Prefix-tree Decoding for Predicting Mass Spectra from Molecules

Samuel Goldman, John Bradshaw, Jiayi Xin, Connor W. Coley

Advances in Neural Information Processing Systems 36 (NeurIPS) — Spotlight, 2023

We develop a new method for predicting mass spectra from molecules via first generating the product formulae (the x axis locations of mass spectra peaks) before secondly predicting their intensities (the peaks’ heights on the y axis). In order to overcome the combinatorial explosion in the possible number of product formulae, we design an efficient and compact representation for these formulae based on prefix trees.

Local Latent Space Bayesian Optimization over Structured Inputs

Natalie Maus, Haydn T. Jones, Juston S. Moore, Matt J. Kusner, John Bradshaw, Jacob R. Gardner

Advances in Neural Information Processing Systems 35 (NeurIPS), 2022

We develop LOL-BO, an algorithm to more efficiently perform latent space optimization (optimization over hard discrete spaces by transforming them into continuous problems) by recognizing latent space optimization is still high dimensional optimization.

Machine Learning Methods for Modeling Synthesizable Molecules

John Bradshaw

PhD Thesis, 2021

My PhD thesis, which brings together some of the papers listed below to describe physically inspired machine learning models for reaction prediction and de novo design.

Barking up the right tree: an approach to search over molecule synthesis DAGs

John Bradshaw, Brooks Paige, Matt J. Kusner, Marwin H. S. Segler, José Miguel Hernández-Lobato

Advances in Neural Information Processing Systems 33 (NeurIPS) — Spotlight, 2020

We describe how complex synthesis plans can be described using molecule synthesis DAGs and design a generative model over this structure. Not only can our model find molecules with good properties, but those that are synthesizable and stable too, allowing the unconstrained optimization of an inherently constrained problem.

A Model to Search for Synthesizable Molecules

John Bradshaw, Brooks Paige, Matt J. Kusner, Marwin H. S. Segler, José Miguel Hernández-Lobato

Advances in Neural Information Processing Systems 32 (NeurIPS), 2019

When designing a new molecule, you not only want to know what to make, but, critically, how to make it! We propose a deep generative model for molecules (Molecule Chef) that provides this information, by first constructing a set of reactants before reacting them together to form a final molecule.

Are Generative Classifiers More Robust to Adversarial Attacks?

Yingzhen Li, John Bradshaw, Yash Sharma

International Conference on Machine Learning (ICML), 2019

We propose and analyze the deep Bayes classifier, an extension of naive Bayes using deep conditional generative models. We show that these models (a) are more robust to adversarial attacks than deep discriminative classifiers, and (b) also allow the principled construction of attack detection methods.

A Generative Model For Electron Paths

John Bradshaw, Matt J. Kusner, Brooks Paige, Marwin H. S. Segler, José Miguel Hernández-Lobato

International Conference on Learning Representations (ICLR), 2019

We devise a method to model a wide class of common reactions through parametrizing movements of electrons, breaking down the modeling of a reaction into simple, intuitive steps. We come up with an approach to generate approximate electron movements from atom-mapped reaction datasets and show how this can be used to train our model.

Adversarial Examples, Uncertainty, and Transfer Testing Robustness in Gaussian Process Hybrid Deep Networks

John Bradshaw, Alexander G. de G. Matthews, Zoubin Ghahramani

Reliable Machine Learning in the Wild - ICML 2017 Workshop, 2017

We look at the properties of GPDNNs, Gaussian processes composed with deep neural networks. We show that unlike regular feedforward neural networks, GPDNNs often know when they don't know. This means that they often avoid making overconfident predictions on out of domain data.