Evolution, Evolvability, Expression and Engineering

Abstract

This thesis describes how to build machines (Engineering) that answer questions about: (a) Evolution & Evolvability and (b) Expression.

In the first part of this thesis, I present a framework for understanding and engineering biological sequences, and solving sequence→function problems by building ‘Complete Fitness Landscapes’ in sequence space. This framework for measuring, modelling and designing biological sequences is built around the idea of learning an ‘oracle’ (typically a deep neural network model that takes a sequence as input and predicts its corresponding function) to traverse these ‘Complete Fitness Landscapes’. Here we develop a (promoter sequence)→(gene expression) oracle and use it with our framework to design sequences that demonstrate expression beyond the range of naturally observed sequences. We also show how our framework can be used to detect signatures of selection on a sequence, and to characterize robustness and evolvability.

The second part of this thesis describes two frameworks for inferring from single-cell and spatial gene expression measurements: ATLAS (A Tool for Learning from Atlas-scale Single-cell datasets) and insi2vec (a framework for inferring from spatial multi-omic and imaging measurements).