Fungal infections are associated with 3.8 million deaths annually [1] and US costs above $19 billion. Limited antifungal options are increasingly undermined by resistance [2]. Prolonged Candida auris outbreaks in NHS hospitals and emerging pan-resistant strains make prediction and prevention urgent.

Current resistance prediction is dominated by genomic markers. However, genotype alone cannot explain why genetically similar fungal cells occupying different pre-treatment organisational configurations die, survive, recover or evolve resistance [2]. This DPhil will develop interpretable, uncertainty-aware multimodal AI that integrates high-content images and quantitative descriptors of cell-surface organisation with temporal responses, treatment context, isolate background, chemical information and genomic data [3]. Cryptococcus neoformans will be the primary system, with evaluation in other fungal pathogens. Training could potentially benefit from all domains of life.

Day-one proxy data include 35,821 RNA-seq samples across 220 fungal species in FungiExp [4], additional transcriptomic studies from FungiNetDB and GEO/SRA, genomic resources from FungiDB and MycoCosm, and curated mutation–drug associations from FungAMR [5]. These span multiple fungal species, antifungal treatments, genetic perturbations and stresses, supporting identification of conserved, compensatory or species-specific regulatory pathways associated with stress responses and drug targets.

Initially the project will focus on predictive models. As new experimental data will come online throughout the project, it will be possible to robustly evaluate predictive performance using time-split data. Versioned releases from parallel experiments will allow computational and experimental work to progress together.

In the final stage, the project will move from prediction to intervention by modelling how chemical and genetic perturbations move cells through the organisational landscapes. By integrating starting configuration, compound structure, treatment context, longitudinal susceptibility and endpoint sequencing, the framework will prioritise small molecules predicted to redirect high-risk configurations towards drug-sensitive regions and limit resistance emergence [3]. Top-ranked candidates will undergo prospective experimental testing.

The translational outputs will be calibrated early-warning models that map cell state to resistance phenotypes, interpretable configuration biomarkers and experimentally testable intervention candidates. Open software and a versioned fungal multimodal resource will provide a foundation for extending Oxford’s world-leading pathogen-genomics and drug-susceptibility data infrastructure beyond mycobacteria into fungal AMR. This will establish a distinctive Oxford capability linking AI, experimental mycology and antifungal discovery.