AI-driven discovery of microbial surface structures that prime the early-life immune system

PROJECT OVERVIEW

Early life is a critical developmental window for the human microbiome and the immune system. From the first bacterial colonisers at birth to the establishment of a stable microbial community by around three years of age, an increasingly diverse group of microbes stimulates and instructs the developing mucosal immune system, calibrating immune responses towards tolerance, barrier protection, and homeostasis. When microbial stimulation is excessive, insufficient, or imbalanced, immune priming may be disrupted, resulting in immediate and long-term health consequences.

For decades, microbiome research has focused on which bacterial species colonise the gut and which genes or metabolic pathways they encode. These approaches have transformed the field but do not capture the physical reality of immune system education, which is driven by interactions between three-dimensional protein structures displayed on or secreted by microbes (the microbial surfaceome) and those on host epithelial or immune cell surfaces. Until recently, this layer was largely out of reach: experimental structure determination was too slow to scale to the millions of proteins in a microbiome. However, computational advances (e.g., AlphaFold) now make large-scale protein structure prediction possible, as recognised by the 2024 Nobel Prize in Chemistry. These new opportunities for structural prediction, combined with the growing availability of human microbiome datasets that can be represented as graphs and advances in model interpretability, have created new opportunities to apply geometric deep learning to capture the biological complexity of microbial communities and uncover the structural patterns and mechanistic principles underlying host–microbe interactions.

This project will bring these recent computational advances together to investigate gut microbiome development at the host–immune interface in preterm infants, with and without necrotising enterocolitis (NEC), a devastating, often fatal condition of the immature preterm gut. NEC is characterised by gut inflammation and tissue death, in which disrupted microbial colonisation is strongly implicated. Specifically, we will construct an infant 3D surfaceome atlas spanning early postnatal development, capturing the microbial surface structures most directly exposed to the host immune system. Structural prediction will enable the assignment of proteins to homologous surface protein families with shared host-facing functions and resolve the large fraction of hypothetical or unannotated microbiome proteins.

By integrating structural embeddings with serial and paired multi-omics data, each infant microbiome will be represented as a graph of surface protein families, encoding their composition, diversity, abundance, and activity over time. Supervised graph classification will be used to distinguish healthy and NEC-associated microbiomes, while Monte Carlo tree search-based explainability will uncover the subgraphs of surface protein structures that contribute most strongly to model predictions.

Thus, by integrating protein structure prediction, graph representation learning, and explainable AI, the project (i) produces the first structural atlas of the developing infant gut surfaceome, (ii) identifies whether and when immune-instructive interactions become disrupted, excessive, or miscalibrated in NEC infants, and (iii) generates mechanistic hypotheses for experimental validation. Ultimately, the project aims to inform the future development of microbiome-based interventions rationally designed to modulate immune priming in preterm infants.

KEYWORDS

Infant gut microbiome, protein structure prediction, graph learning, immune priming, necrotizing enterocolitis

TRAINING OPPORTUNITIES

Interdisciplinary training opportunities across structural bioinformatics, explainable geometric deep learning, and clinical microbiome research are available. The student will gain expertise across the full functional microbiome research pipeline, including protein structure prediction and clustering, microbiome graph modelling, and interpretable GNNs, while developing an understanding of experimental design to generate validation data for computational predictions. Training will be tailored to individual backgrounds, enabling students from machine learning to gain biological expertise and those from biology or bioinformatics to strengthen their modelling skills.

KEY PUBLICATIONS

1. An infant nasal microbial gene atlas uncovers intervention-driven microbiome shifts and salt-resistant pathogen expansion. Cell Host Microbe. 2026

2. Linking microbial genes to plasma and stool metabolites uncovers host-microbial interactions underlying ulcerative colitis disease course. Cell Host Microbe. 2024

3. Bacterial low-abundant taxa are key determinants of a healthy airway metagenome in the early years of human life. Comput Struct Biotechnol J. 2021

4. [External] Evolutionary-scale prediction of atomic-level protein structure with a language model. Science. 2023

5. [External] Multiple protein structure alignment at scale with FoldMason. Science. 2026

THEMES

1. Structural bioinformatics: Structure prediction and fold-level clustering at microbiome scale.

2. Explainable geometric deep learning: Multi-relational graph representation and interpretable GNNs.

3. Early-life microbiome & mucosal immune education: Host–microbe instruction at the gut interface.

4. Neonatal & preterm health: Necrotizing enterocolitis as a clinically urgent testbed.

CONTACT INFORMATION OF ALL SUPERVISORS

Marie Pust 

Stephen Sansom