Abstract
Early cancer detection via liquid biopsies is often hindered by sparse sampling and the high noise floor of benign DNA shedding. This paper presents a statistical framework that utilizes Signature Theory to analyze the dynamics of circulating tumor DNA (ctDNA). By characterizing the geometric "signature" of biomarker trajectories, the proposed method distinguishes malignant growth from benign equilibrium with high accuracy. Tested on simulated longitudinal data, the approach achieves over 94% detection accuracy using only seven blood samples collected over several years, demonstrating robustness to irregular sampling intervals.
Introduction
Liquid biopsies, which detect circulating tumor DNA (ctDNA) shed by tumors into the bloodstream, offer a non-invasive path toward early cancer detection. However, clinical implementation faces two primary hurdles: data scarcity and statistical noise. Patients typically provide blood samples at irregular, infrequent intervals (e.g., annually), and benign cells also shed DNA, creating a noisy background. Traditional mechanistic models often require dense data to estimate tumor growth rates accurately. This work addresses these challenges by shifting focus from parameter estimation to the geometric properties of the data path using Signature Theory.
Biological and Mathematical Modeling
The authors model the population of cancer cells and the resulting ctDNA levels as a stochastic Birth and Death Process (BDP). In this framework, malignant tumors are characterized as "supercritical" (growing), while benign states are modeled as being in equilibrium.
The dynamics of ctDNA shedding and elimination are captured through a series of differential equations and stochastic variables, summarized in the notation table below.
Signature Theory and Statistical Testing
The core innovation lies in the application of the Signature Transform, a tool from the field of "Rough Paths" in stochastic analysis. The transform converts a time-series signal into a sequence of iterated integrals, capturing the "shape" of the data path. Crucially, signatures are invariant to reparametrization, making the method naturally robust to the irregular sampling intervals common in healthcare.
The researchers derived theoretical probability distributions—specifically the Skellam distribution—for these signature coefficients under a "null hypothesis" representing the benign state. A particularly powerful feature identified is the "Levy Area" (the anti-symmetric part of the second-order signature), which detects the "twist" in data indicating a transition from stable benign shedding to explosive malignant growth. If a patient's signature deviates significantly from the theoretical benign distribution, a malignant state is flagged.
Experimental Setup
Due to the restricted nature of real-world longitudinal ctDNA data, the authors validated their pipeline using simulated datasets generated via the BirDePy Python package. The study included 4,000 benign trajectories (steady state) and 36,000 malignant trajectories (transitioning to aggressive growth). To mimic realistic clinical scenarios, each "patient" was sampled only seven times over a 2,000-day period—approximately one sample every 300 days.
Results
The proposed pipeline was evaluated against several signature coefficients ($T_2, T_{12}, T_{21}, T_{LA}$). Despite the extreme sparsity of the data (7 samples over ~5.5 years), the results were highly significant:
- Accuracy: > 94% across all tests.
- Precision/Recall: Approximately 0.999 for most tests.
- F1-Score: 0.999.
These results suggest that the geometric signature of the data path provides a much stronger signal for malignancy than individual data point values or traditional growth rate estimations.
Conclusion and Future Work
This study demonstrates that Signature Theory can effectively bridge the gap between sparse clinical data and complex biological dynamics. While the current results are based on simulated data, they provide a strong theoretical foundation for early detection. Future work will focus on validating the model against real-world biological noise and "batch effects," as well as exploring higher-order signature coefficients to further improve detection lead times.
Summary updated