Measuring Cancer by Its Chaos, Not Its Quantity

June 10, 2026

liquid-biopsy epigenetics quantum-machine-learning cancer-detection computational-biology

The most promising liquid biopsy tests detect cancer best when it’s already at Stage IV — which is roughly the opposite of what early detection means. GRAIL’s Galleri test, the most prominent multi-cancer screening product on the market, achieves 90% sensitivity at Stage IV. At Stage I, that number drops to 18% across all cancer types. You are more likely to miss an early-stage cancer than catch it. The test is, functionally, a very expensive way to confirm what a symptomatic patient probably already suspects.

This isn’t a failure of execution. It’s a failure of premise.

The Wrong Question

Current liquid biopsy approaches are mostly asking: how much tumor DNA is floating in the blood? The logic is intuitive — cancer cells shed DNA, so more cancer means more DNA, which means a stronger signal. But early-stage tumors are small. They’re poorly vascularized. They shed almost nothing. The signal is buried under a flood of normal cfDNA from healthy tissue undergoing its own routine cell death. You’re trying to hear a whisper in a stadium.

The biological deck is stacked against this approach in a deeper way too. Fast-growing, aggressive, highly vascularized tumors shed the most DNA. So tests optimized for abundance will find those. Slow-growing indolent tumors — the kind you actually have time to intercept — barely register. The test is inadvertently biased toward finding the cancers that are already winning.

What the field is converging on instead is a different question entirely: not how much but how disordered.

Measuring Chaos

Cancer development is fundamentally stochastic. Tumor cells accumulate epigenetic errors rapidly and incoherently — the methylation patterns that regulate gene expression become chaotic, inconsistent from cell to cell, untethered from the tight regulatory logic of healthy tissue. Researchers at the Johns Hopkins Kimmel Cancer Center formalized this intuition into something called the Epigenetic Instability Index (EII). Rather than looking for absolute changes in methylation levels, the EII measures the random variation in those patterns — the degree of epigenetic noise.

The results are striking. By analyzing over 2,000 cancer methylation datasets, the team identified 269 specific genomic regions where this stochastic variability acts as a reliable signal. Applied to early-stage lung and breast cancers, the EII achieved roughly 81% sensitivity for Stage IA non-small cell lung cancer and 68% for early-stage breast cancer — both at 95% specificity. Compare that to the 18% Stage I sensitivity for abundance-based tests. The difference isn’t incremental. It’s a different paradigm.

The insight feels almost philosophical once you sit with it: we’ve been searching for a specific broken rule, when we should have been measuring the collapse of rule-following itself.

Why Quantum Kernels Fit This Problem

Here’s where it gets interesting to me as someone who thinks about computation. The EII approach generates high-dimensional, sparse data — you’re looking at subtle statistical signatures across hundreds of genomic regions, and you’re trying to do it with whatever clinical samples exist, which is never a million patients. Rare cancer subtypes, early-stage cohorts, pediatric cases: the data is always going to be limited. Classical deep learning is not well-suited for this. It needs scale. Feed it a small, high-dimensional dataset and it will overfit or fail to generalize.

Quantum kernel methods approach this differently. Instead of learning a function by brute-force gradient descent over massive data, they encode inputs into a high-dimensional Hilbert space using quantum circuits, then evaluate similarity between data points in that space. The key property is that you can explore an exponentially large feature space with relatively few parameters and relatively little training data. Researchers at the University of Chicago demonstrated this concretely — using quantum machine learning to distinguish tumor-derived exosomes from healthy ones by their electrokinetic properties, outperforming classical methods on minimal training data.

That’s the actual value proposition: not quantum as a buzzword, but quantum as a structural fit for the specific geometry of the problem. Small cohorts. High dimensionality. Subtle, non-linear patterns. The EII sits squarely in that space.

I want to be honest about the caveats here because the hype usually isn’t. Current quantum hardware is noisy and limited. The “barren plateau” problem — where gradient signals vanish in deep quantum circuits — is real and unsolved in general. The data loading bottleneck alone can negate theoretical speedups. What’s more realistic near-term is hybrid quantum-classical architectures, where the quantum processor handles specific high-dimensional kernel evaluations while classical systems manage everything else. That’s not as sexy but it’s probably what actually ships.

What Changes If This Works

I’ve been thinking about what it would mean to reframe cancer screening around disorder rather than quantity. It would change what we’re sequencing for: not hunting rare mutations but reading epigenetic entropy, which means native DNA sequencing via Nanopore long-reads becomes more relevant than PCR-amplified short reads, because bisulfite conversion — the standard way to detect methylation — degrades already-fragile cfDNA. It would change the computational stack: small-cohort, high-dimensional methods over massive-dataset classical approaches. And it would change what “early” means in early detection — possibly reaching back to precancerous lesions before a tumor microenvironment even forms.

The uncomfortable question I keep returning to: if we’ve spent decades optimizing for the wrong signal, how many other diagnostics are similarly misframed? How many conditions are we measuring by their loudest symptom when the meaningful signal is in the noise structure?

That’s probably worth sitting with longer than the answer deserves.