
In 2020, companies spent an average of
17% of revenue on digital initiatives and
4% of revenue on R&D4
IBM Research| ©2023 IBM Corporation 3
1 Accenture, Digital Transformation in the Lab (2020)
2 Aarts, A. A. et al. Science 349, 943 (2015)
3 Hines, W. C. et al. Cell Rep. 6, 779–781 (2014)
4 Gartner, Digital Technologies in R&D (2019), S&P 500 (2020)

AI breakthroughs for language are changing scientific
discovery
4
AI
Data AI Models
Question Answering
Improved
Language
Tasks
Sentences
Words
Letters
Language
AI
Data AI Models
Improved
Chemistry
Tasks
Properties/Reactions
Molecules
Atoms
Chemistry
Generative modeling and transformers are achieving new breakthroughs in chemistry
Natural Language Processing (NLP) Accelerated Chemistry
Transformers Transformers

Generative Toolkit for Scientific Discovery (GT4SD)
Open-source library to accelerate hypothesis generation in scientific discovery
A winner of the
2022 IEEE Open Software
Services Award
GT4SD makes generative AI algorithms and
models easier to use in scientific discovery
https://github.com/GT4SD/gt4sd-core Applications include hypothesis generation for
inverse design and discovery of materials and
therapeutics like antivirals and antimicrobials
Example molecules generated using GT4SD
Test
Hypothesize
Study
1. Train generative models
2. Create inference pipelines
3. Run inference pipelines
W Manica et al., npj Comput. Mater. 9, 69 (2023)
AI for Scientific Discovery | IBM Research | ©2023 IBM Corporation 6

Chemistry as a language
C C 1 ( C C C C ( N 1 O C ( O C 1
= C C = C ( [N+] ( = O ) [O-] ) C
= C 1 ) = O ) ( C ) C ) C
C C 1 ( C C C C ( N 1 O ) ( C ) C ) C . O ( C ( = O ) Cl )
C 1 = C C = C ( [N+] ( = O ) [O-] ) C = C 1
Products Reactants + Reagents
English Spanish
Translation
retrosynthesis
synthesis
“Sentence of atoms”
Textual representation (SMILES)
CC1(CCCC(N1OC(OC1=CC=C([N+](=O)[O-])C=C1)=O)(C)C)C CC1(CCCC(N1O)(C)C)C O(C(=O)Cl)C1=CC=C([N+](=O)[O-])C=C1

Automating lab synthesis and experimentation
with help of AI
7
W Vaucher et al., Nat. Comm. 12 (2022) 2573
O = C ( N C c 1 c c c c ( Cl ) c 1 ) c 1 c c c ( C Br ) c c 1
1. AI-based chemical reaction prediction
synthesis
+
retrosynthesis
“translation”
transformer
N C c 1 c c c c ( Cl ) c 1 . O = C ( Cl ) c 1 c c c ( C Br ) c c 1
W Schwaller et al., Chem. Sci. 9 (2018) 6091
W Schwaller et al., ACSCent. Sci. 5 (2019) 1572
W Schwaller et al., Chem. Sci. 11 (2020) 3316
2. Chemical procedures from text (Paragraph2Actions)
translation
The TFA was removed in vacuo and a
saturated solution of NaHCO3 was added.
Concentrate(),
Add(name=‘saturated solution of NaHCO3’)
W Vaucher et al., Nat. Comm. 11 (2020) 3601
C(=NC1CCCCC1)=NC1CCCCC1.ClCCl.CC1(C)CC(=O)Nc2cc(C(=O)O)ccc21.Nc1ccccc1>>CC1(C)CC(=O)Nc2cc(C(=O)Nc3ccccc3)ccc21
3. Chemical procedures from reactions (Smiles2Actions)
1. ADD $1$
2. ADD $4$
3. ADD $2$
4. ADD $3$
5. STIR for @3@ at #4#
6. FILTER keep precipitate
7. RECRYSTALLIZE from ethanol
8. YIELD $-1$
30k+
global users
via cloud
40+ million
reaction
predictions
Automated synthesis via cloud + robotic lab
https://rxn.res.ibm.com/

Making AI for chemistry available to everyone
http://rxn.app.accelerate.science/

Multi-modal foundation models will
accelerate fundamental research tasks
W Christofidellis et al., ICML 2023
0
An example from chemistry…

Up to 70% of experimentation is not
reproducible because of flawed
experimental data or metadata1
Only one-third to one-half of original
findings are also observed in replication
studies2
Studies can last years instead of months
due to small differences in protocols3
In 2020, companies spent an average of
17% of revenue on digital initiatives and
4% of revenue on R&D4
1 Accenture, Digital Transformation in the Lab (2020)
2 Aarts, A. A. et al. Science 349, 943 (2015)
3 Hines, W. C. et al. Cell Rep. 6, 779–781 (2014)
4 Gartner, Digital Technologies in R&D (2019), S&P 500 (2020)

Vendor lock in for devices, poor
interoperability, and no APIs provided
Documentation burden on laboratory
scientists
Data and meta data not captured,
reporting bias
Need to improve data quality!

Building on the Experience of Domain Experts
Domain Specific Models
Expert Systems
Machine Learning
Deep Learning
User Driven Ecosystem
Multiple models for specific
tasks
Intelligent Ecosystem
Orchestration of domain
specific models based on
context
Domain Experts
Generate Data Requires data
?

Creating the Lab that Learns Key innovations Benefits for the lab
• AI foundation models for
automatic
documentation of
manual procedures and
validation of outcomes
• Hybrid and multi-cloud
computing to
automatically integrate
all data and metadata of
any experiment
• Capture all details
needed to fully capture
and describe an
experiment
• Minimize the time wasted
using different tools to
organize data
• Reproduce any version of
an experiment at any
point in time
• Discover patterns by
continuously learning
over all experimental
data
Chip AZ33 was
developed in 1:3
MIPK:IPA for 10
seconds
E-beam
lithography
Details
Manual puddle
development
Details
Video +Gaze
Audio
Text
The AI-enabled lab of the future for a new
era of reproducible and collaborative
experimentation
Multi-modal
data
Foundation model
generates workflows
010011101101
110101011100
101110110101
100011110011
Bytes

Consider a Laboratory Environment
Chemistry
laboratory
Lab user
Action:
Measurement
Problem:
- Recording volume
IBM Research | ©2023 IBM Corporation 'image: Flaticon.com'. This graphic has been designed using images from Flaticon.com 16

Multi-Modal Models for Automated Documentation
Chemistry
laboratory
Lab user
Meas
Lab goggles with video

Meas
Chemistry
laboratory
Lab user
Containerised data watcher
Laboratory Device
Device 1
Automatic step inference
- Predicts the step taken
- Automatic extraction of
- Volume
- Measurement data

Stir
Chemistry
laboratory
Lab user
Containerised data watcher
Laboratory Device
Device 1
Automatic step inference
- Its easy to forget steps
trivial for domain experts
- Automatic
documentation of which
step and in which order

Architecture Overview
An
Chemistry
laboratory
Lab user
Device 1 Device 2
Collect Videos
Collect Device
Measurements
Data Storage
AI Engine
Consolidate Workflow
Actions
Notes
Digit Recognition

Interoperable Across Multiple Devices
Device 1 Device 2
- Vendor independent data streaming
- Interoperable across multiple devices and
networks
- Containerised solution for automatic data
capture

Interoperable Cloud Environments with Multi-Cloud
- Vendor independent cloud solution
- On premise, public cloud, or a combination
- AI models, and data can live in different cloud
environments
Actions
Notes
Digit Recognition

Summary
Experiment
Question
Hypothesize
Study
Analyze
Report
Lab that Learns
- Vendor independent solution for device
integration
- Interoperable across multiple devices and
networks
- Automatically documents steps for
experiments and workflows
- Extract crucial data and meta-data of
relevance to the experiment
- Reproducible workflows

What can we
discover
together?
Acknowledgments
Adriano Martinelli
Alain Vaucher
Alek Sobczyk
Alessandra Toniato
Alice Driessen
Amol Thakkar
Andrea Giovannini
Antonio Foncubierta
Anuv Chakraborty
Artem Leonov
Carlo Baldassari
Dimitrios Christofidellis
Federico Zipoli
Gianmarco Gabrieli
Jannis Born
Joris Cadow-Gossweiler
Mara Graziani
Marianna Rapsomaniki
Marvin Alberts
Matteo Manica
Miruna Cretu
Nicolas Deutschmann
Nikita Janakarajan
Oliver Schilter
Pushpak Pati
Patrick Ruch
Teodoro Laino

Wilhelm Huck
Radboud University

Introducing the Nationaal
Groeifonds ‘Big Chemistry’
consortium

Partners:
TU/e, RUG, Radboud, AMOLF
Fontys Hogeschool Eindhoven
Max Planck Gesellschaft

Complex society,
complex problems

We know we need to change….
….. But change is too slow!

Changing the paradigm of chemical research
7
Alun Aspuru-Guzik Science (2018) 361, 360-365 DOI: 10.1126/science.aat2663

Evolution is the ‘problem solver’ in biology
8

We need an evolution machine
Genotypes are defined as the collection of
all experimental parameters of a system
(i.e. molecular composition, pH,
temperature, etc. etc.)
Phenotypes are defined as the collection of all
experimental properties of a system (i.e.
fluorescence, turbidity, spatiotemporal
patterns shape, etc. etc.)

Chemistry is more than synthesis
Focus on molecular properties
What?
a) Solubility of molecules in water and organic solvents
b) Predicting CMC, surface tension
c) Predicting reactivity
d) Vapour pressures
Possible applications
a) Aiding formulation of stable emulsions
b) Creating a desired smell of mixture of compounds
c) Discovering catalytic activity
d) Automated synthesis using solubility, reactivity, kinetics prediction

The Big Chemistry ecosystem
RobotLab
Central Facility
Max Planck Research Campus
Industry
transforming formulation from an art to a
science-based technology
Start-ups
Specialized CROs
online formulation
Fundamental research
Tue, RUG, RU, AMOLF, Fontys

12
Example: LLM for solubility prediction
MMB was trained on the ZINC database, approx. 1,5 billion molecules

Database: https://zenodo.org/records/5970538
Vermeire et al. J. Am. Chem. Soc. 2022, 144, 24, 10785–10797
13
megaMOLBart trained on AqueousSolu da
compounds)
Promising result:
MMB is as good as high-level theoretical calculations in predicting
solubility….
(trained small regression head, 600k parameters)

megaMOLBart does not understand chem

Broadening: predicting logCMC values
Data:
Manually curated dataset containing 1316 compounds
Type of surfactants:
i. Anionic --- 225 compounds
ii. Anionic-cationic salt --- 13 compounds
iii. Cationic --- 693 compounds
iv.Nonionic --- 366 compounds
v. Zwitterionic --- 19 compounds

Next steps
Explore possibilities for multi-property prediction
solubility + pCMC + surface tension
solubility in multiple organic solvents
Expand experimental datasets
Ensure each additional datapoint yields maximum information
Develop high throughput analytical methods
Beyond pure compounds
Predict properties of mixtures of molecules

acknowledgements
Co-PIs: Bert Meijer, Ghislaine Vantomme, Ben Feringa, Nathalie Katsonis Board Big C
Marcel Wubbolts Radboud Team:
Tal Kachman, Stefan Hödl, Will Robinson, Aigars Piruska, Luc Hermans Peter Korevaar, Jana Roit
Collaborators RUG, Tue, AMOLF, Fontys

Workshop Chemical Robotics ChemAI 231116.pptx

Recommended

Recommended

More Related Content

Similar to Workshop Chemical Robotics ChemAI 231116.pptx

Similar to Workshop Chemical Robotics ChemAI 231116.pptx (20)

Recently uploaded

Recently uploaded (20)

Workshop Chemical Robotics ChemAI 231116.pptx