Cancer Likelihood in Plasma is a concept in computational oncology that refers to the statistical probability of a malignancy being present, estimated from molecular and cellular signals found in blood plasma. Plasma, the liquid component of blood, carries cell-free DNA (cfDNA), circulating tumor DNA (ctDNA), proteins, metabolites, and extracellular vesicles shed by tumors. Advances in high-throughput sequencing and Machine learning have enabled the development of models that integrate these signals to produce a likelihood score, often termed a cancer likelihood score, which can guide clinical decisions regarding further diagnostic imaging or biopsy.
The concept emerged from the broader field of liquid biopsy, which aims to replace or complement invasive tissue sampling. Early work in the 2010s demonstrated that ctDNA mutations could be detected in plasma, but sensitivity for early-stage cancers was limited. The integration of multiple analyte types - such as methylation patterns, fragmentomics, and protein biomarkers - improved performance, leading to multi-cancer early detection (MCED) tests. These tests typically report a likelihood score and a predicted tissue of origin, enabling screening for cancers that lack standard screening modalities.
Biological Basis
Plasma contains a mixture of nucleic acids released from apoptotic and necrotic cells throughout the body. In cancer patients, a fraction of this cfDNA originates from tumor cells and carries somatic mutations, copy number alterations, and aberrant methylation patterns. The concentration of ctDNA varies with tumor burden, vascularity, and turnover rate. Additionally, tumors secrete specific proteins and glycoproteins, such as carcinoembryonic antigen (CEA) and cancer antigen 125 (CA-125), which have long been used as serum biomarkers, albeit with limited sensitivity and specificity.
Recent research has focused on fragmentomics - the analysis of cfDNA fragment size and end motifs. Tumor-derived fragments tend to be shorter and exhibit different end sequences compared to normal cfDNA. Machine learning models can exploit these subtle patterns to discriminate cancer from non-cancer samples. Methylation-based approaches, such as those targeting CpG islands, offer another layer of information, as cancer genomes often display widespread hypomethylation and focal hypermethylation at promoter regions.
Machine Learning Approaches
Estimating cancer likelihood from plasma data typically involves supervised learning. A model is trained on a cohort of known cancer cases and healthy controls, with input features derived from sequencing or proteomic assays. Common algorithms include logistic regression, random forests, and gradient boosting, but deep learning methods have gained traction. Neural networks, particularly residual networks and U-Net variants for image-like methylation maps, can capture complex nonlinear relationships.
Feature engineering is critical. Raw sequencing data are transformed into features such as mutation allele fractions, methylation beta values, fragment size distributions, and copy number profiles. Dimensionality reduction techniques like principal component analysis are often applied before training. The models output a probability score between 0 and 1, with a threshold chosen to balance sensitivity and specificity. Calibration is essential to ensure that the score reflects true likelihood, often achieved through isotonic regression or Platt scaling.
Training requires large, well-annotated datasets. Public resources like The Cancer Genome Atlas (TCGA) provide genomic and clinical data, but plasma-based datasets are scarcer. Several companies, including GRAIL (now part of Illumina), have generated proprietary datasets from clinical trials. The use of AI in this domain has raised questions about interpretability, leading to research on explainable AI methods such as SHAP values to identify which features drive the likelihood score.
Clinical Applications
The primary application is early cancer detection. MCED tests aim to identify cancers at a stage when treatment is more effective. For example, the Galleri test, developed by GRAIL, uses targeted methylation analysis and a machine learning classifier to detect over 50 cancer types. In the PATHFINDER study, the test demonstrated a specificity of 99.5% and a positive predictive value of 43.1%, meaning that among positive results, 43.1% were confirmed to have cancer. The test also predicted the tissue of origin with high accuracy, guiding subsequent diagnostic workup.
Beyond screening, cancer likelihood in plasma is used for minimal residual disease (MRD) monitoring. After curative-intent surgery, serial plasma samples are analyzed for ctDNA; a rising likelihood score indicates recurrence. This approach has been shown to detect recurrence months before radiographic imaging. In metastatic settings, the score can reflect treatment response, with declining ctDNA levels correlating with tumor shrinkage.
The concept also informs risk stratification in asymptomatic individuals with hereditary cancer syndromes. For carriers of BRCA1 or BRCA2 mutations, plasma-based tests may complement existing surveillance protocols, although evidence for clinical utility is still emerging.
Challenges and Limitations
Several challenges impede widespread adoption. Sensitivity for early-stage cancers remains suboptimal, particularly for stage I disease where ctDNA levels are often below the detection limit. Clonal hematopoiesis of indeterminate potential (CHIP) can produce false positives, as mutations in blood cells are mistaken for tumor-derived signals. To mitigate this, assays often sequence white blood cells in parallel to subtract background mutations.
Cost is a significant barrier. Whole-genome sequencing of cfDNA is expensive, and the computational infrastructure required for analysis is substantial. Reimbursement policies are evolving, but many tests are not yet covered by insurance. Regulatory approval varies by jurisdiction; in the United States, the FDA has granted Breakthrough Device designation to some MCED tests, but full approval requires prospective clinical trials demonstrating mortality benefit.
Interpretability and bias are also concerns. Models trained on predominantly European-ancestry populations may perform poorly in other groups, leading to disparities. Efforts are underway to diversify training cohorts and validate tests across ethnicities.
Future Directions
Research is exploring the integration of plasma-based likelihood scores with other data modalities, such as AI-analyzed imaging and electronic health records. Multi-modal models could improve accuracy and provide more personalized risk assessments. The use of large language models to parse clinical notes and extract relevant features is an emerging area.
Technological advances in sequencing, such as nanopore platforms, may reduce costs and turnaround times. AWS Trainium and other specialized hardware could accelerate model inference in clinical settings. Additionally, the development of standardized reference materials and benchmarking frameworks will facilitate comparison of different tests.
Longitudinal studies are needed to establish the clinical utility of plasma-based likelihood scores in reducing cancer mortality. The NHS-Galleri trial in the United Kingdom, which enrolled 140,000 participants, is expected to provide definitive evidence. If successful, cancer likelihood in plasma could become a routine component of preventive healthcare.