Cancer as Evolution: The Combinatorial Trap

This series has been building a worldview of cancer as a complex, dynamical system. Cancer is an evolving population (Post 1). Treatment reshapes that population through selective pressure (Post 2). The tumor and its microenvironment co-evolve as an ecosystem (Post 3). And these dynamics unfold across multiple timescales simultaneously, faster than our current ability to observe and respond (Post 4).

This final post asks the question the series has been converging on: given everything we now understand about how cancer evolves, why haven’t we been able to use that understanding to defeat it? And what will that take?

We have the knowledge, the drugs, and the data. What we lack is the capacity to reason about a problem that is combinatorially vast, with tools that cannot operate at the scale the disease demands.

Start with a simplified version of the decision an oncologist faces. Depending on the cancer, they have somewhere between 10 and 50 approved agents to work with: cytotoxic chemotherapies, targeted inhibitors, immunotherapies, hormonal therapies, antibody-drug conjugates. These can be given alone or in combination, at different doses, in different sequences, and started, stopped, or modified at different points based on response.

The number of possible strategies is astronomical. For 20 agents, the two-drug combinations alone number 190; the three-drug combinations, 1,140. Layer in dose, sequence, and timing, and the space grows combinatorially — far beyond what any trial program could explore.

Clinical trials sample a vanishing fraction of it. A Phase III trial compares one experimental regimen against one standard of care, in a defined population, over years of design, enrollment, and analysis. It answers a single question: is regimen A better than regimen B, on average, here? It cannot tell you which of the thousands of untested regimens might have been better, or which patients within the trial would have done better on a different choice.

Randomized trials remain the most rigorous tool we have for causal evidence. The limitation is structural: the space of possible treatments grows combinatorially, while the method for testing them grows one trial at a time. The gap widens with every new approval.

The treatment space is only half the problem. The other half is how the tumor responds.

For any treatment, the tumor population can adapt along many evolutionary paths. Post 2 described the simplest case: one drug (erlotinib) selecting for one dominant resistance mutation (T790M) in EGFR-mutant lung cancer. In practice, resistance to a given therapy can arise through dozens of mechanisms, genetic and non-genetic, operating in parallel or in sequence.

Consider combination immunotherapy (anti–PD-1 plus anti–CTLA-4) for melanoma. The tumor can escape by losing antigen presentation (B2M mutations, HLA downregulation), losing interferon signaling (JAK1/JAK2 mutations), engaging alternative inhibitory checkpoints (VISTA, TIGIT, LAG-3), recruiting immunosuppressive cells (regulatory T cells, myeloid-derived suppressor cells), excluding T cells through stromal remodeling, or reprogramming the metabolic microenvironment (IDO-driven tryptophan depletion, adenosine accumulation). Each is a different evolutionary trajectory, and several can run in the same tumor at once.

So the clinician faces a compound problem. The treatment space is combinatorially large. The response space is combinatorially large. And the two are coupled: each treatment constrains the tumor’s possible responses, and each response constrains the next treatment. A resistance mechanism selected first-line may confer cross-resistance to second-line agents, or a collateral sensitivity a well-chosen agent could exploit. The optimal strategy depends on predicting which responses follow which choices, across sequential decisions, in a system changing in real time.

The trap closes here: the decision space is too large to search by trial and error, the dynamics too complex to navigate by intuition, and the stakes high. Every treatment reshapes the landscape for every treatment that follows.

Oncology has built sophisticated tools for this. Molecular tumor boards match genomic findings to targeted therapies. Guidelines distill trial evidence into decision trees. Biomarker-driven, basket, and umbrella trials test drugs against molecularly defined subtypes in parallel. These are genuine progress, and they have improved outcomes.

They also share a limitation: they treat cancer as a set of static molecular features rather than a causal and dynamic evolutionary system.

A tumor board sees a BRAF V600E mutation and recommends dabrafenib plus trametinib. Sound, evidence-based medicine — but it doesn’t account for the resistant subclone that may already exist, the cells’ capacity to enter a persister state, the immune context of the microenvironment, or what the tumor will look like when the regimen fails and which options will remain. Guidelines encode what works on average for a molecular profile; they don’t model the individual tumor’s trajectory. Two patients with identical genomic profiles can diverge completely, because their tumors differ in clonal architecture, microenvironment, spatial organization, and epigenetic state — none of which a genomic profile captures.

The ceiling is one of bandwidth, not knowledge or intent. Oncologists know tumors evolve, that resistance is often predictable, that heterogeneity matters. But the number of interacting variables (genomic, transcriptomic, epigenetic, spatial, temporal, ecological) exceeds what any person or committee can integrate in real time, the strategies exceed what any trial program can test, and the dynamics move faster than the decision cycle can adapt.

Stated fully, the problem is this: cancer is a multi-scale evolutionary system whose state is described by high-dimensional, multi-modal measurements that change over time in response to both its own physical dynamics and our therapeutic interventions. Optimal treatment means integrating across every measurement dimension, modeling the dynamics they describe, predicting the tumor’s response to candidate interventions, and choosing strategies that account for the full space of possible trajectories. Furthermore, this must be accomplished for each patient individually, and fast enough to keep pace with a disease evolving in real time.

No human, however expert, can hold all of this at once. No static model can capture its dynamics. No trial program can exhaust its vast search space.

The biological understanding and the measurement capability are both here. What is missing is the capacity to reason about the system at the scale and speed the disease demands. Cancer has been precisely stated, precisely measured, and precisely characterized as the evolutionary system it is.

What remains is building something that can reason about all of it at once. That requires an entirely new paradigm of computational biomedicine. That is also precisely what we are building at Galen Health.

This concludes the Cancer as Evolution series. For the foundational biology, start with Cancer from First Principles. For a primer on measurement technologies, continue with How We Learned to Read Cancer.

First Principles is published by Galen Health, where we create intelligent systems that solve disease from first principles. If you would like to learn more about what we do, reach out.