Ken Pienta, Former Director at The Johns Hopkins University School of Medicine, shared a post on Substack:
“Cancer kills 10 million people a year, and the decision architecture we use to fight it is structurally incapable of telling a clinician what will happen next in the patient sitting in front of them. Here is why.
In 2022, about 20 million people were newly diagnosed with cancer worldwide. About 10 million died of it. Roughly 27,000 deaths every day. The International Agency for Research on Cancer and the 2025 Global Burden of Disease analysis project that, on current trajectory, those numbers reach 35 million new cases and 18 million deaths annually by 2050.
The oncology enterprise has not been idle in the face of those numbers. Over the past twenty – five years we have committed an extraordinary fraction of biomedical research dollars to targeted therapy, immunotherapy, and precision genomics.
In specific tumors and specific patient subgroups, that investment has produced real, durable wins. But it has not bent the aggregate mortality curve, and the projections above assume that it will not. The reason is not a shortage of drugs. The reason is structural. The dominant decision architecture in oncology – the staging systems, the prognostic models, the trial endpoints, even most of the digital twins now being built – was designed for average populations, not individual trajectories.
It is excellent at telling a clinician what has been true for patients like this one in the past. It is structurally incapable of telling a clinician what will happen next in this new patient under each of several candidate treatment paths.
Three specific architectural failures – three blind spots – compound to produce that incapacity. They are not failures of effort or talent. The people building these tools are some of the best clinicians and scientists alive. The tools themselves are missing pieces that, once you see them, are very hard to unsee.
Blind spot 1: the exposome collapsed to a checkbox
Open any oncology staging system, any tumor board workflow, any prognostic calculator in routine clinical use. Look at how it represents what the patient has been exposed to over the course of a lifetime.
You will find three fields. Tobacco, usually in pack-years. Alcohol, usually in drinks per week. Occupational exposure, almost always a binary yes/no for a small handful of industries. That is the entire causal history of carcinogenesis as represented in mainstream oncology decision tools.
It is not enough. It is not close to enough. Carcinogen exposure is not a checkbox. It is a chemistry problem: which of the IARC Group 1 carcinogens this person has been exposed to across a lifetime, in what doses, in what combinations, processed through which genotype – specific enzyme variants, deposited in which tissues. Each step in that chain has measurable variability, and the variability compounds.
We need a tool that lets us measure how much that compounding matters. That runs realistic multi – carcinogen exposure profiles through them, and integrates the result across the cytochrome P450, N-acetyltransferase, glutathione S-transferase, aldehyde dehydrogenase, sulfotransferase, and UDP-glucuronosyltransferase enzyme families that determine whether a given carcinogen is detoxified or activated in a particular person.
We need to expand the exposome past three checkboxes. The most important causal information about why this cancer arose in this person – the information that should be setting the prior for every downstream model of what the tumor is and what it will do – is absent from the current architecture. Single-carcinogen, single-gene, population-average models cannot recover it, because they are not designed to. The variance lives in the interaction structure: which carcinogens the patient encountered together, in which tissues, through which enzymes. Linear single – factor models cannot represent that.
The blind spot is not that we lack the data. We can measure it. Stable isotopes and mass spectrometry data that measure exposures, blood biomarkers, residential exposure history, occupational records, pharmacy fills, genotyping – every piece of this is technically achievable today. The blind spot is that the decision tools do not have a place to put it.
Blind spot 2: the tumor microenvironment as a fixed parameter
Most oncology decision tools – and the majority of oncology digital twins published to date – represent the tumor as a single growing object. Drug sensitivity is a fixed parameter, or an efficacy scalar derived from a population-average trial. The immune compartment, the stromal and vascular environment, the local resource landscape, and the subclonal competition that ultimately decides whether resistance emerges are either absent from the model or collapsed to a single summary statistic.
This is wrong about what a tumor is, and the wrongness has a specific shape.
A tumor is an ecosystem. It is a mosaic of competing subclones preyed upon by anti-tumor immune populations, facilitated by pro-tumor immune and stromal populations, dependent on vasculature that gates drug delivery, and shaped by a resource landscape that determines which phenotypic strategies survive. The interaction matrix among these populations – who suppresses whom, who facilitates whom, at what rates – is not fixed. It evolves under treatment. It shifts across the tumor’s edge-interior axis. It resets after perturbation. A model that treats these dynamics as fixed parameters cannot, by construction, forecast a trajectory or individualize a treatment decision, because the parameters it is treating as fixed are the very variables that determine the answer.
The post you are reading right now is the second in a series. The first laid out the prairie-dog-town homology – the argument that tumors are bounded ecological communities, not invading species. That argument is the same argument I am making here, run forward into the modeling layer. If a tumor is a bounded ecological community, then the structure of the model you use to forecast its behavior had better also be a bounded ecological community, with the relevant species, interactions, and resource flows explicit. Otherwise you are reasoning about a town with a single-population growth equation, and you will get the answer that single-population growth equations always give: too high, too fast, and indifferent to the structure that actually decides the outcome.
There is good news here. The mathematical scaffolding for ecological modeling is mature. Generalized Lotka–Volterra dynamics for the species, evolutionary game theory for the strategies, Bayesian inference for the latent interaction matrix, control theory for the therapeutic decision. None of this needs to be invented. It needs to be assembled and tested at the scale and resolution that oncology demands. That assembly is what our Cancer Ecology Center program is developing.
Blind spot 3: the model is static, not data-assimilating
Even the best mechanistic digital twins published in oncology to date share a third architectural property that quietly defeats them. The majority are calibrated once, at baseline, with whatever data are available at that one moment, and then they run forward. They produce a single forecast trajectory from a fixed initial condition.
That architecture is mismatched to the clinical reality of cancer. A patient with advanced disease typically generates multiple observation modalities – imaging, circulating tumor DNA, tissue genomics, laboratory panels, performance status, increasingly continuous physiologic data from wearables – at irregular, sparse intervals over months to years. Each observation contains real information about what is happening to the latent state of the disease. A model that does not continuously update its estimate of that latent state is throwing away most of the information the clinical workflow is already collecting. We pay for the data and then we discard it.
The right architecture has been understood for decades in other fields. It is called data assimilation, and it is the standard machinery of weather forecasting, financial risk modeling, GPS, and aerospace tracking. You maintain a probability distribution over the latent state of the system rather than a single point estimate. Between observations, you propagate that distribution forward with a dynamical model. When an observation arrives, you update the distribution using a Bayesian filter that weights the new information against your prior belief in proportion to how reliable each is.
Weather forecasting is the cleanest analogue. The reason the seven-day forecast is now better than the three-day forecast was in 1980 is not that meteorologists got smarter. It is that they switched from single deterministic runs to ensembles continuously assimilating satellite, radiosonde, buoy, and aircraft observations. Each new observation tightens the posterior. The forecast improves as the data accumulates.
Oncology has not made that transition. Our digital twins are mostly the equivalent of running a single weather model from Monday’s observations and asking it to predict Friday. They do not assimilate Tuesday. They do not assimilate Wednesday. By Friday they have drifted, and we blame the model rather than the architecture. We are building a Cancer Ecology Digital Twin to explicitly fix this – a state-space model over the ecological variables, a generalized Lotka–Volterra dynamical core, a latent interaction matrix, and a Bayesian ensemble Kalman filter that updates the posterior whenever new clinical data arrive. The design is not novel. The application is.
Why the three compound.
The three blind spots are individually serious. They are catastrophic in combination, and the combination is what we actually use in clinic.
Without an individualized exposure history, the prior for the initial tumor state is wrong. Without an ecological model of the tumor, the forward projection from that wrong prior is also wrong, and wrong in a particular direction – toward homogeneous growth and away from the heterogeneous, structured behavior that determines resistance. And without data assimilation, the errors do not get corrected by the longitudinal observations the workflow generates. They accumulate.
You can see the compounding in real outcomes. Cancer survival improves modestly each year on average. The improvements concentrate in patient subgroups whose tumors happen to match the population averages encoded in the trials. The patients who fall outside those averages – different exposure history, different tumor ecology, different trajectory shape – are the ones for whom the models systematically fail, and they are disproportionately the patients who die. The architectural blind spots are not abstractly wrong. They are wrong precisely where the mortality lives.
None of this is an argument that targeted therapies do not work, or that genomics is overrated, or that precision oncology is a mistake. The argument is narrower. Those tools are necessary and insufficient. The decisions they support are correct on average. They are not correct for the individual sitting in clinic, and the individual sitting in clinic is the unit of oncology that actually matters.
What it would take to close them.
A reframe is not a research program. The reason I am willing to write this in public is that the worldwide research enterprise needs to evolve to close all three blind spots simultaneously. This is what we are trying to do at the Cancer Ecology Center. An ideal state to aspire to:
- A patient’s exposure history is captured at high resolution through carcinogen-flux modeling, multi-enzyme genotyping, and a stable-isotope assay cheap enough to deploy globally:
- that history sets a quantitative prior for risk and tumor state:
- the tumor itself is modeled as a bounded ecological community using the same mathematical machinery ecologists have used for half a century:
- and the full system is wired into a digital twin that maintains a posterior distribution over the latent state and updates it via Bayesian assimilation each time a new observation arrives from the clinical workflow.
Treatment becomes a control problem on that posterior – what intervention, at what magnitude, at what time, moves the trajectory toward the outcome the patient and clinician want – rather than a search through population averages.
The next several posts walk through explaining and solving these issues one at a time.
For this week, the claim is the smaller one. The aggregate cancer mortality curve has not bent, despite extraordinary investment, because the architecture of the tools we use to make decisions is missing three specific pieces. Adding more drugs to the formulary will not fix that. Adding the missing pieces will.”
Other articles about Cancer Types on OncoDaily.