### Iterative Molecular Discovery with Interpretable Deep Learning

Andrew White
University of Rochester
Department of Chemical Engineering

Research Overview
September, 2021

# The Whitelab

## ✨deep learning✨

Given $\vec{x}$ and $y$, find a function $\hat{f}(\vec{x})$ that predicts $\hat{y}$

$\hat{y} = \sigma\left(\mathbf{W}X + b\right)$
• $\sigma$ non-linear activation
• $\mathbf{W}$ trainable weights
• $\mathbf{W}$ trainable bias

I wrote a book on it: whitead.github.io/dmol-book

# Inductive Bias

Explicit choices made in architecture that reflect data-generation or solution space.

# Inductive Bias

## Mol Graph

• atom permutation equivariance
• bond permutation equivariance

## Point Cloud

• atom permutation equivariance
• translation equivariance
• rotation equivariance

# Inductive Bias

## Mol Graph

• atom permutation equivariance
C O N $\hat{f}(x)$
0 1 0 0 0.3
1 1 0 0 0.8
2 1 0 0 0.1
3 1 0 0 0.0
4 1 0 0 0.0
5 1 0 0 0.0
6 1 0 0 0.2
7 1 0 0 0.5
8 1 0 0 0.9
9 1 0 0 0.6
10 0 1 0 0.3

# Inductive Bias

## Mol Graph

• atom permutation equivariance
C O N $\hat{f}(x)$
3 1 0 0 0.0
1 1 0 0 0.8
2 1 0 0 0.1
0 1 0 0 0.3
4 1 0 0 0.0
5 1 0 0 0.0
6 1 0 0 0.2
7 1 0 0 0.5
8 1 0 0 0.9
9 1 0 0 0.6
10 0 1 0 0.3

# Inductive Bias

## Point Cloud

• atom permutation equivariance
• translation equivariance
• rotation equivariance

## Data/Training Tricks

Method Equivariance
Matrix Determinant Permutation Invariance
Eigendecomposition Permutation Invariance
Reduction (sum, mean) Permutation Invariance
Pairwise Vector/Distance Translation/Rotation Invariance
Angles Translation/Rotation Invariance
Trajectory Alignment Rotation/Translation Invariance
Molecular Descriptors All Invariant
Training/Testing Augmentation All Invariant

White, AD Deep Learning for Molecules & Materials 2021

## SMILES


CCCC1=CC=CC=C1C(=O)
N2CCCC[C@H]2C3CCCC3 


Sequence Network
(1D Conv, RNN, Transformer)

## Mol Graph

Graph Neural Networks

## Points

Equivariant Neural Networks

Preferred

If you must

Avoid

# The Power of Text

Text is excellent at predicting scalars — invariant to permutations.* Training is fast, implementations are easy, available in javascript🚀

Via canonicalization

## Predicting Chemical Shifts with GNNs📊

Yang, Z., Chakraborty, M. & White, A. D. Predicting chemical shifts with graph neural networks. Chemical Science (2021).

# Inductive Bias

• Chemical shift is per-atom. Need perm equivariance
• Chemical shift should be position dependent
• Chemical shift should not use descriptors to generalize

# Inductive Bias

• Chemical shift is per-atom. Need perm equivariance
• Chemical shift should be position dependent
• Chemical shift should not use descriptors to generalize

Graph Neural Network with Relative Distances

Graph Neural Network with Relative Distances

Graph Neural Network

Graph Neural Network

Graph convolutional neural network example.

White, AD Deep Learning for Molecules & Materials 2021

## Parameters

• 2,262 proteins, 361 metabolite molecules
• 5 million chemical shifts (526,000 fragments)
• Network Size: 3 edge FC layers (tanh), 4 GCN (relu), 3 FC Layers (tanh)
• Withheld 20,000 fragments (180,000 chemical shifts)
• Adam optimizer (0.001), huber loss, 0.15 dropout, 128 node feature dim, 4 edge feature dim
H RMSD H $$r$$ H$$^\alpha$$ RMSD H$$^\alpha$$ $$r$$ Parameter Number
Perfect 0.176 0.965 0.138 0.967
Model (H) 0.459 0.781 0.264 0.878 1,185,437
Model (all) 0.527 0.718 0.293 0.844 1,185,437
Medium 0.511 0.712 0.290 0.848 297,181
Small 0.501 0.726 0.288 0.849 42,123
Weighted 0.471 0.766 0.274 0.865 1,185,437
SHIFTX+ 0.455 0.787 0.248 0.890
SHIFTX+$$^*$$ 0.378 0.836 0.197 0.932
CS2Backbone 0.716 0.418 0.417 0.708

### Salt Bridge Sensitivity

pip install nmrgnn

## XAI with Counterfactuals🔦

Wellawatte, G. P., Seshadri, A. & White, A. D. Model agnostic generation of counterfactual explanations for molecules. (2021).

What is an explanation of a prediction?

1. Justification: reasoning for using a prediction, like test error
2. Interpretability: "the degree to which an observer can understand the cause of a decision"
3. Explanation: a presentation of information intended for humans that give the context and cause for a prediction

Miller T, Artificial Intelligence 2019

## Instance Explanations

1. Feature Explanation: which features contributed most to outcome?
2. Contrastive Explanation: what would the outcome be if the features were different?
3. Counterfactual Explanation: what are the closest features with a different outcome?

## Counterfactual

My grant proposal was rejected from NSF. If I had correctly formatted my references, it would have been funded.

## Chemistry Counterfactual

Counterfactuals are solution to optimization. Given prediction $x,\hat{f}(x)$, find $x'$ with

Challenging because optimization requires $\nabla_x\hat{f}(x)$

Use STONED Method to enumerate chemical spcae

## Output

Accuracy of method depends on being able to generate close, interpretable structures near $x$.

1. STONED: Generate structures with generative property of SELFIES
2. PubChem: Sample nearby structures deposited in PubChem

Nigam et al., Chemrxiv (2021)

## Alphabet of plausible tokens

pip install exmol

## Descriptor Explanations🍋

Gandhi H, & White, A. D. Local surrogate models for descriptor explanations (2021)

Unpublished

Counterfactuals are fundamental

They are one of the major theories of explanations in philosophy, including within it causality.

What about practicality? Can we act on these explanations?

Woodward J The Stanford Encyclopedia of Philosophy 2016

## Local General Explanation

Explain using local space around prediction with surrogate interpretable model

$\hat{y} = \beta \mathbf{X}$

where $\mathbf{X}$ are descriptors like solubility or number of hydrogen bonds

Riberio MT, et al. KDD Proceedings 2016

## Weighted Local Surrogate Model 🍋

Which features were significant for weighted linear regression around base?

$\DeclareMathOperator*{\argmax}{arg\,max} t_j = \frac{\hat{\beta_j}}{\sigma_{\beta_j}},\quad \hat{\beta} = \argmax_{\beta} \sum_i w_i\left(\hat{y}_i - \beta\mathbf{X}\right)^2$ $w_i = s(x_i, x_b)$

## What descriptors?

1. 160 MACCS Keys -- Are there nitrogen containing rings?
2. Solubility
3. Number of hydrogen bond donors
4. Number of heavy atoms

## Collaborative ML🤗

Zhu W, White AD, Luo J Federated Learning of Molecular Properties in a Heterogeneous Setting (2021).

## deep learning = big data

Chemical data is valuable and siloed. Enough data exists, but it is not shared

## Federated 𓃏 Learning

Rather than share training data, share your model

 Dataset Centeralized Training Federated Learning $$\alpha$$ MolNet* FedChem*$$_{\textrm{ours}}$$ FedAvg FedProx FedFocal FedVAT FLIT$$_{\textrm{ours}}$$ FLIT+$$_{\textrm{ours}}$$ FreeSolve$$\Downarrow$$ 0.1 1.40 1.430 1.771 1.693 1.686 1.371 1.634 1.228 0.5 1.445 1.376 1.322 1.299 1.366 1.127 1 1.223 1.216 1.294 1.150 1.277 1.061 Lipophilicity$$\Downarrow$$ 0.1 0.655 0.6290 0.6361 0.6403 0.6403 0.6556 0.6563 0.6392 0.5 0.6306 0.6365 0.6365 0.6333 0.6368 0.6270 1 0.6505 0.6474 0.6474 0.6488 0.6443 0.6403 ESOL$$\Downarrow$$ 0.1 0.97 0.6570 0.8016 0.7702 0.8022 0.7776 0.7788 0.7642 0.5 0.7524 0.7382 0.7708 0.7243 0.7426 0.7119 1 0.7056 0.6828 0.6822 0.7253 0.6705 0.6998 QM9$$\Downarrow$$ 0.1 2.35 0.0890 0.5889 0.6036 0.6164 0.5606 0.5713 0.5356 0.5 0.5906 0.5751 0.6059 0.5656 0.5658 0.5222 1 0.5786 0.5691 0.5822 0.5602 0.5621 0.5282

## Treating Heterogeneous Data

Focus on points of high error and disagreement

## Future of Deep Learning

Easy sharing of models, rather than data, to build global predictive models. Already used in location data in phones, training on image recognition, speech.

Perspective on GPT-3/Codex in chemistry