Biology is entering a new era defined not just by data, but by intelligence. Over the past decade, single-cell sequencing and spatial transcriptomics have made it possible to map tissues at unprecedented resolution, cell by cell, molecule by molecule, and in spatial context. Yet as these datasets grow in scale and complexity, the central challenge has shifted: converting multi-scale omics data into knowledge that is robust, generalisable, and clinically actionable.
Here, I will introduce SemanticST, our AI-based framework for scalable spatial transcriptomics analysis that preserves subtle biological signals. SemanticST learns biologically meaningful tissue representations, enabling the discovery of rare cellular micro-niches, fine-grained tissue architecture from ultra-high-resolution datasets such as Xenium, while maintaining interpretability, an essential requirement for clinical trust. At the same time, SemanticST, like most current AI methods in biology, remains primarily task-specific: it performs within defined settings, but reliable generalisation across technologies, cohorts, and patient populations remains a broader challenge. This limitation is driving a major shift toward foundation models for biology: models that learn transferable biological representations and can adapt across tasks, datasets, and modalities. To assess the state of the field, we systematically benchmark six state-of-the-art foundation models across the tasks. Our results show promise, but also persistent gaps in domain generalisation, interpretability, and robustness to platform variability.
We then connect these issues to a high-stakes clinical application: immunotherapy response prediction including immune checkpoint inhibitors, where model failures directly affect patient stratification and treatment decisions. Through large-scale benchmarking across independent cohorts, we show that current transcriptomics-based models still struggle with reliable cross-cohort performance, underscoring the gap between research-grade models and clinically deployable AI.
Looking forward, the future of AI_in_Biology will not be defined simply by larger models or more data, but by biologically grounded intelligence, models that understand biology across scales, modalities, and patient populations.