Transforming Unimodal Silos into Computable Medicine

Session Focus

Mahmood framed the discussion around two ideas.

The first is scale. Patients are evaluated at many levels: records, scans, tests, treatments and outcomes, alongside biological information such as genomics, plus data collected for research rather than care, including spatial transcriptomics and proteomics. All of it is generated for different purposes.

The second is the evolution of machine learning over roughly five years, which he summarized as a transition from supervised bespoke models trained for specific tasks, to unimodal foundation models representing complex data in forms channelable to many tasks, and now to multimodal foundation models incorporating several or all modalities to represent the entire patient. The same shift is occurring on the research side, combining spatial transcriptomics and proteomics representations.

Layered on that are reasoning models from frontier labs, which can interact with representation learning models to produce additional insight, plus agentic frameworks and an emerging move toward simulation and world models.

What Spatial Organization Reveals

Tiffany Chen MD noted that most of the panel were pathologists, and that the bread and butter of pathology lies in tissue. Pathologists have always seen tissue spatially even without framing it that way, because a diagnosis does not rest on what a single cell is. It rests on where that cell sits and what it is engaging with. The difference between an invasive tumor and a precursor lesion is precisely where the cell is.

Her underlying claim is that biology is a continuum. Cells are not in one bucket or another; they change constantly, as do their relationships and interactions, across tissue and across time.

The immunoscore illustrates the point. It is a spatial assay examining where T cells sit, inside the tumor or at the tissue edge. Identifying a cell as a T cell at single-cell resolution does not provide the full picture, while its position predicts patient outcome in ways traditional staging has not achieved.

The ambition is to move beyond defining what each cell is and where it sits toward understanding why it is there and what it is going to do.

Reading Tissue at Different Depths

Asked how proteomics and spatial transcriptomics get used by pharma and in discovery, Chen described them as the same view of tissue at different depths.

Her analogy: the H&E slide is Google Maps, showing who is there, the cars, cells, people and houses. Spatial transcriptomics reveals what instructions they hold and what they plan to do. Spatial proteomics shows what machinery they have built, what they intend and whether they are building toward it. That combination is the best way to understand what a cell is actually doing.

All three levels matter both for understanding the biology and for predicting whether a patient will respond to a therapeutic, which in turn guides trial design and ultimately how patients receive a drug.

Foundation Models as Computational Sensors

Rowland Pettit MD, PhD, who recently completed clinical pathology residency at Mass General Brigham, framed the problem as actionability. The volume of data generated exceeds what any clinical care team can appreciate, synthesize or deploy at the point of care.

His framing of a foundation model is as a computational sensor. Humans have senses that sample the environment; a foundation model ingests data types across modalities. Structured data includes laboratory values, medications and vital signs. Unstructured data includes clinical text, diagnostic reports and long oncology progress notes. Each holds a different piece of the patient’s trajectory and must be appreciated with nuance to become actionable.

His first application is resolving fragmented ontologies, and his example was deliberately mundane. Identifying a patient with type 2 diabetes currently depends on data existing in incompatible formats. It might appear as clinical text reading T2DM or type 2 diabetes, as the ICD-10 code E11.9, or as a hemoglobin A1c above 6.5 on a chart.

Making that actionable previously required structuring exercises and layers of logic to determine who the diabetic patients are. A medical foundation model can ingest those elements as they exist, as messy as they are, and appreciate in a high-dimensional representation space that this patient has diabetes.

That, in his view, is the fundamental principle of the current era: ingesting disparate data types, unifying representation to understand medical concepts, and making them actionable.

The Broader Frame

Jorge Reis-Filho MD, PhD, previously chief AI and innovation officer at AstraZeneca, described his current work as delivering biomedical superintelligence through the convergence of AI and biomedical technologies, compounding innovations becoming available across the field.

Celestine Schnugg described Boom Capital as an anti-thesis firm, operating on the principle that once you understand something you have already missed it. The firm has backed roughly 60 companies, about 70 percent in life sciences and tech bio, and is beginning to build disease-specific translational biobanks.

Key Takeaways

1. Position predicts outcome where cell identity alone does not, which is why spatial assays outperform traditional staging.

2. Biology is continuous rather than categorical, and models can now learn from that continuum.

3. Three depths of tissue reading: H&E shows who is present, transcriptomics shows intent, proteomics shows capability.

4. Foundation models function as computational sensors, ingesting structured and unstructured data across modalities.

5. Fragmented ontologies are the first solvable problem. One diagnosis can appear as text, a code or a lab value, and a model can unify them.

6. Reasoning models layered onto representation models are where additional clinical insight is expected to come from.