Summary
A medRxiv preprint presents IMPGAT, a method that infers cell-level flow cytometry gating from sample-level cell-frequency annotations. In leukemia datasets, its reported results were comparable to expert variability in ALL and to a supervised model in AML.
A medRxiv preprint introduces IMPGAT, a machine-learning framework that reconstructs cell-level flow cytometry gating from sample-level records. The authors evaluated it using 711 measurements from 268 people with acute lymphoblastic leukaemia (ALL) and acute myeloid leukaemia (AML), covering 240 million single cells. They report that its results were comparable to human expert variability in ALL and to a supervised model in AML.
Learning cell populations from sample records
Flow cytometry measures physical and fluorescent properties of individual cells. Analysts use these marker patterns to assign cells to populations—a process called gating—and calculate the populations’ proportions in a sample. Those proportions can contribute to clinical assessments, but archived records often preserve sample-level summaries without standardized labels identifying the population of each individual cell.
IMPGAT is designed to learn from that gap. Instead of training on explicit cell-by-cell labels, it uses existing sample-level cell-type frequency distributions: information about how much of a sample belongs to each population. The framework treats the cells in a sample as a group and uses an induced set attention mechanism to connect their marker profiles to the populations represented in the aggregate annotations. According to the authors, training requires neither manual single-cell labels nor unsupervised clustering.
This approach could make historical flow cytometry records more usable for developing automated gating logic. It does not replace the underlying measurements; it seeks to infer how cells should be grouped from information already recorded about each sample.
Results in ALL and AML
The preprint reports two comparisons. For ALL, IMPGAT’s estimates of cell-type abundance showed no statistically significant difference from human inter-expert variability. That result places the model’s quantification in relation to variation between experts; it is not a claim that every predicted cell assignment matches an expert’s decision.
For AML, the framework’s quantification of blasts—immature blood cells that can accumulate in leukaemia—was reported to be on par with a recently established supervised state-of-the-art model trained using explicit single-cell annotations. The abstract does not provide a numerical performance estimate, so the comparison is best understood as the authors’ reported overall finding rather than a quantified margin.
The authors also found that disagreements with expert gating mainly involved events experts had labelled as non-cellular. They link this pattern to the model’s lack of explicit physical prior knowledge, such as geometric cues traditionally used in scatter plots to identify debris or doublets. They report that model certainty was associated with phenotypically borderline cases, where drawing a manual boundary can itself be difficult.
The work is a preprint and its reported evaluations concern ALL and AML cohorts. The authors describe applications such as immune monitoring and drug screening as broader potential uses. The AML measurements and annotations are publicly available through a cited Zenodo dataset; the ALL dataset is restricted under data-protection rules and may be shared subject to institutional and ethical approvals.