PaperlayerView exact revision on arXiv

Source-linked AI summary

Interpretable Multi-Instance Learning Enables Early Prediction of Key Molecular Alterations from Routine Flow Cytometry in Acute Myeloid Leukemia

Jonathan Legrand, Aguirre Mimoun, Baudouin Denis de Senneville, Audrey Bidet, Pierre-Yves Dumas, Christèle Etchegaray

arXiv:2609.18825v1cs.LG

TL;DR

Molecular testing for AML mutations can return weeks after treatment decisions, although routine flow cytometry may carry predictive signal. The paper applies interpretable decision-tree MIL to cell-level cytometry data; on FLT3-ITD, cytometry-based classifiers generally outperformed the clinical random-forest baseline.

  • Problem

    Molecular test results for actionable AML alterations may arrive weeks after initial management, while routine flow cytometry has been reported to contain mutation-correlated cell-expression profiles.

    Source: Introduction · p. 3

  • Method

    The paper models each patient sample as a cell bag and uses decision-tree MIL to infer mutation status from cell-level predictions, benchmarking against clinical-variable random forests and a multitube CNN.

    Source: Introduction · pp. 4–5, Proposed MIL model · p. 8

  • Results

    For FLT3-ITD, cytometry-based classifiers except the cytometry-only CNN significantly outperformed the random-forest baseline, with no significant pairwise difference among cytometry-based classifiers.

    Source: FLT3-ITD mutation detection · p. 17

  • Takeaways & Limitations

    The authors present routine flow cytometry as a basis for predicting NPM1 and FLT3 status and report that cell-level model strategies align with established hematological knowledge.

    Source: Introduction · pp. 4–5

Abstract

from arXiv · show

Background: Molecular testing for NPM1 and FLT3-ITD mutations guides critical early treatment decisions in acute myeloid leukemia (AML), but results can take weeks, long after these decisions must be made. Flow cytometry, already performed within hours of admission as part of routine care, may carry enough signal to predict these mutations directly, without added cost or delay. Methods: We developed an interpretable multi-instance learning classifier based on a decision tree, in which each patient sample is modeled as a collection of individual cells and mutation status is inferred from cell-level predictions. The model was benchmarked against a random forest trained on clinical variables and a deep convolutional neural network adapted for multitube flow cytometry data. Performance was assessed by cross-validation on a discovery cohort of 197 patients and tested on an independent cohort of 161 patients, using the area under the receiver operating characteristic curve (AUROC) and positive predictive value. Results: In cross-validation on the discovery cohort, the MIL model achieved mean AUROCs of 0.96 (SD=0.05) for NPM1 and 0.86 (SD=0.10) for FLT3-ITD, outperforming the clinical baseline and matching deep learning approaches. The model then successfully generalized to the independent test cohort of 161 patients, reaching AUROCs of 0.90 (NPM1) and 0.82 (FLT3-ITD), with positive predictive values of 0.87 and 0.68, respectively. Cell-level interpretation recovered established immunophenotypic signatures (CD33${}^{+}$ /CD34___ for NPM1-mutated cases, CD33${}^{+}$ /low side-scatter for FLT3-ITD), directly linking model predictions to known biology. Conclusions: These results show that an interpretable model applied to data already collected in routine care can predict AML molecular status within hours, offering a practical route to earlier, biology-informed treatment decisions.

Authors

Molecular testing for AML alterations can be costly and delayed, while routine flow cytometry may support earlier prediction. The paper presents an interpretable cell-level MIL approach and evaluates it on an independent cohort.

Authors

  • The listed authors are affiliated with Bordeaux university, research, and hospital institutions.Source: Authors · p. 1

Introduction

  • Routine flow cytometry may predict NPM1 and FLT3 status without added cost and within the first days of care, addressing molecular-test delays.Molecular results may arrive weeks after initial management, when urgent treatment decisions are already needed.Source: Introduction · p. 3
  • The paper's stated contribution is reliable prediction of NPM1 and FLT3 mutation status from routine clinical flow cytometry assays.Source: Introduction · p. 4
  • The decision-tree MIL framework models samples as bags of cells and infers patient-level labels from cell-level predictions, benchmarked against clinical-variable random forests and a multitube deep CNN.The comparison addresses trade-offs among model complexity, interpretability, and performance.Source: Introduction · pp. 4–5
  • The work evaluates its MIL framework on an independent dataset to assess generalizability and operational utility in clinical settings.Source: Introduction · p. 5
  • The authors report that the model's decision strategies align with hematological knowledge and identify cell-level features associated with FLT3-ITD and NPM1 predictions.Source: Introduction · p. 5

Patient cohorts

The study assembled exploration and independent test cohorts with matched multitu​​be cytometry setups. It describes sample preparation, flow-data processing, and clinical-variable handling.

Exploration and validation datasets

  • The exploration cohort included 193 patients for NPM1 and 197 for FLT3, and the independent DATAML test cohort comprised 161 patients.The exploration cohort came from the ICAML study at Bordeaux University Hospital; test-cohort samples had matching tubes, antibodies, and fluorochromes.Source: Exploration and validation datasets · p. 6

Bone marrow sampling and preparation

Instrument configuration

  • Flow cytometry used two Navios instruments configured for harmonized 10-color, 12-detector acquisition, with daily performance checks and manually established compensation matrices.Source: Instrument configuration · pp. 6–7

Flow cytometry data pre-processing

  • FCS preprocessing removed margin events, corrected fluorescence spillover with file-specific compensation matrices, and applied an arcsinh(x/5) transform; quality-control plots were manually inspected.Source: Flow cytometry data pre-processing · p. 7
  • For each patient, uniformly sampled cells per tube were standardized using training-set means and standard deviations computed per tube and marker.The two cell-sampling sizes were K=100 and K=10 000; standardization statistics were computed within each training fold or on the full exploration set for the final model.Source: Flow cytometry data pre-processing · p. 7

Clinical data pre-processing

Classification models for FLT3-ITD and NPM1 mutation detection

The proposed MIL approach treats each patient's cells as a labeled bag and trains interpretable decision trees at the cell level. Cell predictions are pooled into patient scores, with an optional clinical-data aggregator.

Proposed MIL model

  • Each patient is represented as a bag of cells labeled by patient mutation status, and interpretable decision trees predict bag labels from individual cells.The paper describes this as single-instance learning within the MIL literature.Source: Proposed MIL model · p. 8

Model training

  • Training assigns each cell its patient's mutation label, combines cells from training samples, and fits a separate tree-based classifier for each tube.Source: Model training · p. 8

Inference and score aggregation

  • At inference, mean pooling aggregates each tube's K cell scores, then average pooling combines the three tube scores into a patient-level score.The supplied passage ends before specifying the final expression for this aggregation.Source: Inference and score aggregation · p. 8

Integration of clinical data

  • Clinical features can be integrated by feeding them and tube-level scores into a logistic-regression aggregator rather than directly pooling tube scores.The aggregator is trained on tube scores generated by stratified internal 5-fold cross-validation to prevent leakage.Source: Integration of clinical data · p. 8
  • Cell-level interpretation traces the decision-tree path associated with mutation detection and maps cell prediction scores onto corresponding flow-cytometry biplots.This visualizes the cellular phenotypes contributing most to mutation detection.Source: Integration of clinical data · p. 9

C. Cell-level interpretation.

The paper describes how decision-tree paths support cell-level interpretation, how model settings were tuned, and how clinical and deep-CNN alternatives were configured.

C. Cell-level interpretation.

  • Interpretation identifies the decision path associated with the highest mutation probability and maps individual-cell scores to color intensity on biplots.Red rectangles mark gating regions whose boundaries come from learned marker split thresholds.Source: C. Cell-level interpretation · p. 10

Hyper-parameter tuning on the exploration cohort

  • A stratified 5-fold grid search compared tree depth, minimum leaf size, pruning strength, and tube-pooling functions using mean AUROC.Tuning used K=100 cells per tube to reduce computation time; the tested depths were 3 or 4, and leaf sizes were 1, 100, or 500.Source: Hyper-parameter tuning on the… · p. 10

Alternative approaches

Deep CNN for flow cytometry data

  • The CNN used depth-wise separable convolutions for multitube data, retained the original optimal hyperparameters, and accepted 10 000 cells.For cytometry-plus-clinical experiments, clinical features were concatenated with the convolution-derived embedding before the fully connected layers.Source: Deep CNN for flow cytometry data · p. 11
  • The CNN processes each cell with tube-wise 1D convolutions, averages cell representations into a patient embedding, and optionally concatenates tabular clinical data before fully connected layers.Its input is structured as (C, K, D): tubes, cells per tube, and features per cell.Source: Deep CNN for flow cytometry data · p. 11

Experimental evaluation protocol

The evaluation first compared candidate models under default settings, then optimized the MIL model and assessed it on an independent test set. Cross-validation used stratified folds, while external-test metrics were estimated across repeated random cell-sampling trials.

Experimental evaluation protocol

  • Candidate models were first compared with default settings on the exploration dataset, then the MIL model was optimized and evaluated on an independent test set.The two-step protocol assessed relative performance under comparable conditions before testing generalization.Source: Experimental evaluation protocol · p. 12

Training and inference via cross-validation

  • Stratified 10-fold cross-validation used 10,000 cells per sample for cytometry-based models; mean AUROC and standard deviation were reported across folds.Class importance was weighted by inverse class frequencies for the CNN and clinical baseline; MIL used mean aggregation at cell and tube levels and min_samples_leaf=1,000.Source: Training and inference via… · p. 12

Statistical comparison of classifiers

  • Pairwise permutation tests compared classifiers without assuming normality or independence between cross-validation folds.The tests addressed conditions not inherently guaranteed in cross-validation.Source: Statistical comparison of classifiers · p. 12

Inference on external test set

  • External-test AUROC, PPV, and NPV were averaged over 100 independently seeded cell-sampling trials, with evaluation repeated separately for tubes A, B, and C.The tube-wise analyses assessed retained discriminative power and transfer of learned biological patterns.Source: Inference on external test set · p. 13

Hardware and implementation

Flow cytometry preprocessing and model implementation used specified R and Python packages, with deep-CNN training on a local GPU. The exploration-cohort table reports demographic and clinical features and identifies the statistical tests used for group comparisons.

Hardware and implementation

  • Flow cytometry data were preprocessed in R with flowCore and PeacoQC, while Python implementations used PyTorch for the deep CNN and scikit-learn for random forest classification, cross-validation, and threshold optimization.The scripts used for the analysis are available in a public repository.Source: Hardware and implementation · p. 13
  • The deep CNN was trained on local GPU hardware; other computations used local community hardware.Source: Hardware and implementation · p. 13

Description of the exploration cohort

Comparison of classifiers performances

Across cross-validation, cytometry-based classifiers achieved high performance for NPM1 and FLT3-ITD detection, with MIL among the strongest models. Cell-level interpretations identified shared and tube-specific phenotypic patterns, while the discussion notes challenges from variability and AML heterogeneity.

NPM1 mutation detection

  • NPM1: MIL achieved mean AUROCs of 0.96 (SD = 0.05) with cytometry alone and 0.95 (SD = 0.06) with clinical variables, versus baseline 0.62 (SD = 0.10).All cytometry-based classifiers significantly outperformed baseline (P < 0.05, ΔAUROC > 0.33), with no significant pairwise differences among them.Source: NPM1 mutation detection · p. 15, Figure 4 · p. 16

FLT3-ITD mutation detection

  • FLT3-ITD: MIL achieved mean AUROCs of 0.86 (SD = 0.10) with cytometry alone and 0.87 (SD = 0.10) with clinical variables, versus baseline 0.70 (SD = 0.10).All cytometry-based classifiers except the cytometry-only CNN significantly outperformed baseline (P<0.05, ΔAUROC > 0.14); pairwise differences among cytometry-based classifiers were not significant.Source: FLT3-ITD mutation detection · p. 15, Figure 4 · p. 16

Hyper-parameter tuning and model performance

Cell-level interpretation on the test dataset

Discussion

  • Molecular status informs AML risk stratification and treatment decisions, but technical variability and biological heterogeneity make robust genotype–phenotype associations difficult.The discussion describes flow-cytometry inference as a potential complementary approach to NGS for faster, more accessible characterization.Source: Discussion · p. 22

Comparison of classifiers and methodological considerations

The tree-based MIL model recovered clinically relevant molecular information from routine flow cytometry and performed comparably to the deep CNN on the two mutation tasks. Adding clinical metadata did not significantly improve prediction performance.

Comparison of classifiers and methodological considerations

  • The proposed tree-based MIL model demonstrated performance comparable to the existing deep CNN across the two tasks implemented in this study, using cell-level classification followed by sample-level pooling.Source: Comparison of classifiers and… · p. 23
  • The proposed tree-based MIL model achieved performance comparable to the existing deep CNN across the two tasks implemented in the study, despite using cell-level classification followed by sample-level pooling.Source: Comparison of classifiers and… · p. 23
  • The tree-based model and deep CNN both classify cells before pooling predictions at the sample level, but transfer of the comparable performance to other mutation tasks remains untested.Rigorous comparison with prior studies is limited when they use different datasets.Source: Comparison of classifiers and… · p. 23
  • Adding clinical metadata, including bone marrow blast percentage, did not significantly improve performance in any prediction task.The authors note that myelogram results could add workflow delays and complexity.Source: Comparison of classifiers and… · p. 23

Interpretability and biological plausibility of tree-based MIL models

The tree-based MIL model exposes cell-level decision rules that recover known mutation-associated phenotypes and distinguish subpopulations within mutation classes. Its explicit rules are immediately interpretable, unlike the posthoc procedure described for the deep CNN.

Interpretability and biological plausibility of tree-based MIL models

  • The tree’s explicit decision rules directly inform mutation probabilities, although the learned tree may vary with the particular cell-sampling realization.Each sampling realization may introduce new relevant information.Source: Interpretability and biological… · p. 23
  • NPM1-mutated cases showed a CD33 high/CD34 low profile, with tube-specific refinements involving CD14, CD64, CD36, CD65, CD38, and CD2.In tube A, the CD33⁺/CD34⁻ population split into monocytic and non-monocytic branches, consistent with distinct differentiation subtypes.Source: Interpretability and biological… · p. 24
  • The model also identified a potential NPM1–CD36 association, which the authors suggest may reflect the frequent monocytic subtype of NPM1-mutated leukemia.Monocytes commonly express CD36.Source: Interpretability and biological… · p. 24
  • For FLT3-ITD, the model identified high CD33 and low side scatter, while the CD34⁻ association disappeared after excluding NPM1-mutated patients.This supports the authors’ hypothesis that the initial CD34⁻ association reflected mutation co-occurrence rather than a direct FLT3-ITD effect alone.Source: Interpretability and biological… · p. 24
  • The proposed tree classifier is immediately interpretable, whereas the cited CNN interpretation relies on computationally intensive posthoc up-sampling and has uncertain reliability in hematological malignancies.The posthoc method requires a forward pass for each iterative cell up-sampling step.Source: Interpretability and biological… · p. 25

Towards prediction in operational setting

The study reports mutation-specific predictive values, identifies a limited set of markers for a possible compact panel, and demonstrates generalization to routine clinical testing. Cross-laboratory transfer and the surface-marker-only scope remain practical boundaries.

Performance and PPV

  • The decision threshold was optimized using Fβ with β = 0.5, weighting precision twice as heavily as recall to penalize false positives more given their clinical implications.Source: Performance and PPV · p. 25
  • 0.87 PPV for NPM1 characterization (all tubes combined) contrasted with FLT3-ITD PPV up to 0.68 using tube A, which limits stand-alone confirmation for FLT3-ITD.The authors suggest NPM1’s relatively homogeneous phenotype may contribute to its higher PPV.Source: Performance and PPV · p. 25
  • NPV reached 0.87 for NPM1 (tube B) and 0.92 for FLT3 (all tubes combined), suggesting potential use as rapid rule-out tools.The passage connects mutation exclusion with decisions about intensive chemotherapy for some poor-prognosis alterations.Source: Performance and PPV · p. 25

Designing evidence-based panels for molecular characterization

  • Eleven markers—CD33, CD34, CD14, CD65, CD38, CD2, CD36, CD64, CD45, CD123, and side scatter—were sufficient to identify cell populations associated with the highest mutation probabilities.The authors suggest a dedicated compact panel could simplify workflows, reduce costs, and facilitate broader implementation.Source: Designing evidence-based panels for… · p. 26

Model generalization across cohorts

  • Despite differences between the research exploration cohort and fully routine test cohort, the tree-based MIL model generalized to the testing cohort.The routine cohort had inherent variability and operational constraints, while the research cohort had greater standardization and more cells per sample.Source: Model generalization across cohorts · p. 26
  • Cross-laboratory transfer remains challenging because laboratories differ in fluorochromes and sample preparation protocols.The authors call for evaluation across distinct protocols and domain-adaptation strategies.Source: Model generalization across cohorts · p. 26
  • The study used conventional panels with 9 to 10 surface markers per tube and no intracellular staining, leaving intracellular signaling and metabolic features unmeasured.The authors suggest higher-dimensional spectral cytometry could improve prediction, particularly for FLT3-ITD and unexplored mutations.Source: Model generalization across cohorts · pp. 26–27
  • The observed phenotypically distinct NPM1-predicted subpopulations raise the possibility that sufficiently large, well-annotated datasets could eventually resolve mutational subclones within patient samples.The authors motivate further multicentric-cohort and single-cell-genotyping research.Source: Model generalization across cohorts · p. 27

Authors contribution

The contribution statement assigns study and analysis responsibilities; the authors report no competing financial interests. The datasets are requestable but not publicly available, and the study reports ethics approvals and patient consent.

Authors contribution

  • CE, PYD, and BDS contributed to conceptualization and reviewing/editing; CE and BDS also handled funding acquisition and supervision.AM contributed to conceptualization, data curation, investigation, and reviewing/editing.Source: Authors contribution · p. 28

Conflict of interest

Data availability statement

  • The analyzed datasets are not publicly available for privacy or ethical reasons but may be obtained from the corresponding author on reasonable request.Source: Data availability statement · p. 28

Ethics approval and consent to participate

  • The study followed the Declaration of Helsinki, received the stated French regulatory approvals for the validation cohort, and obtained informed consent from included patients.The exploration cohort was collected as part of ICAML study NCT03449745.Source: Ethics approval and consent to… · p. 29
Loading 2609.18825v1…