Life Sciences & Pharma

Client Context

A biotech company developing a novel biomarker panel for early-stage ovarian cancer detection needed to validate the panel's diagnostic accuracy. They had a case-control study with 180 patients (90 confirmed ovarian cancer, 90 benign pelvic masses) and candidate biomarker measurements for six proteins.

The Challenge

The company needed to determine the optimal combination of biomarkers for maximum sensitivity and specificity, account for the prevalence of ovarian cancer in the target screening population (approximately 1 in 10,000), and produce results suitable for both a regulatory submission to the MHRA and an investor data room. Previous analyses by an in-house team had produced overoptimistic estimates due to overfitting on the same dataset used for feature selection.

Our Approach

We used cross-validated logistic regression with LASSO regularisation to select the biomarker combination, preventing overfitting by embedding feature selection within each fold of a 10-fold cross-validation. We computed receiver operating characteristic (ROC) curves and area under the curve (AUC) for both the training and held-out folds. Positive and negative predictive values were calculated at realistic disease prevalence. We also performed sensitivity analyses with different cut-point strategies (Youden's index, closest-to-corner, and clinical cost ratio). All analyses were conducted in R with the caret and pROC packages.

Results

The optimal panel of four biomarkers achieved a cross-validated AUC of 0.91 (95% CI: 0.86 to 0.95), with sensitivity of 84% and specificity of 87% at the Youden-optimal cut-point. At the population prevalence of 0.01%, the positive predictive value was 0.07%, confirming that the panel was suitable for triage rather than population screening. The MHRA accepted the validation methodology, and the company used the findings to secure Series A funding.

Client Testimonial

"AnalyticsScholar's approach to validation was exactly what regulators and investors needed to see. Their transparency about the limitations — particularly the PPV at low prevalence — built trust rather than creating problems."

— Chief Scientific Officer, OncoDetect Ltd

[Explore More Case Studies → /case-studies/]

[Discuss Your Life Sciences Project → /contact/]