Multivariate and biomarker analysis

Multivariate biomarker analysis is needed when the diagnostic or prognostic information lies not in a single marker but in the combination of several. It applies when you have a broad set of clinical, laboratory or molecular variables and are looking for the panel with better diagnostic, prognostic or predictive performance than any marker on its own. A marker-by-marker test ignores the correlation and the interaction between variables, raises the risk of spurious findings from multiple comparisons and does not estimate the combined performance of the panel. The service identifies that informative combination while controlling overfitting, and quantifies its true performance, not the apparent performance in the set where it was discovered.

You need a matrix of candidate variables (clinical, laboratory or molecular) measured in a cohort with a well-defined outcome of interest. The client measures and collects the markers; the service identifies, validates and interprets them. Ideally, the sample is large enough to split into a discovery set and a validation set, or at least to support honest cross-validation. The database does not need to arrive fully clean: missing data, basic consistency and the quality of the measurements are reviewed before any variable selection. If the database needs a full structural audit, it is worth planning a prior data quality phase. What does condition the scope is the imbalance between the number of variables and the number of samples, which is addressed explicitly.

It doesn't invalidate it, but it dictates the method. When there are far more candidate variables than samples, almost any combination looks predictive by chance, and that is the setting in which a false finding is most easily produced. Depending on the design and the sample size, the service can apply variable selection with overfitting control, such as Boruta, recursive feature elimination (RFE) or LASSO regularisation, and, where appropriate, dimensionality reduction (PCA) or regularised discriminant models, to identify the informative subset. The key is not just to select but to validate that selection: the stability of the panel is assessed under resampling, so that the variables chosen do not depend on one particular split of the data. The honest limit: with very few samples, the result is hypothesis-generating, not confirmatory.

Performance is measured by combining discrimination and calibration, not with a single statistic. Discrimination, the ability to separate classes, is quantified with the ROC curve and the area under the curve (AUC), together with sensitivity, specificity, predictive values and likelihood ratios. But a high AUC does not ensure that the predicted probabilities are reliable. That is why calibration is also assessed: whether an estimated risk of 30% corresponds to an observed frequency close to 30%. A model that discriminates well but is poorly calibrated can lead to the wrong management decisions. Cut-off points are set on clinical criteria, not by maximising a statistic. The cost of a false negative rarely equals that of a false positive.

It is not possible to ensure that a panel will keep exactly the same performance in an external cohort. What can be done is to reduce the model's optimism by separating discovery from validation from the outset and not letting the validation set contaminate any modelling decision. Most panels that fail do so because they were fitted so closely to the set in which they were discovered that they do not generalise: the published performance is that of the training set itself. The service works with honest cross-validation (variable selection happens inside each fold, not before) and, where data are available, with external validation in an independent cohort. The distinction is deliberate: cross-validation estimates generalisation within the same population; external validation tests it in another. Without that separation, the reported performance tends to be optimistic and cannot be read as solid evidence of generalisation.

You receive an interpretable result, not a black box. A predictive model that is hard to interpret faces more barriers to adoption in clinical research and is harder to defend to reviewers. That is why, alongside performance, we provide a reading of how each variable contributes to the prediction, using explainability techniques such as SHAP, which break down the weight of each marker in each case. On top of this comes the biological interpretation of the finding, grounded in training in clinical biochemistry and immunology: the biological hypothesis that a cytokine, an immunological parameter or a metabolite might point to when it carries weight in the model. That mechanistic reading is what helps separate a marker with biological meaning from a spurious correlation that passed the statistical filter by chance.

Multivariate and biomarker analysis is used when you start from a bounded set of candidate variables, such as clinical or laboratory variables or a targeted molecular panel, and the aim is to identify and validate the combination with diagnostic or prognostic performance. Applied bioinformatics and omics data analysis fits when the starting point is a high-throughput experiment (transcriptomics, proteomics, metabolomics) with tens of thousands of variables and specific needs for processing, normalisation and false discovery control at that scale. The two can be combined in the same project, for example an omics screen that feeds a reduced panel which is then validated, but they do not answer the same question: omics analysis explores the full space; biomarker analysis validates a candidate panel methodologically for research, publication or preliminary clinical decision-making.

No. The service covers the identification, methodological validation and characterisation of the diagnostic performance of a biomarker or panel: discrimination, calibration, robustness and external validation. That is traceable, reproducible scientific evidence. It does not include the formal analytical and clinical validation required for a regulatory agency to approve a diagnostic test: that is a regulatory process with its own framework, and it is outside ADI's scope. The boundary is clear: the service produces the methodological evidence that supports a biomarker and underpins a publication or a research decision; it neither issues nor manages the regulatory conformity of a diagnostic product. If you need the latter, you will find a methodological evidence base here, but not the regulatory validation or the management of the procedure.