Interpretable Multi-Instance Learning Enables Early Prediction of Key Molecular Alterations from Routine Flow Cytometry in Acute Myeloid Leukemia
Jonathan Legrand (IMB, Aguirre Mimoun (CHU Bordeaux), Baudouin Denis de Senneville (IMB, MONC), Audrey Bidet (CHU Bordeaux), Pierre-Yves Dumas (CHU Bordeaux, Inserm U1312 - BRIC), Christ\`ele Etchegaray (MONC, et al.
cs.LG
Quick verdict
MIL predicts NPM1/FLT3-ITD from routine AML flow cytometry (AUROC 0.90/0.82) via interpretable biology-consistent rules; single-center study, modest FLT3 PPV.
Summary
The paper proposes a decision-tree-based multi-instance learning (MIL) pipeline that predicts NPM1 mutation and FLT3-ITD status directly from routine multi-tube flow cytometry in acute myeloid leukemia (AML). Each patient is a bag of cells; a per-tube decision tree is trained on cell-level labels inherited from the patient's mutation status, and cell-level scores are mean-pooled to tube and then patient level. On a 197-patient exploration cohort (ICAML, Bordeaux), 10-fold CV AUROCs were 0.96 (NPM1) and 0.86 (FLT3-ITD), significantly exceeding a clinical-variable random forest and statistically indistinguishable from a CNN adapted from the cytometry literature. An independent 161-patient cohort (DATAML, 2013-2018) yielded AUROCs 0.90/0.82 with PPVs 0.87/0.68. Learned decision rules recovered established immunophenotypes (CD33-high/CD34-low for NPM1; CD33-high/low side scatter for FLT3-ITD), and a control retraining excluding NPM1-mutated patients showed the CD34-low FLT3 signal reflects mutation co-occurrence. Limitations: both cohorts come from a single center with matched instruments and panels, so cross-laboratory transfer is untested; the all-tubes FLT3-ITD PPV is only 0.55, below the best single tube quoted in the abstract; test-cohort prevalence, confidence intervals, and calibration are not reported in the main text; the CNN baseline retained its original hyperparameters after architectural adaptation. Code is public; data available on request.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- A decision-tree MIL pipeline predicts NPM1 and FLT3-ITD status from routine flow cytometry alone within hours, achieving external test AUROCs of 0.90 and 0.82.
- The interpretable tree-based model performs on par with an adapted deep CNN while providing immediate rule-based interpretation.
- Cell-level decision rules recover established immunophenotypic signatures (CD33-high/CD34-low for NPM1; CD33-high/low side scatter for FLT3-ITD), grounding predictions in known biology.
- Adding clinical variables (age, sex, blast percentage) does not significantly improve any prediction task.
- High NPVs (0.87 for NPM1, 0.92 for FLT3-ITD) support use as a rapid rule-out tool, although FLT3-ITD PPV (0.55-0.68) precludes stand-alone confirmation.
Strengths
Genuinely independent external validation
The 161-patient DATAML test cohort was collected 2013-2018, years before and under fully routine conditions unlike the research-standardized ICAML exploration set; performance is reported per tube and over 100 cell-resampling trials (Table 2), not just cross-validation.
Interpretation is tied to published biology, not merely post-hoc
Learned decision paths reproduce CD33-high/CD34-low for NPM1 (consistent with refs 10, 13) and CD33-high/low side scatter for FLT3-ITD (ref 11); the novel CD36 association is given a plausible monocytic-differentiation explanation (refs 13, 24) rather than overclaimed.
Explicit confounder check for mutation co-occurrence
Honest comparative statistics and negative results
Actionable clinical framing
Weaknesses
Single-center data limit the generalizability claim
FLT3-ITD operational precision is weak and the headline number is the best tube
Statistical reporting gaps
Baselines may be under-tuned in both directions
Interpretation robustness is unquantified
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 5 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
Can flow cytometry predict NPM1 or FLT3-ITD mutations in AML without waiting for genetic testing?
This study reports that a machine-learning model applied to routine multi-tube flow cytometry predicted mutation status with external test AUROCs of 0.90 (NPM1) and 0.82 (FLT3-ITD) in 161 patients, using data already collected within hours of admission. It is positioned as an early triage aid: positive FLT3-ITD predictions (PPV 0.55-0.68) would still require confirmatory molecular testing.
Which cell surface markers indicate NPM1-mutated AML on flow cytometry?
According to the learned model, the core signature is high CD33 with low CD34, refined per tube by CD14, CD64, CD36, CD65, CD38, CD2 and CD45. This matches previously published NPM1 immunophenotypes and suggests a compact dedicated panel of about 11 markers (also including CD33, CD34 and side scatter for FLT3-ITD) could suffice.
How accurate is this model and is it ready for clinical use?
External AUROCs were 0.90 for NPM1 and 0.82 for FLT3-ITD. The positive predictive value reached 0.87 for NPM1 and the negative predictive value 0.92 for FLT3-ITD, supporting rapid rule-out use. However, FLT3-ITD PPV was only 0.55-0.68 and validation was single-center, so it is not yet a stand-alone diagnostic test.
Is the code or data for this study publicly available?
The analysis code is public on GitHub (github.com/jonathan-legrand/cytometry-molecular-characterization). Patient data cannot be shared publicly for privacy and ethical reasons but is stated to be available from the corresponding author on reasonable request.
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 (decision trees, cell-level bi-plots, performance box plots) could not be analyzed; assessment relies on captions and in-text descriptions.
- Supplementary material (Tables S2/S4/S5/S8, Figures S1/S6-S9, Notes S1/S3) was not in the extracted text; test-cohort prevalence and tuning details are unverified.
- All quantitative claims were taken from the extracted text; the public GitHub repository was not inspected.
- This is an arXiv preprint (2609.18825); peer-review status and target journal 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.