Hannes Whittingham Email and Phone Number

A machine learning and AI role centering around the ground-up construction of several machine learning models for chemical property prediction.Productionisation of models for company-wide real-time use by integration into AstraZeneca's chemical property modelling platform; use of models to steer generation of proposed druglike compounds by REINVENT, a generative AI platform built to suggest new drug-like compounds. Role also involved giving in-depth presentations on recent advances in AI, such as AlphaFold 1, AlphaFold 2 and ChatGPT in implementation-level detail, covering aspects such as self-attention, transformers, and the tuning of LLMs. An additional project, in which I created a large-scale data mining and wrangling process to provide a dataset used to find drug candidates to fulfil a novel and unusual role, was nominated for a company-wide R&D award and won a department-wide award.

Multiple projects in machine learning and exploratory data analysis and visualization. Indicative diagnosis models for Asthma and COPD: successful models built from demographic and questionnaire data after assessing a broad variety of supervised learning algorithms. Communication of model capabilities to a non-technical audience.Wearable Data Analysis: Learned to use AWS ML services; constructed automated data ingestion and cleaning pipeline for wearable data from COVID-19 vaccine trial, run daily. Exploratory analysis and data visualization; constructed patient biometric profiles; created combined timeline view of patient journey integrating wearable and clinical data.Provided in-depth ML tutoring for summer students (Gaussian Process Regression)

Two-year scheme comprising three 8-month deep learning research projects and a series of professional development courses. 20-page contribution to published textbook, ‘The era of Artificial Intelligence, Machine Learning and Data Science in the Pharmaceutical Industry’.Project 1 - Multitask Neural Networks and Proteochemometric Models for Activity Prediction across the Kinome (Prediction of enzyme inhibition from molecular structure)Compilation, cleaning and merging of data; cluster-based dataset partition for fairer assessment of model generalisation ability; hyperparameter optimization of two network architectures; experience with PyTorch. Best-performing model ensemble improved significantly on a similar recent publication. Released as CLT for company-wide use, with automated retraining tools for maintenance.Project 2 - U-nets for Differential Phase-Contrast (DPC) Image SegmentationImplementation of U-net to segment objects in DPC images. Scaling of the training pipeline with parallelized on-the-fly preprocessing and multi-GPU training. Improved performance by implementation of novel approach using custom loss function and smoothed training mask in correctly finding boundaries between objects.Project 3 (present) - Protein Structure Prediction using Graph Neural NetworksExplored an alternative to part of DeepMind’s AlphaFold workflow. Cleaning of PDB data; creation of novel graph representation of proteins; Heavily adapted and customised Gated Graph Neural Network architecture to enable training on very large, memory-intensive protein graphs. Mature use of Tensorflow: use of sparse tensors, flexible input shapes, model extensions using self-proposed tensor operations, custom loss function, loss mask for unknown coordinates.