A brief overview of how our AI can help automate the materials discovery process, covering a wide range of problems, from drug design to crystal plasticity.

1. via Learning to represent, predict,
generate and explain
A/Prof Truyen Tran
Head of AI, Health and Science
AI for automated
materials discovery

2. 26/05/2023 2
Solves
Inspires
AI/ML
Memory
Learning
Reasoning
Computer
vision
Human-AI
Teaming
Optimisation
Industry
Health
Software
Drug
discovery
Materials
science Manufact
uring
Business
processes
Energy
What we do @A2I2

3. 26/05/2023 3
Agrawal, A., & Choudhary, A. (2016). Perspective: Materials informatics and big data: Realization of the “fourth paradigm” of
science in materials science. Apl Materials, 4(5), 053208.
The 5th paradigm
(2020-present)
• Advanced deep learning
• Massive data simulation
• Powerful Foundation
Models
https://www.microsoft.com/en-us/research/blog/ai4science-to-empower-the-fifth-paradigm-of-scientific-discovery/

4. Challenges
Materials science: Materials discovery is
very slow and extremely costly.
Automated chemist: Chemical interaction
and reaction prediction is key for advancing
chemistry, but extremely challenging.
26/05/2023 4

5. Materials discovery as smart
search over in exponential
space
5
#REF: Gómez-Bombarelli, Rafael, et al. "Automatic chemical design using a data-driven
continuous representation of molecules." ACS Central Science (2016).
Photo credit: wustl.edu
Molecular search space: 1023 to 1060
| Knowledge-driven
| AI-driven

6. Space of innovation
• Molecular space exploration
• Small, medium, large, supra
• Molecular interaction
• Network, docking
• Chemical reaction, retrosynthesis
• Catalyst, yield, free-energy
• Crystal space exploration
• Alloy space exploration
• Microstructures
• Knowledge extraction, coding, expression,
manipulation
27/05/2023 6
• Representation
• Graphs, geometry, periodicity, token
• Materials manifold
• Learning, attention and memory
• Self-supervised, supervised, reinforcement
• Transfer, zero-shot, few-shot, adaptation learning
• Learning to reason
• Reasoning
• Optimisation
• Extrapolation, generation
• Abductive, inductive, deductive reasoning
Materials AI/ML
Image: Shutterstock

7. AI/ML Topics
26/05/2023 7
REPRESENTATION PREDICTION OPTIMIZATION &
GENERALIZATION
EXPLANATION

8. Molecule → fingerprints
26/05/2023 8
#REF: Duvenaud, David K., et al.
"Convolutional networks on graphs for
learning molecular fingerprints." Advances
in neural information processing systems.
2015.
• Graph → vector. Mostly discrete. Substructures
coded.
• Vectors are easy to manipulate. Not easy to
reconstruct the graphs from fingerprints.
Kadurin, Artur, et al. "The cornucopia of meaningful leads: Applying deep adversarial
autoencoders for new molecule development in oncology." Oncotarget 8.7 (2017): 10883.

9. Source: wikipedia.org
Molecule → string
• SMILES = Simplified Molecular-Input Line-
Entry System
• Ready for encoding/decoding with
sequential models (seq2seq, RL,
Transformer).
• BUT …
• String → graphs is not unique!
• Lots of string are invalid
• Precise 3D information is lost
• Short range in graph may become long range in
string
26/05/2023 9
#REF: Gómez-Bombarelli, Rafael, et al. "Automatic chemical design using a
data-driven continuous representation of molecules." arXiv preprint
arXiv:1610.02415 (2016).

10. 26/05/2023 10
Molecule → graphs
• No regular, fixed-size structures
• Graphs are permutation invariant:
• #permutations are exponential function of #nodes
• The probability of a generated graph G need to be
marginalized over all possible permutations
#REF: Pham, T., Tran, T., & Venkatesh, S. (2018).
Relational dynamic memory networks. arXiv
preprint arXiv:1808.04247.
Input
process
Memory
process
Output
process
Controller
process
Message
passing
• Multiple objectives:
• Diversity of generated graphs
• Smoothness of latent space
• Agreement with or optimization
of multiple “drug-like” objectives

11. Representing proteins
• 1D sequence (vocab of size 20) –
hundreds to thousands in length
• 2D contact map – requires
prediction
• 3D structure – requires folding
information, either observed or
predicted. Now available thanks
to AlphaFold 2.
• NLP-inspired embedding
(word2vec, doc2vec, glove,
seq2vec, ELMo, BERT, GPT).
26/05/2023 11
#REF: Yang, K. K., Wu, Z., Bedbrook, C. N., & Arnold, F.
H. (2018). Learned protein embeddings for machine
learning. Bioinformatics, 34(15), 2642-2648.

12. 26/05/2023 12
Crystal structure
• Definition:
• Crystal structure is the repeating arrangement in the 3D space of atoms
throughout the crystal.
• Crystal structure is presented by the arrangement of atoms within the unit cell.
• The atom interacts with atoms within unit cell and adjacent unit cells.
Crystal structure Ac₂AgIr Unit cell of crystal structure Ac₂AgIr
Slide credit: Tri Nguyen

13. 26/05/2023 13
Crystal structure representation
• Crystal structure input:
• Atom type
• Atom coordinates
• Periodic lattice
• Multi-graph representation to model the periodic interaction
Slide credit: Tri Nguyen

14. Representing microstructures of crystal
mixture
26/05/2023 14
Generate prior 𝛽 grains
Add transformation phases
Generate dual phase models Feature information
Phase %
Orientation
Grain size
Distance to
triple point
• Input information for each
microstructure saved per voxel
• Saved data considers local
environment
Volume domain: 106
voxels
Slide credit: Sterjovski and Agius

15. 26/05/2023 15
Topics
REPRESENTATION PREDICTION OPTIMIZATION &
GENERALIZATION
EXPLANATION

16. Molecular properties prediction
• Traditional techniques:
• Graph kernels (ML)
• Molecular fingerprints
(Chemistry)
• Modern techniques
• Molecule as graph: atoms as
nodes, chemical bonds as
edges
26/05/2023 16
#REF: Penmatsa, Aravind, Kevin H. Wang, and Eric Gouaux. "X-ray
structure of dopamine transporter elucidates antidepressant
mechanism." Nature 503.7474 (2013): 85-90.

17. A graph processing machine for molecular
property prediction
26/05/2023 17
#REF: Pham, T., Tran, T., & Venkatesh, S. (2018).
Relational dynamic memory networks. arXiv
preprint arXiv:1808.04247.
Input
process
Memory
process
Output
process
Controller
process
Message
passing
Unrolling
Controller
Memory
Graph
Query Output
Read Write

18. Multi-target prediction
26/05/2023 18
Possible
targets
Molecular
graph
#REF: Do, Kien, et al. "Attentional Multilabel
Learning over Graphs-A message passing
approach." Machine Learning, 2019.

19. Predict multiple properties
26/05/2023 19
#REF: Do, Kien, et al. "Attentional Multilabel Learning over Graphs-A message passing
approach." Machine Learning, 2019.

20. Chemical-chemical interaction via
Relational Dynamic Memory Networks
26/05/2023 20
𝑴1 … 𝑴𝐶
𝒓𝑡
1
…
𝒓𝑡
𝐾
𝒓𝑡
∗
Controller
Write
𝒉𝑡
Memo
ry
Graph
Query Output
Read
heads
#REF: Pham, Trang, Truyen Tran, and Svetha Venkatesh. "Relational
dynamic memory networks." arXiv preprint arXiv:1808.04247(2018).

21. Drug and protein binding
26/05/2023 21
Drug molecule
- Binds to protein
binding site
- Changes its target
activity
- Binding strength is
the binding affinity
Protein
- May change its
conformation due to
interaction with drug
molecule
- Its function is altered due
to the present of drug
molecule at its binding site
Image credit: Lancet
Slide credit: Tri Nguyen

22. 26/05/2023 22
GEFA: Drug-protein binding as graph-in-
graph interaction
Protein graph
Drug graph
A
K
L
A
T
A
Drug
Graph-in-Graph
interaction
Nguyen, T. M., Nguyen, T., Le, T. M., & Tran, T. (2021). “GEFA: Early Fusion Approach in Drug-Target
Affinity Prediction”. IEEE/ACM Transactions on Computational Biology and Bioinformatics
Slide credit: Tri Nguyen

23. 26/05/2023 23
GEFA (cont.)
Nguyen, T. M., Nguyen, T., Le, T. M., & Tran, T. (2021). “GEFA: Early Fusion Approach in Drug-Target Affinity
Prediction”. IEEE/ACM Transactions on Computational Biology and Bioinformatics
Slide credit: Tri Nguyen

24. Predicting stress-strain curve from
crystal mixture
• Transformer to leverage
long-range
dependencies between
voxels
• Input: Feature vectors
per voxel.
• Output: Strain curve per
voxel.
26/05/2023 24

25. 26/05/2023 25
Topics
REPRESENTATION PREDICTION OPTIMIZATION &
GENERALIZATION
EXPLANATION

26. Molecular generation
• The molecular space is estimated to
be 1e+23 to 1e+60
• Only 1e+8 substances synthesized thus
far.
• It is impossible to model this space
fully.
• The current technologies are not
mature for graph generations.
• But approximate techniques do
exist.
26/05/2023 26
Source: pharmafactz.com

27. Combinatorial chemistry
• Generate variations on a template
• Returns a list of molecules from this template that
• Bind to the pocket with good pharmacodynamics?
• Have good pharmacokinetics?
• Are synthetically accessible?
26/05/2023 27
#REF: Talk by Chloé-Agathe Azencott titled “Machine learning for therapeutic
research”, 12/10/2017

28. 26/05/2023 28
Retrosynthesis
prediction
• Once a molecular structure is
designed, how do we synthesize it?
• Retrosynthesis planning/prediction
• Identify a set of reactants to synthesize a
target molecule
• This is reverse of chemical reaction
prediction
Picture source: Tim Soderberg, “Retrosynthetic analysis and metabolic pathway prediction”, Organic Chemistry With a Biological Emphasis,
2016. URL: https://chem.libretexts.org/Courses/Oregon_Institute_of_Technology/OIT%3A_CHE_333_-
_Organic_Chemistry_III_(Lund)/2%3A_Retrosynthetic_analysis_and_metabolic_pathway_prediction

29. GTPN: Synthesis via reaction
prediction as neural graph morphism
• Input: A set of graphs = a
single big graph with
disconnected components
• Output: A new set of
graphs. Same nodes,
different edges.
• Model: Graph morphism
• Method: Graph
transformation policy
network (GTPN)
26/05/2023 29
Kien Do, Truyen Tran, and Svetha Venkatesh. "Graph Transformation Policy Network for Chemical Reaction
Prediction." KDD’19.

30. 26/05/2023 30
Alloy design generation
• Scientific innovations are expensive
• One search per specific target
• Availability of growing data
Nguyen, P., Tran, T., Gupta, S., Rana, S., Barnett, M. and Venkatesh, S., 2019, May. Incomplete conditional density estimation for fast materials discovery.
In Proceedings of the 2019 SIAM International Conference on Data Mining (pp. 549-557). Society for Industrial and Applied Mathematics.

31. 26/05/2023 31
Inverse design
• Leverage the existing data
and query the simulators in
an offline mode
• Avoid the global
optimization by learning the
inverse design function f -1(y)
• Predict design variables in a
single step

32. 26/05/2023 32
Incomplete conditional density estimation
• Multimodal density estimation given
incomplete conditions
• However, integrating over h is still intractable, we
approximate the expectation by a function evaluation at
the mode

33. 26/05/2023 33
Generated alloys
example
• Known-alloy dataset:
15,000 variations from 30
known series of
Aluminum alloys
• BO-search dataset:
15,000 variations from
1000 found alloys by
Bayesian optimization
• Input: phase diagram |
Output: element
composition

34. 26/05/2023 34
Crystal structure generation
• Application in structure discovery: battery, aerospace
materials, etc.
• The stability of a solid-state crystal structure is connected
to its formation energy
• Target:
• Generate crystal-like structure
• Has low formation energy
• Diversity set of crystal structure candidates for active learning
Slide credit: Tri Nguyen

35. 26/05/2023 35
GFlownet
• GFlownet learns to generate the composition object:
• From the starting state, policy network output the
probability distribution over building blocks
• Select building blocks randomly based on the output
probability distribution and create a new state →
calculate the new probability distribution
• Repeat until reaching the terminal state
• Getting the reward from the environment (sparse
reward)
• The complete set of actions from starting state to
terminal state is a trajectory
• Flow is a non-negative function defined on the set of
complete trajectory
• GFlownet is trained by matching the flow going
through state: in-flow = out-flow
Slide credit: Tri Nguyen

36. 26/05/2023 36
GFlownet
• Advantage of GFlownet
• Diverse set of candidates → avoid
getting stuck in multi-modal
distribution (e.g stability/energy
landscape of crystal structure)
• Can sample in proportion to a given
reward function (crystal structure
generation: formation energy)
Slide credit: Tri Nguyen

37. 26/05/2023 37
Crystal structure generation with
GFlownet
• State:
• Multi-graph representation for structure:
• Node: atoms
• Node feature: element type, fraction coordinate
• Edges: built using near-neighbor-based method CrystalNN with search cut-off starting
from 13 and increasing to 20
• Edge feature: cell-direction vector ‘to_image’, bond distance
• 3D grid space: currently occupied and available position to insert new atom
• Action:
• Available fraction coordinate on a 3D grid.
• The chosen element
Slide credit: Tri Nguyen

38. 26/05/2023 38
GFlownet -
Forward policy
• Policy network:
calculate the probability
distribution over actions
Examples of
generated crystal
structure
Slide credit: Tri Nguyen

39. 26/05/2023 39
Topics
REPRESENTATION PREDICTION OPTIMIZATION &
GENERALIZATION
EXPLANATION

40. Explaining DTA deep learning model:
feature attribution
26/05/2023
Explainer
Protein
Drug
Deep learning model
Affinity = 6.8
Show the contribution of
each part of input to the
model decision
Does not show the
causal relationship
between the input and
the output of model
40
Slide credit: Tri Minh Nguyen

41. 26/05/2023 41
Drug agent
Protein agent
Actions:
Removing atoms,
bonds
Add atoms, bonds
Actions:
- Substituting one
residue of other
types with Alanine
DTA model
Drug-Target
Environment
Reward
+ Drug representation similarity
+ Protein representation similarity
+ Δ Affinity
Communicate to form a
common action
MACDA: MultiAgent Counterfactual Drug-target Affinity framework
Nguyen, T.M., Quinn,
T.P., Nguyen, T. and
Tran, T., 2022. Explaining
Black Box Drug Target
Prediction through
Model Agnostic
Counterfactual
Samples. IEEE/ACM
Transactions on
Computational Biology
and Bioinformatics.
Slide credit: Tri Minh Nguyen

42. 26/05/2023 42
The road ahead
Image source: pobble365

43. Grand framework
• Two-step paradigm:
• Step 1: Compress ALL materials knowledge into a giant model.
• Data, context as episodic memory | Model weights as semantic memory.
• Step 2: Decompress knowledge into something new.
• This requires learning to reason – learn how to manipulate existing knowledge.
• Search, plan are reasoning. Both aims to minimize an objective (e.g., matching or energy).
26/05/2023 43
Picture taken from (Bommasani et al, 2021)
• Learning to reason (zero-shot) all.
• Eying few-shot capability (e.g., materials
prompting).
• Leverage LLMs capability.

44. Prediction versus understanding
• We can predict well without understanding (e.g.,
planet/star motion Newton).
• Guessing the God’s many complex behaviours versus
knowing his few universal laws.
• → Automated laws discovery!
• → Abductive reasoning.
26/05/2023 44