In summary we propose to use discovery proteomics data directly for diagnostic and in future prognostic procedures, bypassing extensive assay development. ADAPT-MS is the framework we developed to enable this vision.

The beauty is that with every measured sample, the prior knowledge growth and enables better and more classification tasks. With growing databases these can be tailored to ever more specific subpopulations fitted to covariates and local specific effects.

This procedure again brought up wider implications: if we can classify one diagnostic question from discovery proteomics data, we can do this for all. The precondition is prior data for this question. With recent population based proteomics efforts this seems to be in reach in not to distant future.

Based on the overlap of this feature list and the available proteins per sample we train sample specific classifiers for each later measured sample where this decision is needed for. This makes the procedure robust and performant.

So we set out to build an ML architecture that can be employed to single sample discovery data without the need to process the sample within a study. The solution for us was a refitting procedure: we preselect features from study data for a specific task, e.g. benign from malignant classification.

After measuring a recent study and building successful classifiers with ML we asked ourselves: could we measure another patient sample now or in half a year and classify with the same performance? Admittedly, the answer is: it is complicated.