diffcyt: Decide Whether You Are Testing Cell Abundance or Cell State

The original diffcyt framework offers two discovery questions. Choosing one explicitly makes a high-dimensional cytometry analysis easier to interpret and audit.

Two different questions can share a dataset

Imagine a methods meeting in which one person asks whether a cell population becomes more frequent and another asks whether its signalling changes. Both may point to the same cytometry files, yet they are asking different questions. A useful analysis plan states the target comparison before choosing a striking visualization.

What diffcyt contributes

Weber, Nowicka, Soneson and Robinson's 2019 Communications Biology paper combines high-resolution clustering with statistical tests for differential abundance and differential state. By default, cell-type markers define clusters; cluster counts support abundance tests, while median transformed cell-state marker values support state tests. The framework accommodates design variables such as paired samples and batch.

The authors evaluate semi-simulated benchmarks and reanalyse experimental datasets. They report limitations: strong batch effects may split similar cells; median summaries can hide distribution detail; population annotation remains partly subjective. An anti-PD-1 dataset reanalysis was sensitive to clustering seed. They frame the method as discovery requiring targeted confirmation, not a diagnostic classifier.

Original CFCM conceptual review checklist: Name the question; Specify the feature; Plan confirmation. No experimental plots or patient data. — Original CFCM conceptual diagram. Educational synthesis; not clinical validation.

Make the question visible in the handover

In a local analysis proposal, use one sentence to name the biological comparison and another to name the measured feature. For example, keep the question about population frequency separate from the question about a marker within that population. Avoid describing both with the single word 'activation' unless the report defines precisely what was compared.

Then ask a colleague to reconstruct the design from the proposed output tables. Can they identify the sample, the comparison, the feature and the multiplicity adjustment? Are exploratory choices recorded? A complete handover should preserve the reasoning behind a result as well as the result itself.

Plan a result that needs follow-up

Before running a discovery analysis, reserve space in the report for uncertain labels, sensitivity analyses and hypotheses requiring confirmation. A ranked cluster is a starting point for interpretation. For clinical use, establish an appropriate validation plan rather than treating publication of a computational framework as validation of your own pipeline.

Daily Learning Pill

Write 'more cells' and 'different state within cells' on separate lines. If the team cannot choose which line a result belongs to, revisit the analysis question.

Discuss your analysis plan using the CFCM cytometry resources. This original educational interpretation was prepared by ChatGPT/Codex, not an independent scientific reviewer, and provides no clinical validation.