Precision medicine in oncology is entering a new phase. For Pablo Tamayo, Professor of Genomics and Precision Medicine at UC San Diego School of Medicine, the next frontier lies not simply in generating more molecular data, but in building models that can integrate that data in ways that are transparent, biologically meaningful and clinically useful.
Tamayo sees two major challenges ahead: one algorithmic and one data-driven. On the algorithmic side, he argues that the field needs more effective multi-omic models capable of integrating genomic, transcriptomic, proteomic and clinical information. These models should not depend solely on low-level features such as individual mutations or single-gene expression changes. Instead, they should use higher-level representations, including pathways, biological processes and cellular circuitry.
This shift, he explains, could make models more robust, less prone to overfitting and easier for researchers and clinicians to interpret. “Interpretable multi-omics at scale” means building systems whose internal logic can be understood without requiring every user to be a machine learning expert. The goal is not to reduce models only to what biology already knows, but to ensure that new data-driven patterns can still intersect with existing biological and clinical knowledge.
For real-world treatment decisions, Tamayo emphasizes the importance of evidence. Rather than relying on opaque black-box predictions, he points to inferential frameworks that quantify how much evidence different biomarkers, pathways or representations provide for a given outcome, such as response to a therapy. Bayesian statistics and information theory, he suggests, offer useful ways to formalize this process.
AI and foundation models may have an important role to play, but Tamayo is cautious about where they are most useful. In his view, AI may be especially valuable as an interface that allows researchers and clinicians to interrogate complex models in natural language: asking why a prediction was made, exploring what-if scenarios, or identifying patterns of evidence across patients. However, in clinical prediction, where datasets are often small, noisy and fragmented, models must be tightly constrained, parsimonious and interpretable.
Tamayo’s perspective is shaped by his experience developing widely adopted tools such as Gene Set Enrichment Analysis and GenePattern. One lesson from that work is that successful tools need to combine sophistication with usability. They should automate parts of the analysis, but also help users build intuition and think carefully about the biological meaning of their results.
Looking ahead, Tamayo expects progress in precision medicine to be incremental rather than immediate. He argues that the field needs a more strategic view: instead of asking whether to use AI, researchers should ask what clinical problem they are trying to solve, whether the right data exist, and what kind of model can realistically support confident therapeutic decisions.
For precision oncology to move beyond molecular profiling, the next decade will require better data access, stronger integration across modalities, interpretable model architectures and a clearer consensus on the problems that matter most. Only then can multi-omic data more reliably support confident, evidence-based treatment decisions.
