When I began working in cancer genomics, the goal seemed clear: catalogue mutations, define subtypes, and link them to outcomes. We were building increasingly detailed snapshots of tumours, and for a long time that worked. But as datasets have grown in scale and complexity, a different picture has started to emerge – one that computation is helping to bring into focus: tumours don’t sit still.

Across genomics, single-cell and spatial data, what we see is not a stable system but one in constant flux. Cells shift between phenotypic states, adapt to their environment, and respond to pressures like therapy in often reversible ways. Yet many of our models and assumptions remain based on static categories. This mismatch is becoming a central challenge in cancer research.

Rethinking cell identity in cancer

Single-cell technologies were expected to define the “parts list” of a tumour. Instead, they have shown how blurry those parts are. Rather than discrete cell types, we often observe continuous spectra of cell states.

Take the epithelial-to-mesenchymal transition (EMT), a process where epithelial cells become more mobile and invasive. Instead of distinct categories, EMT unfolds along a continuum of intermediate phenotypes. In our own computational modelling of these transitions, we found that cells rarely commit fully, instead occupying hybrid states that are highly context-dependent, often transient and notably difficult to predict.

This suggests that focusing on what a cell is at a single time point may be less informative than understanding what states it can access, and under what conditions. This is where computation becomes essential. The trajectories cancer cells follow are not directly observable; they must be inferred from high-dimensional data. By modelling how cells move through state space, we shift from asking not just where they are, but where they might go next – a subtle, but important step towards prediction.

Context shapes cell behaviour

Another shift has come from spatial data, with spatial omics and imaging highlighting that location within the tissue strongly influences behaviour. Using methods from geography and ecology, we can now map how cells interact across space.

In our analyses of breast tumours, we found that stromal cells such as myofibroblasts don’t just influence nearby cancer cells; they can exert effects across surprisingly long distances, reshaping entire tissue regions and promoting invasive behaviour. More broadly, we see that a cell’s neighbourhood can be as informative as its genome. By combining graph neural networks with spatial statistics, we can predict cancer cell states from their immune and stromal context. Often, the microenvironment explains more variation than genomic alterations. Stable states, like mesenchymal ones, are easier to predict, while hybrid states remain harder to capture.

This reframes plasticity as something that can be quantified. Rather than treating it as biological noise, we can begin to measure which cells are more likely to change state, and under what conditions. This is particularly relevant upon treatment, where transient, drug-tolerant states enable cell survival and later relapse.

These observations support a more ecological view of cancer, where cell states emerge from interactions within structured environments. They also raise the possibility of identifying “plastic niches”: tumour regions where cells are more likely to transition between states and develop resistance. Targeting these high-risk spatial zones could allow us to anticipate and prevent resistance rather than react to it.

From snapshots to trajectories

If spatial data has changed how we think about context, time remains the dimension we understand least. Much of cancer biology still relies on static snapshots, even though the processes are dynamic.

We now know that “normal” tissues are not genetically pristine. Healthy cells accumulate mutations and expand clonally long before cancer appears, making the transition to malignancy more gradual than once thought. Similarly, treatment does not simply select resistant clones; it can reshape cell states, with cells moving through transient adaptive phases before becoming stably resistant. Without longitudinal data, these trajectories are largely inferred rather than observed.

Linking these stages together – normal tissue, pre-cancer, treatment and relapse – requires models that integrate data across both time and space. While longitudinal datasets remain limited, computational approaches are increasingly able to infer these trajectories by combining genomic, transcriptomic and spatial information.

The promise and limits of AI

Advances in AI, particularly foundation models trained on large-scale biological data, offer new ways to capture complex patterns across scales. These models can, in principle, learn subtle regulatory programmes that traditional approaches may miss.

But there is a catch. They are good at recognising stable cell types, yet struggle with dynamic, plastic states central to cancer progression. In our work developing models for cellular plasticity, we find their performance is highly context-dependent. For instance, a model trained to capture EMT induced by TGFβ may fail to capture the same biological process under EGF stimulation. Even when adapted, their ability to generalise across different experimental conditions remains limited.

This highlights a broader challenge: in cancer, where cell behaviour highly depends on context (tissue, microenvironment, treatment) and no dataset will ever capture all possible conditions, increasing model complexity does not guarantee deeper biological insight. Simpler models can sometimes perform as well or better. The most effective approaches will likely combine methods, using AI for pattern discovery and large-scale simulations alongside mechanistic, interpretable models.

At the same time, AI is beginning to connect molecular data with clinical practice. Models integrating histopathology and omics can infer molecular features directly from routine tissue sections, potentially bringing insights on cell states, spatial organisation and plasticity into standard clinical settings.

Towards predictive models

All of this points towards a common goal: moving from describing tumours to predicting their behaviour. Can we identify which pre-cancerous lesions will progress? Anticipate how tumours will adapt to therapy? Or even steer cells towards more stable, targetable states?

Answering these questions depends not only on analysis but also on how data is generated. Datasets that capture space, time and treatment together – still relatively rare – will be crucial for building predictive models.

Computation is central to this shift. Not simply by analysing larger datasets, but by providing frameworks to model how cancer evolves and responds to intervention. The challenge is no longer measuring cancer, but modelling its possible futures.

By modelling cancer as a dynamic, evolving system, we move closer to understanding not just what we see in a biopsy, but what might happen next – and that is what will ultimately improve detection and treatment.