A spectroscopic liquid biopsy for the earlier detection of multiple cancer types

Earlier detection of cancer is critical to improving survival because earlier diagnosis and treatment increase the opportunity to control disease progression. Early-stage cancers are more often amenable to surgical resection, reducing or avoiding the toxicity of radiotherapy or chemotherapy, whereas late-stage disease is associated with poorer prognosis. Most cancers are not screened for due to low prevalence and high cost per cancer detected. Many cancers present with non-specific early symptoms that can be overlooked. A rapid liquid biopsy platform capable of supporting clinicians in diagnosing different cancers, particularly in patients with non-specific symptoms or cancers not targeted in screening programmes, could be transformational.

Study design and cohort: A large-scale discovery study included 2092 patients, of whom 1542 had confirmed diagnoses of one of eight cancers (brain, breast, colorectal, kidney, lung, ovarian, pancreatic, or prostate). The non-cancer cohort comprised asymptomatic controls (NC; n=91) and symptomatic non-cancer patients (NCS; n=436) presenting with generic symptoms and benign conditions. Samples: Patient serum samples were analysed using the Dxcover® Cancer Liquid Biopsy platform based on FTIR spectroscopy. Each patient sample yielded nine spectra. Spectroscopy captures a pan-omic biochemical signature reflecting tumour- and immune-derived markers. Machine learning and validation: A nested cross-validation strategy was used to develop classifiers and prevent data leakage. For model development, patients were randomly split into training and test sets with a 70:30 split, repeated five times. The inner CV (fivefold on the 70% training set) was used to tune hyperparameters and select probability thresholds to target desired sensitivity or specificity. The outer CV evaluated performance on the held-out test sets. Per-patient predictions were obtained by aggregating spectrum-level predictions via majority vote across the nine spectra. Spectra from the same patient were never present in both training and testing. Organ-specific classification: Separate classifiers were trained for each cancer type versus symptomatic non-cancer controls (NCS). Ovarian cancer was compared against female-only NCS (NCS-F), and prostate cancer against male-only NCS (NCS-M). Thresholds were selected to achieve minimum 90% sensitivity or specificity during cross-validation, and additional operating points targeting ≥45% sensitivity or specificity were explored. Pooled classification: Exploratory pooled models were trained to classify ‘cancer’ versus asymptomatic non-cancer (C vs NCA) and ‘cancer’ versus all non-cancer (C vs NC), with ROC analyses and classification metrics reported, and detection rates stratified by cancer stage. Feature importance: Post hoc analyses identified discriminative spectral regions (e.g., Amide I/II bands, phosphodiester stretching, lipid-associated bands) contributing to classifications.

- Organ-specific performance versus symptomatic non-cancer controls (AUC): brain 0.90, breast 0.76, colorectal 0.91, kidney 0.91, lung 0.91, ovarian 0.86, pancreatic 0.84, prostate 0.86. - Examples of tuned operating points: sensitivity-tuned lung (93% sensitivity/78% specificity) and kidney (92%/79%); specificity-tuned brain (74%/91%) and colorectal (77%/90%). - Pooled classification, cancer vs asymptomatic non-cancer (C vs NCA): AUC 0.94. Sensitivity-tuned: 98% sensitivity at 59% specificity. Specificity-tuned: 99% specificity at 57% sensitivity. Stage-specific detection: at high sensitivity, Stage I 99%; at high specificity, Stage I 64% (and overall early-stage sensitivity 57%). - Pooled classification, cancer vs all non-cancer (C vs NC): AUC 0.85. Sensitivity-tuned: 90% sensitivity at 60% specificity. Specificity-tuned: 95% specificity at 40% sensitivity. - Stage detection rates reported high across stages under sensitivity-tuned settings and remained substantial under high-specificity settings, indicating strong early-stage detectability. - Feature importance highlighted biologically relevant spectral regions, notably Amide II (~1530 cm−1), Amide I (1600–1700 cm−1), phosphodiester stretching (~1260, ~1074, ~1167 cm−1), carbohydrate-associated bands (~1025, ~1061 cm−1), and lipid-associated high-wavenumber bands (~2861–2947 cm−1), with cancer-type specific patterns.

The study addresses the need for earlier multi-cancer detection by leveraging a pan-omic spectroscopic signature from serum that captures both tumour- and host-response information. By training and validating machine-learning classifiers against symptomatic non-cancer controls—a more clinically representative and challenging comparator set—the approach demonstrated robust discrimination across eight cancer types and strong performance in pooled ‘any cancer’ detection. The tunability of operating points along the ROC curve allows adaptation to different clinical pathways that may prioritise high specificity (to minimise false positives and downstream procedures) or high sensitivity (to maximise early detection). Performance compares favourably with more complex and costly liquid biopsy methods, and integration with orthogonal biomarkers such as cfDNA and clinical risk data could further enhance accuracy. Importantly, early-stage detection rates (including Stage I) were high under sensitivity-tuned settings, supporting the potential clinical value for earlier intervention and triage of symptomatic patients.

The Dxcover® spectroscopic liquid biopsy shows promise for earlier detection of multiple cancer types, including high sensitivity to Stage I–II disease. The platform is rapid, low-cost, minimally invasive, and can be tuned for sensitivity or specificity to fit diverse diagnostic pathways. While further prospective validation, larger cohorts, and exploration of combined testing strategies are needed, the technology could facilitate efficient triage and earlier diagnosis, potentially improving patient outcomes.

- Information content: Spectroscopy does not provide genomic data to guide treatment decisions; complementary DNA-based testing may be required downstream. - Control cohort composition: All cancer types were compared against a shared symptomatic non-cancer group; cancer-specific symptomatic controls (e.g., benign disease mimics) were not used, which could refine accuracy estimates. - Sample size and balance: Although prior work suggests ~200 samples per class can be adequate, larger and more balanced cohorts would increase robustness. - Generalisability and prevalence: As an early-stage discovery study, real-world prevalence and deployment settings remain to be defined; positive predictive values will vary with population. - Data access: Under GDPR, raw datasets are not publicly available, limiting external replication with the exact data. - Minor inconsistencies and typographical variations in reported device naming across sections; prospective, standardised protocols are needed in future trials.