Superior precision of clinical predictions after CD3-relativisation to align flow cytometry data

Flow cytometry captures subtle changes in immune cell distributions caused by disease, but its full potential in medical decision-making is presently limited by technical variability across instruments, sites and time. To accelerate development of generalisable diagnostic, prognostic or predictive clinical tests, we assembled a benchmark flow cytometry dataset over 20 months using 6 cytometers at 4 independent laboratories in Spain and Germany. Cohorts were amalgamated using a new alignment strategy, CD3-relativisation.

Four hundred and eighty-two clinical flow cytometry samples from 381 healthy donors and a further 100 samples from post-surgical patients admitted to intensive care were stained with a 10-colour T cell marker panel. To align data from different cohorts, we introduced CD3-relativisation, a method for normalising fluorescence intensities per-channel against CD3 signals. The quality of data alignment was evaluated using optimal transport distances (OTD), clustering consistency and predictive performance.

Our fully annotated data resource (Zenodo 17094078) revealed systematic biases in flow cytometry measurements across time, locations and cytometers. CD3-relativisation minimised these biases without sacrificing biological information. Cell clustering performance and sample-to-sample variability improved after relativisation. Consequently, we were able to predict CMV-IgG serostatus, age and sex with superior precision without relying upon external calibrators, measurement of paired samples, batch definitions or data sharing. Models established in healthy control populations were transferable to a cohort of critically unwell, post-surgical patients.

Our CD3-relativised benchmark dataset establishes a robust standard for evaluating computational methods in clinical cytometry, especially their stability over time and generalisability between instruments, laboratories and clinically heterogeneous populations.

This work was supported by the BMS-Foundation (FA-19-009), BZKF (BF/04/R/Hutch), EU-H2020 (Immutol_101080562, exTra_101119855, PAVE_861190), BMBF (01KD2206I) and DFG (403161218).