Search

Machine learning with deep neural networks ("deep learning") allows for learning complex features directly from raw input data, completely eliminating hand-crafted, "hard-coded" feature extraction from the learning pipeline. This has lead to state-of-the-art performance being achieved across several---previously disconnected---problem domains, including computer vision, natural language processing, reinforcement learning and generative modelling. These success stories nearly universally go hand-in-hand with availability of immense quantities of labelled training examples ("big data") exhibiting simple grid-like structure (e.g. text or images), exploitable through convolutional or recurrent layers. This is due to the extremely large number of degrees-of-freedom in neural networks, leaving their generalisation ability vulnerable to effects such as overfitting. However, there remain many domains where extensive data gathering is not always appropriate, affordable, or even feasible. Furthermore, data is generally organised in more complicated kinds of structure---which most existing approaches would simply discard. Examples of such tasks are abundant in the biomedical space; with e.g. small numbers of subjects available for any given clinical study, or relationships between proteins specified via interaction networks. I hypothesise that, if deep learning is to reach its full potential in such environments, we need to reconsider "hard-coded" approaches---integrating assumptions about inherent structure in the input data directly into our architectures and learning algorithms, through structural inductive biases. In this dissertation, I directly validate this hypothesis by developing three structure-infused neural network architectures (operating on sparse multimodal and graph-structured data), and a structure-informed learning algorithm for graph neural networks, demonstrating significant outperformance of conventional baseline models and algorithms.

Nanobodies (Nbs) are a class of single-domain antibody derived from the immune systems of camelid species. They achieve binding affinities and specificities to target antigens comparable to those of classical antibodies (Abs), despite being ten times smaller (~15 kDa) and having only three variable loops. This raises the question of how these binding affinities and specificities are achieved in such a compact molecule. To address this, a novel dataset of Nb-antigen co-crystal structures are assembled and investigated. Findings are presented in three main chapters, with a fourth describing a collaborative Nb engineering project. First, the sequence and structural diversity of Nb domains is analysed and compared with that observed in a comparative set of Ab domains. Nbs are shown to display greater structural diversity across all three loops, in conjunction with enhanced conservation across the framework regions. Second, the way in which Nbs bind antigens is investigated. Class-averaged properties of antigen-contacting residues (the `paratope'), show Nb paratopes are more variable than those of Abs. This is true for both the distribution of paratope residues across the domain, and the types of residues used at interfaces. Notably, Nbs deviate from the `loops = paratope' assumption which is true for Abs; an insight which has implications for Nb selection, modelling and engineering strategies. Third, the sequence-structure-paratope relationship of Nb CDR H3 loops is interrogated. Previous chapters show the H3 loop is the most structurally diverse region, and critical in determining antigen-binding specificity. Here, Nbs are clustered into three distinct classes based on H3 loop structural features. The classes have distinct sequence features and use different distributions of paratope residues; suggesting loop conformation and antigen-binding orientation may be inferred from Nb sequence. Finally, an anti-GFP Nb is engineered by structure-guided mutagenesis for enhanced binding to a closely related antigen. The engineered Nb is being trialled for use in a novel super resolution microscopy method.