GeoTMI: Predicting Quantum Chemical Property with Easy-to-Obtain Geometry via Positional Denoising

1. Van Thuy Hoang
Network Science Lab
Dept. of Artificial Intelligence
The Catholic University of Korea
E-mail: hoangvanthuy90@gmail.com
240302
Hyeonsu Kim et.al., NIPS23

2. 2
Graph Convolutional Networks (GCNs)
 Generate node embeddings based on local network neighborhoods
 Nodes have embeddings at each layer, repeating combine messages
from their neighbor using neural networks

3. 3
Higher-order Graph Neural Networks
 Higher-order Graph Neural Networks | Semantic Scholar

4. 4
Quantum chemical calculations
 Quantum chemical calculations

5. 5
Overview
 Geometry optimization to obtain 𝑿 from 𝑋
 To obtain 𝑋 , a geometry optimization process with a starting
geometry 𝑋 should be preceded, which is also based on quantum
calculation

6. 6
Geometry optimization to obtain X
 Most 3D GNN studies have achieved great success in many quantum
chemical property (𝑌) prediction tasks using 𝑋.
 The usage of 𝑋 is infeasible in real-world applications.
 To tackle this, propose a novel training framework for 3D GNNs,
GeoTMI, which aims to predict high-level quantum properties from
an easy-to-obtain low-level geometry (𝑋).
 present a theoretical basis for fully exploiting such easyto-obtain
geometries to predict accurate target properties.

7. 7
The goal of GeoTMI
 obtain a proper representation Z˜ in predicting Y , by aligning it into
Z that contains more enriching information for Y .
 Propose a training framework for learning a proper representation
for predicting Y from X˜, which can be done by maximizing the MI
between the variables, Iθ(Z˜; Y ). This is somewhat similar to the
objective of the general supervised learning which predicts Y from X˜.
 However, the training only with X˜ to predict Y could be erroneous
because there is no guarantee a model utilize the proper information
resided in both X˜ and X.

8. 8
Introduce Z
 Implies undesirable information of Z˜ in predicting Y that is not
relevant to Z.
 X is sufficient information for the prediction of Y , and thus should
be minimized to zero in the optimal case
 In summary, the training process is about finding optimal model
parameters

9. 9
Overall framework
 The encoder design involves 3D GNN layers for both X and X˜,
sharing model parameters
 3D GNN models utilize roto-translational invariant 3D information as
their inputs

10. 10
Overall framework
 By maximizing three-term mutual information, we can account for
the physical inductive bias while accurately predicting Y
 However, because the I is intractable, we adopted a tractable lower
bound (LB) of it.

11. 11
Experiments
 Task 1. Molecular property prediction (QM9M)
 Task 2. IS2RE prediction (OC20)
 Task 3. Reaction property prediction (Reaction barrier height)*

12. 12
Molecular property prediction
 GeoTMI achieved consistent performance improvements across all
properties and models using X˜, demonstrating effectiveness of
GeoTMI.

13. 13
Experiments: IS2RE prediction (OC20)
 Both Noisy Nodes and GeoTMI show performance improvements
over the baseline Equiformer*, but GeoTMI achieves better
performance gains across all metrics.

14. 14
Conclusion
 We propose GeoTMI, a model-agnostic training framework designed
to exploit easy-toobtain geometry for accurate prediction of
quantum chemical properties.
 We envision that the GeoTMI becomes a new solution to solve the
practical infeasibility of high-cost 3D geometry in many other
chemistry fields.