Josh Southern

Subgraph GNNs with Walk-Based Centrality

Read Paper

We propose an expressive and efficient approach that combines the strengths of two prominent extensions of Graph Neural Networks (GNNs): Subgraph GNNs and Structural Encodings (SEs). Our approach leverages walk-based centrality measures, both as a powerful form of SE and also as a subgraph selection strategy for Subgraph GNNs.

Understanding Virtual Nodes

Read Paper

Work which explores how virtual nodes effect oversquashing and node heterogeneity for Graph Neural Networks

Curvature Filtrations for GGME

Read Paper

Collaboration with the AIDOS lab at Helmholtz Munich. Exploring the stability and expressivity of discrete curvature on graphs, and combining it with topological data analysis to obtain robust, expressive descriptors for Graph Generative Model Evaluation.

Evaluation Metrics for Protein Structure Generation

Read Paper

Collaboration with the Correia lab at EPFL developing better metrics for protein generation

Exploring “dark-matter” protein folds using deep learning

Read Paper

Collaboration with the Correia lab at EPFL developing a model to addresses the backbone designability problem by using protein sketch templates

Single domain scaffolding of non-overlapping protein epitopes

Read Paper

Collaboration with the Correia lab at EPFL and the Baker Lab describing an approach to simultaneously scaffold multiple functional sites in a single domain protein using deep learning

Genomic-driven nutritional interventions

Read Paper

using biological networks to map the space of food molecules and provide genomic-driven nutritional interventions to radiotherapy-resistant rectal cancer patients

Protein Docking

Read Paper

Collaboration with the Correia lab at EPFL. Creating a physics-informed deep neural network for rigid-body protein docking

Limbic

Paper

The first hire at this London based startup in October 2018. I have been working as a Machine Learning Researcher building emotion recognition technologies using data collected from a wearable device. We have implemeted this technology as a real-time product which increases data quality in mental health treatment.

Music and the brain

Read Thesis

Used tools from Information Theory (PMIME) to create a time-varying causal network between EEG time series. I then used network science approaches, particularly the Shannon entropy of the distribution of community sizes, to analyze differences in brain function for musicians when they are either improvising or playing a set piece.

Presented at NetSci 2019.

Emotion from smartwatch data

Read Paper

Extended work on emotion recognition from medical grade ECG devices to PPG collected from a smartwatch. Made statistical comparisons of the ECG and PPG time-series and used Transfer Learning to try and improve model performance.

Poster presented at ACII 2019.

Bayesian deep learning

Read Paper

Bayesian deep learning for emotion recognition. Used an LSTM-FCN to improve state of the art accuracy for predicting emotional valence from unimodal heartbeat data. Incorporated Bayesian considerations into the model using Monte-Carlo dropout for uncertainty quantification and to tune the accuracy and coverage.

I'm currently a Machine Learning Scientist at Prescient Design. I work on developing and applying advanced machine learning methods to Large Molecule Drug Discovery. Previously, I was a PhD student at Imperial College London, supervised by Prof. Michael Bronstein and Dr. Kirill Veselkov and part of the AI4Health CDT. I am interested in geometric deep learning, foundation/generative models and their application to drug discovery.

Prior to my PhD, I did my Masters degree in Applied Mathematics, also at Imperial College London, and then spent two years as a Machine Learning Researcher at a Health Tech startup working with wearable device data, trying to benefit the mental health space. Outside of academia, I am obsessed with sport (particularly tennis), pubs and jazz. Feel free to reach out if you want to discuss ideas, have questions about my research or want to recommend places in San Francisco or London!

Get in touch

Publications

2023

2023

2022

2022

2020

2019

© 2019