FlowLOT: Linearized Optimal Transport for Flow Cytometry Analysis
Naqib Sad Pathan, Mohammad Shifat-E-Rabbi, Kristofor E. Pas, Ivan Medri, Bartek Rajwa, Gustavo K. Rohde
q-bio.QM
Quick verdict
FlowLOT embeds flow-cytometry samples via linearized OT for few-shot AML classification and ungated MRD estimation; promising, but MRD validation is small.
Summary
FlowLOT represents each flow cytometry sample as an empirical probability distribution over cells and embeds it into a fixed-length vector via linearized optimal transport (LOT) against a per-tube reference patient sample. Within this linear space the paper unifies three tasks: few-shot AML classification using a nearest-subspace classifier, class-difference visualization by inverting embedding-space paths back to marker space, and distribution-to-scalar regression (PLS) for measurable residual disease (MRD). On FlowCAP-II (359 patients) and BLAST110 (110 patients), LOT-based classifiers match or exceed deep baselines (AttentionMIL, CytoSet, DGCNN, PointNet++, CellCNN, FlowSOM), reaching 94.3±1.9% balanced accuracy with 16 patients/class and 98.0±1.7% with 8/class, without preprocessing, gating, or GPU training. PLS on Tube-1 embeddings correlates with blast/WBC percentage at r=0.816 (BLAST110, 5-fold CV) and r=0.785 under cross-cohort transfer to LAIP29 (n=35); at the 0.1% LAIP threshold the model reports 72% sensitivity at 100% specificity. Main limitations: FlowCAP-II rows in Table 1 are numerically identical across all three cell-subsampling depths (likely a table error), tube aggregation is described inconsistently (majority voting vs. score fusion), the MRD operating point is tuned and evaluated on the same 35-sample cohort, no significance tests or CIs accompany the ~60 baseline comparisons, and sensitivity to the reference-sample choice is unexamined. Code and data are public.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An LOT embedding relative to a single reference sample converts unordered flow cytometry samples into fixed-length vectors, enabling closed-form classification, inversion-based visualization, and regression within one framework.
- In few-shot settings, simple classifiers on LOT embeddings match or exceed deep MIL/point-cloud baselines: 94.3±1.9% balanced accuracy on FlowCAP-II with 16 patients/class and 98.0±1.7% on BLAST110 with 8/class.
- The embedding is invertible, exposing marker-level shifts (reduced SSC, increased FSC, dimmer CD5, increased CD34/CD117/HLA-DR) along the class-discriminative axis consistent with AML blast immunophenotype.
- PLS regression on Tube-1 LOT embeddings quantifies blast/WBC burden (r=0.816 in-cohort; 0.785 cross-cohort) and detects 0.1% LAIP positivity with 72% sensitivity at 100% specificity, without gating or predefined phenotypes.
- The framework is intrinsically robust to axis-aligned scaling/translation artifacts, reducing the need for normalization — explicitly excluding spectral unmixing effects.
Strengths
Unified distribution-level framework with genuine interpretability
The same LOT embedding supports classification (Table 1), inversion-based visualization of marker shifts along the class-discriminative axis (Fig. 3: reduced SSC, increased FSC, dimmer CD5, increased CD34/CD117/HLA-DR, consistent with immature myeloid phenotype), and PLS-based MRD regression — one representation, three tasks, no per-task feature engineering.
Strong few-shot accuracy with a parameter-free classifier
LOT+NSC reaches 94.3±1.9% (FlowCAP-II, k=16) and 98.0±1.7% (BLAST110, k=8) balanced accuracy, competitive with six deep baselines, using closed-form SVD class subspaces with no iterative training, hyperparameter tuning, or GPU (Results, 'Classification Performance').
Unusually careful scoping of robustness claims
Cross-cohort MRD transfer without gating
Weaknesses
Table 1 internal integrity problems
Contradictory aggregation protocols
Small, optimistically tuned MRD evaluation
No inferential statistics across many comparisons
Unexamined reference sensitivity and possible confounding
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is FlowLOT and how does it work?
FlowLOT is a method for analyzing flow cytometry data. Each patient sample (thousands of cells measured over several markers) is treated as a probability distribution. A chosen reference sample is mathematically 'transported' onto each sample via optimal transport, producing a fixed-length vector that preserves the distribution's geometry. Because all samples are embedded in the same linear space, standard tools — classifiers, regression, visualization — apply directly, and results can be mapped back to the original marker space for interpretation.
How accurately does FlowLOT detect acute myeloid leukemia (AML)?
In few-shot experiments, FlowLOT with a nearest-subspace classifier reached 94.3% balanced accuracy on the FlowCAP-II benchmark using 16 labeled patients per class, and 98.0% on the BLAST110 cohort using 8 per class. This was achieved without manual gating, data preprocessing, or GPU training, and matched or exceeded deep-learning baselines such as AttentionMIL, CytoSet, and DGCNN.
Can FlowLOT quantify measurable residual disease (MRD) without manual gating?
The authors apply PLS regression directly to the embeddings to estimate blast percentage, obtaining a correlation of about 0.82 with measured burden on held-out BLAST110 samples and 0.79 when transferring to an independent cohort (LAIP29). At the clinically used 0.1% LAIP threshold, they report 72% sensitivity and 100% specificity — but the threshold was tuned on the same small 35-patient cohort used for evaluation, so independent validation is still needed.
Is FlowLOT code and data publicly available?
Yes. Code is on GitHub (naqibUVa/FlowLOT_v1). The BLAST110 and LAIP29 datasets are on Zenodo (record 11046402), and the FlowCAP-II AML dataset is on FlowRepository (FR-FCM-ZZYA). The method required no additional preprocessing beyond what these public datasets provide.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 could not be inspected from the extracted text; quantitative judgments rely on captions, body text, and Table 1 values.
- Table 1 shows FlowCAP-II rows identical across N=500/1000/5000; this may be a source-table error rather than an extraction artifact, but the true per-N results are unknown.
- Confidence in the exact FlowCAP-II aggregation protocol is reduced by the contradictory majority-voting/score-fusion descriptions.
- Review is based on an arXiv v1 preprint; peer-review status and any supplementary material beyond the mathematical appendix are unknown.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.