AI Meets the Blood Drop: Scientists Map the Rules for Merging Liquid Biopsy Data in Cancer Research

A single drop of blood can carry traces of a tumor: fragments of its DNA, proteins shed from its cells, vesicles loaded with molecular cargo, and even whole circulating tumor cells drifting through the bloodstream. Turning those scattered molecular whispers into reliable clinical predictions, however, demands far more than simply feeding the data into an algorithm. A new methodological article published in Frontiers in Artificial Intelligence on 23 October 2026 lays out an end-to-end roadmap for combining artificial intelligence with liquid biopsy features in cancer studies, and reveals that the field is far less mature than the buzzwords suggest.

The work, led by Mireia Andueza and Núria Malats of the Spanish National Cancer Research Centre (CNIO) together with collaborators including Klaus Pantel and Catherine Alix-Panabières of the European Liquid Biopsy Society, is explicitly a methods paper. Its goal is not to announce a new cancer test but to diagnose the state of the research itself. Liquid biopsy, a concept introduced in 2010, refers to the analysis of biological fluids such as blood, urine, and saliva for tumor-derived biomarkers, including circulating tumor cells (CTCs), cell-free DNA (cfDNA), circulating tumor DNA (ctDNA), tumor-derived extracellular vesicles (EVs), tumor-educated platelets (TEPs), and microRNAs. These analyses hold promise for early detection, identification of the tissue of origin, monitoring of minimal residual disease, prognosis prediction, therapeutic target identification, uncovering resistance mechanisms, and surveillance of treatment response, all through examinations that are less invasive and more cost-efficient than tissue biopsies.

To gauge how often researchers actually integrate these multiple molecular layers into a single artificial intelligence model, the team ran a PubMed search covering human studies published between January 2020 and January 2026. Combining terms for cancer, liquid biopsy, machine learning, deep learning, or artificial intelligence, and early detection, minimal residual disease, or prognosis yielded 144 records. After screening, and with four additional articles found through reference lists, only 25 studies made the final descriptive analysis. Five were prospective and 18 retrospective; 19 reported early cancer detection outcomes, including multi-cancer early detection. The numbers expose a striking gap: only six of the 25 studies performed external validation on a completely independent dataset, and none assessed calibration or clinical utility. The authors caution that the search was narrative rather than systematic, so the count should not be read as a definitive measure of the entire literature.

That scarcity, the authors argue, partly reflects how demanding this kind of research is. Every stage of the workflow carries methodological traps. It begins with defining the study objective and clinical context, whether early detection, prognosis, or treatment monitoring, because that choice dictates the design, the patient selection criteria, and the endpoints. Among the 25 studies reviewed, selection criteria were described in 22. Most relied on retrospective case-control designs, which are cheaper and faster but prone to bias and incomplete clinical data, whereas prospective studies allow real-time, controlled data collection that strengthens validation.

Sample size is another critical and frequently neglected element. Only two papers in the search estimated sample size through power analysis, targeting area under the curve values of 0.95 and 0.98. The article points to established tools such as the pmsampsize package in R and Stata for conventional prediction models, while stressing that machine learning and deep learning settings complicate the picture: effective model complexity depends on signal strength, sparsity, regularization, and the learning algorithm itself, and oncological evidence suggests machine learning approaches may require substantially larger development samples than conventional regression. For validating an already-trained model, methods such as Sample Size Analysis for Machine Learning (SSAML) offer model-agnostic estimates designed for precision rather than model fitting.

Once data are collected, preprocessing becomes the battleground where overoptimistic results are often born. The authors detail strategies for handling missing data, from MissForest and k-nearest neighbors imputation to Multiple Imputation by Chained Equations, and draw an important distinction between missing values within a data layer and missing modalities, where an entire assay is unavailable for some patients. Reconstructing a whole unmeasured layer rests on strong assumptions about the relationships between modalities, and models that natively accommodate missing modalities may be preferable. Zeros in the measurements also demand care, since a lack of biomarker detection does not necessarily imply its absence; true biological zeros should be preserved while values below detection limits should be treated as censored measurements. Batch-effect correction, meanwhile, must be learned on training data only and applied unchanged to validation samples, because estimating correction parameters on the full dataset constitutes a form of data leakage.

On the modeling side, the article walks through dimensionality reduction techniques such as principal component analysis, UMAP, and multiomics factor analysis, alongside feature selection methods ranging from simple filters to embedded approaches like LASSO and tree-based algorithms. Multimodal fusion, the process of combining omics layers, comes in early, late, and intermediate flavors. In one multi-cancer early detection study cited by the authors, an early fusion strategy using extreme gradient boosting on concatenated features of targeted and genome-wide methylation, ctDNA copy number alterations, fragment length, and end motifs actually outperformed a late fusion stacking ensemble, a reminder that the fusion strategy itself can be decisive. Emerging tools are also surveyed: transformer-based models have been applied directly to cfDNA fragments, and large language model foundations such as LLaMA have been fine-tuned to predict cancer from cfDNA, though their advantages for multiomics integration remain unproven given small sample sizes and heavy computational demands.

Perhaps the sharpest message concerns evaluation. Discrimination, the ability to distinguish between outcome states, must be paired with calibration, the agreement between predicted probabilities and actual outcomes, measured through observed-to-expected ratios, calibration slopes, and plots. The authors illustrate why with a sobering example: an early detection model with an area under the curve of 0.90, run at 95% specificity and 60% sensitivity on a population with 1% disease prevalence, would flag roughly 6 true positives alongside 50 false positives per 1,000 people screened. High discrimination alone does not guarantee clinical usefulness. Decision curve analysis and net benefit calculations, they argue, should become standard, as should interpretability methods such as SHAP, with the caveat that feature importance reflects prediction contribution, not proof of causal biology.

The article closes with an eight-point checklist urging researchers to define objectives clearly, understand the statistical properties of their data, choose appropriate validation strategies, keep preprocessing confined to training data, report software versions and random seeds for reproducibility, and pursue external validation, model updating, and clinical utility assessment. The authors also flag unresolved challenges, including fragmented ethical and legal approval processes across Europe that delay data sharing, and the promise of generative artificial intelligence for data augmentation and imputation. For a field hoping to turn drops of blood into life-saving predictions, the roadmap is now written; the harder task will be following it.

Subject of Research: Artificial intelligence methods for integrating multi-layer liquid biopsy biomarkers in cancer research

Article Title: Learning from drops: AI-guided integration of liquid biopsy features in cancer studies

Article References: Andueza, M., Villoslada-Blanco, P., De Dreuille, B., Alonso, L., Sabroso-Lasa, S., Pantel, K., Alix-Panabières, C., López De Maturana, E., & Malats, N. (2026). Learning from drops: AI-guided integration of liquid biopsy features in cancer studies. Frontiers in Artificial Intelligence, 9, Article 1937336. https://doi.org/10.3389/frai.2026.1937336

Image Credits: AI Generated

DOI: 10.3389/frai.2026.1937336

Keywords: liquid biopsy, artificial intelligence, machine learning, cancer detection, multiomics integration, ctDNA, external validation, model calibration, data preprocessing, clinical utility, deep learning, study design