Deep learning based individualized cross-platform molecular subtype classification of B-lineage acute lymphoblastic leukemia Authors: Bowen Cui, Huiying Sun, Shuang Zhao, Rongrong Fan, Jianan Rao, Wenyan Wu, Ying Zhong, Ronghua Wang, Ying Wang, Qiaoqiao Shi, Yuxuan Guo, Jiale Lin, Yuanlu Huang, Yuxuan Han, Hui Liu, Xiaolong Chen, Shuhong Shen, Han Wang, Yu Liu Published: 2026 ABSTRACT AbstractMolecular subtypes of B-cell acute lymphoblastic leukemia (B-ALL) are essential in modern clinical treatment. However, the fast emerging subtypes and the requirement of complex multi-omic diagnostics are continuously challenging the clinical subtypes classification in sensitivity, cost, and turnaround time. We develop B-cell Acute Lymphoblastic Leukemia Subtype Identification based on gene eXpression (BALL6), a robust deep learning framework for cross-platform B-ALL subtyping. BALL6 utilizes a recurrent neural network trained on rank-transformed expression values of feature genes, capturing subtype signals while inherently minimizing technical noise. We implement in BALL6 the rank-based augmentation framework which further enhances its performance on data-limited or imbalanced datasets. BALL6 includes two integrated models: an AL model distinguishing B-ALL, T-ALL, and AML, and a B-ALL model identifying the most updated 20 established molecular subtypes. BALL6 demonstrates robust accuracy across multiple independent datasets, achieving 99.38% (AL model) and 93.84% (B-ALL model) accuracy on previously unseen data. Notably, BALL6 is robust to missing values, which enables its cross-platform application. This is demonstrated through reliable subtype prediction by BALL6 with sparse gene expression profiles from scRNA-seq data. BALL6 is open-sourced with an accessible web tool (https://cccg.ronglian.com/#/analysis), facilitating its broad applications in leukemia research and clinical diagnostics.Subject terms: Cancer, Computational biology and bioinformatics, Oncology Introduction Acute leukemia (AL) is a hematologic malignancy of critical clinical importance, particularly as a leading cause of cancer-related death in children1. AL is broadly classified into three major lineages based on cell origin: B-cell ALL (B-ALL), T-cell ALL (T-ALL), and acute myeloid leukemia (AML). Among these, B-ALL is the most common type of pediatric leukemia and represents a highly heterogeneous disease with significant variability in clinical behavior and treatment response2–5. Current clinical diagnostics and stratification of leukemia primarily rely on the integration of morphology, immunophenotype, cytogenetics, and molecular biology (MICM) framework. However, this approach faces restrictions in multiple aspects. Fluorescence in situ hybridization (FISH) panels, used for rapid detection of recurrent fusion genes, are intrinsically limited to predefined targets. For complex subtypes like Ph-like and KMT2A-rearranged B-ALL, the extraordinary diversity of fusion partners and breakpoint locations often exceeds the detection scope of conventional FISH probes6,7. Similarly, although G-band karyotyping is the widely established clinical method for detecting chromosomal abnormalities and characterizing subtypes such as hyperdiploidy (HYPER), hypodiploidy (HYPO), and intrachromosomal amplification of chromosome 21 (iAMP21)8–10, its sensitivity can be constrained by sampling limitations and resolution. Especially, the masked hypodiploid ALL, which is an ALL subtype with poor prognosis, could be misdiagnosed as the HYPER subtype with favorable prognosis by karyotyping, posing an important limitation of this test. Recent advances in next-generation sequencing (NGS) have revolutionized our understanding of B-ALL biology, enabling the recognition of over 20 distinct molecular subtypes2,11,12 with profound prognostic implications for risk stratification and therapy. The NGS based tests are also rapidly applied in molecular subtype stratification in clinic. However, this usually requires the integration of multi-omics data including whole genome sequencing (WGS), whole exome sequencing (WES), and transcriptome sequencing (RNA-seq) in these clinical tests. Although this strategy is the most comprehensive way of characterizing the genome aberration profiles of tumor, the requirement of multiple sequencing experiments inevitably compromises the turnaround time in clinic and substantially increases experiment and computing costs, thereby undermining their clinical feasibility. Meanwhile, transcriptional profiling offers an alternative approach, as gene expression signatures can encapsulate high-dimensional, subtype-specific information. Previous studies have leveraged this to develop gene expression-based subtype classifiers3,13,14. Nevertheless, two major limitations exist for current methods. Firstly, they are hypersensitive to technical variations arising from experimental platforms, gene expression quantification algorithms, and prone to the missing values in gene expression profiles which is not unusual when applying this analysis on cross-platform data. Moreover, the rapidly emerging new subtypes are not implemented in these methods, limiting their application in current real-world clinical settings. Consequently, achieving accurate, reliable, and platform-agnostic classification of the full spectrum of B-ALL subtypes remains an unmet need in clinic. To address these limitations, we turned to deep learning, which excels at learning noise-resistant representations from high-dimensional biological data and can flexibly accommodate novel subtype patterns15–18. We therefore developed BALL6 (B-cell Acute Lymphoblastic Leukemia Subtype Identification based on gene eXpression), a deep neural network-based tool with advanced data augmentation technique for acute leukemia subtyping. Our framework comprises two specialized models: 1) The AL model, trained to distinguish the three major lineages (AML, T-ALL, B-ALL), and 2) the B-ALL model, designed for precise classification of the 20 most up-to-date B-ALL subtypes (Supplementary Data 1). This includes not only the established high-risk subtypes such as Philadelphia chromosome positive (Ph), Ph-like, KMT2A-rearranged, but also subtypes particularly difficult to define by current RNA-seq based bioinformatic analysis, including HYPER, HYPO, and iAMP21, and the subtypes defined by gene dysregulation (e.g., DUX4-rearranged, ETV6::RUNX1-like, MYC or NUTM1 activation) and rare, newly reported entities (e.g., PAX5alt, IKZF1-N159Y and ZEB2/CEBP). Crucially, to enhance robustness and mitigate technical batch effects inherent to different sequencing platforms and bioinformatic pipelines, BALL6 utilizes rank-transformed gene expression values as input. This approach effectively decouples classification from absolute gene expression measurements while providing strong data-adaptability. Compared to existing classifiers such as ALLSorts14, Allspice19, and ALLCatchR20, BALL6 offers more comprehensive subtype coverage and achieves superior accuracy on unseen independent cohorts. Notably, we showed that BALL6 performed well on sparse gene expression profiles with high proportion of missing values, providing robust accurate predictions with cross platform data, even with single cell RNA-seq (scRNA-seq) data. These features make BALL6 a robust, accurate, and clinically adaptable solution for comprehensive acute leukemia subtyping. Results BALL6 model design The develop of BALL6 consisted of major components, the rank-transformation based data augmented framework (RTDAF) and a neural network model for subtype prediction from RNA-seq data (Fig. 1A–C). The RTDAF component aimed to maximize phenotype-relevant signal extraction from RNA-seq data while minimizing technical artifacts. We implemented a multi-step preprocessing pipeline in this step to build the training dataset (Fig. 1A and Supplementary Data 2, 3). Firstly, to mitigate the impact of extraneous signals prevalent in RNA-seq data, we applied a series of feature engineering techniques including statistical methods (e.g., ANOVA, differential gene expression), regression-based approaches (e.g., LASSO), and machine learning models (e.g., random forest). We also included literature-supported and manually curated genes exhibiting pronounced phenotypic traits in this step. Finally, a total of 476 genes were selected which possess adequate predictive power for molecular subtype classification (Methods and Supplementary Data 4). Subsequently, to minimize biases introduced by various sequence technical or quantification analytical procedures affecting gene expression values, we performed rank transformation on the expression values (Methods). Rank-transformed data significantly reduces batch effects while preserving biologically relevant heterogeneity between sample subtypes. (Fig. 1D and Supplementary Fig. 1). Overall architecture of BALL6. A, B Design of BALL6. Using raw gene expression as feature selection input, 476 genes were selected and performed rank-transformation before forwarded to the embedding layer and the LSTM recurrent neural network model. C BALL6 classified the input sample into three major lineages in AL model and 20 most up to date B-ALL subtypes in the B-ALL model. Prediction probability was visualized as polar plots, with the probability represented by the radius of each sector and the final predicted subtype showed in red. A B-ALL case of ETV6::RUNX1 subtype (SCMC-001749) was shown as an example. D Scatterplot shows the top two principal components from PCA of the raw expressions (left) and rank-transformed expressions (right) profiles. The top 2000 most-variable genes were included in this analysis. The ranked gene expression data were fed into the neural network model of BALL6. Data were processed through an embedding layer to learn distributed representations of each gene and effectively capture higher-dimensional latent features. This was followed by two Long Short-Term Memory (LSTM) layers, which not only processed the high-dimensional expression patterns but also explicitly leveraged the sequential nature of the ordered gene data to enhance predictive efficiency. BALL6 employed two models for different prediction tasks: the AL model and the B-ALL model. The AL model was trained on the St. Jude Cloud cohort2,13 consisting a total of 1606 ALs to classify samples among the three major lineages of acute leukemia. Meanwhile, the B-ALL model was trained on SCMC B-ALL cohort consisting a total of 1,009 B-ALLs to predict 20 well-defined subtypes, including HYPER, HYPO, iAMP21, Ph, Ph-like, KMT2A-rearranged, MEF2D-rearranged, HLF-rearranged, MYC-rearranged, PAX5 P80R, IKZF1 N159Y, DUX4-rearranged, ETV6::RUNX1/-like, PAX5 alterations, TCF3::PBX1, ZNF384-rearranged/-like, NUTM1-rearranged and ZEB2/CEBP fusion, as shown in Fig. 1C and Supplementary Data 1. Critically, the subtype annotation for each sample within the SCMC cohort underwent rigorous manual curation, integrating comprehensive clinical data, karyotyping results, and high-throughput sequencing analysis (Manuscript in preparation and Supplementary Data 1 and 3). Both of the AL model and the B-ALL model in BALL6 use the same core architecture, differing only in their dimension of final output layer. Hyperparameter optimization was performed across a defined search space to identify optimal configurations. The optimal hyperparameters were selected and used to train the final BALL6 model (Methods). Data augmentation for BALL6 training To improve the robustness of BALL6 against technical variability, we developed and implemented in BALL6 training a data augmentation technique specifically designed for rank-based gene expression data (Methods and Supplementary Fig. 2A). This approach strategically introduces controllable perturbations at both global and local scales into the rank-transformed expression profiles to simulate inherent variability in real-world data, while avoiding distortion in the newly generated data and effectively retaining subtype-specific information (Supplementary Fig. 2B, C). Critically, this approach trains BALL6 to maintain predictive accuracy against technical deviations which may be caused by differences in tumor purity, sequencing platforms, or quantification methods21,22. We applied this data augmentation technique to the two collected fully independent RNA-seq expression datasets for BALL6 training. The St. Jude Cloud cohort included 1606 patients (284 AML, 1223 B-ALL, and 99 T-ALL) while the SCMC cohort consisted 1144 patients (1009 B-ALL and 135 T-ALL). For each sample, gene expression levels were quantified using both FPKM and DESeq2 pipelines aligned to hg19 and hg38 reference genomes. Applying our data augmentation technique to these multi-source datasets, we generated simulated rank-transformed profiles, achieving an 8-fold expansion of the original data and using for BALL6 training. This yielded an augmented training set of 6424 profiles from St. Jude Cloud and 9152 profiles from SCMC for training of BALL6. BALL6 achieves enhanced subtype prediction accuracy BALL6 demonstrates exceptional predictive accuracy and robustness. Both models exceeded 99.8% accuracy on their respective training sets (Supplementary Fig. 3). On the validation set combined original and noise-perturbed data, the AL model achieved 99.45% accuracy, and the B-ALL model achieved a classification accuracy of 98.70%, as shown in Fig. 2A. Additionally, the one-vs-rest AUROC on the validation set reached ≥0.998 for the AL model and ≥0.999 for the B-ALL model (Supplementary Fig. 4A, B). BALL6 demonstrated robust performance regardless of the bioinformatic processing pipeline: samples processed using either the hg19 or hg38 reference genome (98.39% vs. 99.01%) and quantified using either DESeq2 or FPKM (98.75% vs. 98.65%) were accurately classified into the correct B-ALL subtypes. The performance of BALL6. Scatter plot with boxplot overlay showing the prediction probabilities of AL model (A) and the B-ALL model (C) of BALL6 in validation dataset. Colors represent data origin (raw expression profiles or profiles after data augmentation). Shapes stand for whether the prediction result was correct for each sample. Comparison of prediction accuracies of BALL6 with most recently published tools in AL (B) and B-ALL (D) subtype prediction tasks in the SCMC, St. Jude Cloud13, and Bld.202023 datasets. Published tools tested included ALLCatchR, ALLSorts, and Allspice. Asterisk (*) indicates that the dataset was used for training in this model. Line graphs show the robustness of BALL6 in feature ablation analysis in AL model (E) and B-ALL model (F). The accuracy of BALL6 after randomly removing a specified number of genes from the input was shown on y-axis, while the number of genes removed was labeled on y-axis. Each point in the plot is the average accuracy with 50 repetitions. Owing to the complete non-overlap between the AL and B-ALL model training sets, we were able to leverage the St. Jude Cloud B-ALL samples – entirely unseen during B-ALL model training – to evaluate subtype classification accuracy. Similarly, the SCMC cohort data was used to assess the AL model’s lineage classification performance. Remarkably, despite trained on class-imbalanced dataset (T-ALL: 6.16%, AML: 17.7% of 1606 St. Jude Cloud patients), the AL model of BALL6 achieved 99.38% accuracy in distinguishing B-ALL, T-ALL, and AML across the unexposed SCMC cohort. This performance significantly surpasses Allspice19 (94.20% accuracy), the only existing classifier capable of the same task (Fig. 2B). Furthermore, the B-ALL model attained an exceptional accuracy of 93.84% on the completely unseen St. Jude Cloud B-ALL samples (Fig. 2C, D, Supplementary Fig. 5). Despite inherent differences in subtype distribution between two cohorts, BALL6 matched or surpassed the performance of the most recently published tools ALLCatchR20 (89.86%), ALLSorts14 (93.36%), and Allspice19 (67.58%). Crucially, it is noteworthy that all these comparator tools incorporated the St. Jude cohort into their training sets to varying extents. To further evaluate the performance of BALL6, an independent dataset of 138 RNA-seq datasets (Bld.2020 cohort, 115 B-ALL and 23 T-ALL) collected from diagnosed and relapsed B-ALL patients23 were obtained for verification. BALL6 showed outstanding prediction performance and robustness when applied to this testing set, with accuracies of 98.56% and 93.04% achieved for the AL and B-ALL models, respectively, and one-vs-rest AUROC values exceeding 0.993 and 0.979, respectively (Fig. 2B, D, Supplementary Fig. 4C, D, and Supplementary Data 5). By contrast, the accuracy of the similar tools ALLCatchR, ALLSorts, and Allspice in B-ALL subtype prediction was 88.70%, 86.96%, and 66.96%, respectively (Fig. 2D). In three independent cohorts, BALL6 achieved superior or competitive subtype‑level accuracies across B‑ALL subtypes relative to three existing classifiers (Supplementary Fig. 6). Notably, BALL6 showed good performance when predicting the aneuploid subtypes. Among the cases from the validation cohort with matched RNA-seq and whole genome sequencing (WGS) data, a total of 27 samples were predicted as HYPER (hyperdiploidy, n = 24) or HYPO (hypodiploidy, n = 3). All the predictions were successfully validated by the copy number variation (CNV) results from the WGS data (Supplementary Data 5). Remarkably, karyotyping testing failed to detect the HYPER subtype in a total of 13 out of these 24 samples, four were reported as normal and 9 failed the karyotyping test. This further demonstrated the power of BALL6 in predicting the aneuploidy subtypes of B-ALL. Feature ablation analysis demonstrates the robustness of BALL6 We tested the robustness of both AL and B-ALL models in BALL6 to missing gene expression data, which was common when applying the prediction to different experimental platforms. To do this, we evaluated the prediction performance with different numbers of censored gene values mimicking the missing gene expression values. Results showed that the prediction accuracy for AL model was over 98.0% for all three disease types when there were 50 genes ablated from the total of 229 genes used in this model (Fig. 2E). Meanwhile, the B-ALL model of BALL6 was subtype dependent, as shown in Fig. 2F. Subtypes of MEF2D, TCF3::PBX1, and DUX4 exhibited exceptional robustness to sparse data. The model maintained both of precision and recall exceeding 80% even when transcription values for 100 randomly genes (representing 34% of all required genes) were omitted (Fig. 2F & Supplementary Fig. 7). On the other hand, the ZNF384, KMT2A, and HYPER subtypes were more vulnerable to the censoring of gene expressions, which can only achieved precision and recall rate above 80% under less than 30 omitted genes (10%). BALL6 predicts molecular subtype from single cell RNA-seq data BALL6 also demonstrates robust applicability to single-cell data. We evaluated its performance on 29 scRNA-seq samples from 10 B-ALL patients with different subtypes24,25 (Supplementary Fig. 8), comprising ETV6::RUNX1 (n = 14), Ph (n = 6), MEF2D (n = 4) and KMT2A rearranged (n = 5). After cell type annotation against the human bone marrow reference, we identified five B-cells clusters yielding a total of 74,212 high-confidence B cells at an average of 2559 cells per sample (Methods and Fig. 3A). We firstly applied BALL6 to pseudobulk expression profiles, derived from aggregating all B-cells within each sample. These profiles captured a median of 257 detectable genes (out of 292 gene in B-ALL model) per sample (Supplementary Fig. 9A). Strikingly, BALL6 achieved 100% subtyping accuracy across 22,015 cells from 10 diagnostic-phase samples (Fig. 3B). In relapsed samples, it successfully predicted the subtype for 7 out of 9 cases (77.78%), accounting for 14,603 out of 16,620 (87.86%) single cells, demonstrating its robustness to potential low tumor purity. Application of BALL6 on scRNA-seq data. A UMAP visualization of scRNA-seq gene expression data of B-ALLs anchored to the human bone marrow reference using Azimuth. The B cells were grouped into five clusters and labeled with different colors. B Bar plot shows the accuracy of BALL6 in molecular subtype prediction with pseudobulk data from scRNA-seq. B-ALLs were grouped as diagnosis (10 cases) and relapse (9 cases). The number of single cells was shown on y-axis. Colors indicate whether the prediction was correct. C Scatter plot showing the relationship between the number of cells (x-axis) and the number of detectable genes (expression >0, y-axis) in each cluster. Colors indicate whether the prediction was correct for this cluster of cells. The gray line marks 194 detectable genes (≈2/3% of the 292 genes required for prediction). D UMAP visualization of scRNA-seq data from a primary KMT2A-rearranged B-ALL sample. Colors represent BALL6 subtype predictions at single-cell resolution. E Paired scRNA-seq UMAP plots of diagnosis (left) and relapsed (right) KMT2A-rearranged B-ALL samples. Each dot represents a single cell. Furthermore, we computed pseudobulk expression profiles for each of five B-cell cluster in diagnostic and relapsed samples. Reduced cell counts within clusters revealed a correlation between predictive performance and the number of expressed genes. (Fig. 3C). When clusters contained >194 detectable genes (≈2/3 of all 292 genes for prediction), the B-ALL model achieved 95.43% accuracy, whereas clusters containing detectable genes below this threshold showed only 20.96% accuracy (Fig. 3C). The number of detectable genes was the dominant factor affecting prediction accuracy compared to patient’s subtype or disease phase (Supplementary Fig. 9B). This indicated that BALL6 using pseudobulk strategy can achieve accurate classification even after losing one-third of expression information, and its performance scales with the number of detected genes. Moreover, we evaluated the performance of BALL6 for subtype prediction at single cell resolution. Given that data sparsity (with a median of 23 genes detected in each single cell) in scRNA-seq could substantially impairs classification, we focused this analysis on diagnostic and relapsed samples of KMT2A- and MEF2D-rearranged subtypes, where BALL6 demonstrated strong performance in gene ablation experiments. We used a public non-leukemic/leukemic classifier25 to screen leukemic B-cell prior to BALL6 prediction. Comparing predictions with the ground-truth subtypes revealed striking results: among the three KMT2A scRNA-seq data, BALL6 yielded 90.39% overall accuracy (3197/3537 cells), with 98.76% accuracy in one diagnosis sample (Fig. 3D), 74.89% and 72.80% in a matched diagnosis-relapse pair (Fig. 3E). In a matched MEF2D diagnosis-relapse pair achieved maximum 51.18% accuracy (Supplementary Fig. 10). Collectively, these results further demonstrated BALL6’s capability of subtype prediction with low information density data. Web tool deployment We developed a user-friendly web server (https://cccg.ronglian.com/#/analysis) for applying BALL6 to user provided dataset (Fig. 4A). Users can easily obtain prediction results by uploading expression values or the rank order of gene expressions. Both AL model and B-ALL model run automatically after upload, and subtype predictions and probabilities are presented shortly in tables and interactive polar plots (Fig. 4B). Development of web utility of BALL6. A Homepage of the BALL6 web utility. Users can submit the expression values or ranks of 476 genes to obtain prediction results within a dozen seconds. B Results page of BALL6 prediction. The prediction results from both the AL model and the B-ALL model are shown in tabular (left) and radar chart formats (right). Discussion The identification of molecular abnormalities in acute lymphoblastic leukemia has greatly facilitated the implementation of precision therapy in clinical practice6,8,11,26–28. Despite the significant potential of gene expression profile data from transcriptome sequencing, challenges remain in phenotype discovery. Technical noise in sequencing data and biases introduced by diverse analytical pipelines significantly limit the broad applicability of existing models. This often results in high accuracy on technically matched samples but limited generalization to independent datasets. We present BALL6, a tool for accurate B-ALL subtype classification based on a neural network model. Utilizing the rank order of gene expression rather than continuous values for subtype prediction not only reduces computational complexity and mitigates batch effects but also greatly enhances the model’s cross-platform robustness. While data augmentation techniques are widely used in deep neural networks, their application to expression profile data is still lacking. By adding noise perturbation to rank-ordered data with the data augmentation technique, we were able to expand the training dataset, partially alleviating the impact of class imbalance on model learning. Within this Rank-Transformation-Data-Augmented framework, factors detrimental to model performance – such as batch effects and data distribution biases inherent in expression profiles – are minimized. Empirical evaluations show that BALL6 consistently achieves high and stable accuracy, regardless of whether input data are processed using the hg19 or hg38 reference genomes, or quantified via DESeq2, FPKM, or TPM pipelines. Accordingly, our framework delivers inherently platform‑agnostic performance, ensuring exceptional generalizability across distinct clinical cohorts. While RTDAF enables the model to achieve certain learning capability in classification tasks with limited sample sizes, it must be noted that the model robustness remains inevitably constrained by scarce data. Although BALL6 currently shows good performance for predicting subtypes with small sample size such as MYC and ZEB2/CEBP, larger datasets of these subtypes would further improve the performance. Meanwhile, the neural network architecture deployed in BALL6 also enhances its potential for information mining. Compared to neural networks like MLPs, the LSTM framework is more adept at capturing relational features among genes when processing rank-ordered expression, and unlike attention-based networks, it does not require large-scale training data or substantial computational costs. Ultimately, this design allows BALL6 to learn sufficient knowledge from a cohort with imbalanced subtype distribution and enables the accurate prediction of 20 up-to-date B-ALL molecular subtypes including those that were difficult to detect in clinic. Noteworthy, all results in the BALL6 training set (SCMC cohort) underwent rigorous manual review, and subtype definitions were established based on clinical risk stratification and treatment considerations, leading to more scientifically grounded classifications. This ensures that BALL6 predictions directly relevant to patient management in clinic. Importantly, we showed in the gene ablation experiments the strong robustness of BALL6 to missing values. We observed that this robustness varied across subtypes, indicating a potential different biological complexity of each subtype. Such variation may be underpinned by chromosomal instability, epigenetic plasticity, and clonal evolution29–31. Notably, BALL6 could successfully predict subtypes from sparse gene expression profile like scRNA-seq. Despite never being trained on this type of data, BALL6 was able to successfully perform subtype prediction from single-cell data, notably achieving a 95.43% accuracy when cell clusters contained over 194 detectable genes. Remarkably, taking the KMT2A- and MEF2D-rearranged lineages as examples, BALL6 accomplished reliable subtype predictions at individual cell resolution even when only 23 genes were detected per cell. This ability is intrinsically linked to its use of rank-transformed expression as model input which helps BALL6 focus on gene-gene relationships rather than absolute expression values, thereby facilitating deeper knowledge extraction from low-density data. The single-cell prediction task performed in this study serves as an extreme stress test to validate the model’s robustness against highly data sparsity and the prediction accuracy at individual cell resolution. Further extensive validations are necessary. For example, we currently only analyzed scRNA-seq data from 4 B-ALL subtypes. More comprehensive investigation will be needed in the future. Single-cell technologies enable researchers to investigate multiple aspects such as early relapse prediction, cellular differentiation trajectories, and tumor heterogeneity, thereby contributing to improved clinical outcomes31–34. The ability to predict molecular subtype at single cell level could provide additional information in scRNA-seq analysis of B-ALLs. We developed BALL6 as a free, open-source software. We also developed an accompanying web-based platform to facilitate the use of this tool. Users can rapidly obtain AL and B-ALL subtype predictions for a sample by submitting gene expression data for only 476 genes. However, several practical limitations of BALL6 must be carefully considered in real-world clinical interpretation. First, sample quality and low tumor purity can inevitably affect rank orderings of gene transcription, thereby potentially confounding the model’s predictive precision. Second, because BALL6 operates within a defined diagnostic space of 20 established subtypes, any newly emerging or unclassified rare molecular subtypes remain beyond its predictive capacity. Importantly, real-world clinical management is inherently a multi-dimensional, comprehensive decision-making process. Incorporating alternative genomic modalities, such as somatic mutations and gene fusions, is necessary for proper diagnosis and clinical decision-making. Nevertheless, as a subtyping tool dependent exclusively on RNA-seq gene expression data, BALL6 can be seamlessly integrated into existing precision diagnosis workflows. Compared to the traditional MICM framework, BALL6 offers a more comprehensive and precise identification of complex molecular subtypes. Furthermore, relative to multi-omics NGS pipelines, achieving this comprehensive subtyping through a single, transcriptomics-only approach substantially reduces sequencing costs and accelerates the turnaround time. We believe the robustness and user-friendliness of BALL6 make it a valuable tool for supporting clinical diagnostics and advancing precision medicine in acute leukemia. Methods RNA-seq data processing and ethics approval Gene expression data from the SCMC cohort were used for training and validating the BALL6 B-ALL model (Supplementary Data 2). To enhance the model’s resilience to biases arising from different bioinformatic pipelines, we processed the data using both the hg19 and hg38 human reference genomes for alignment. Gene expression was then quantified using both DESeq235 (variance-stabilizing transformation) and FPKM (Fragments Per Kilobase of transcript per Million mapped reads) analyses. This approach generated 4036 expression profiles derived from the original 1009 B-ALL patient samples. The combined dataset was randomly split into 80% for training and 20% for validation, maintaining original subtype proportions in each subset. The study was conducted in accordance with the Declaration of Helsinki. The analysis of RNA-seq data from SCMC cohort was approved by the Shanghai Children’s Medical Center Institutional Review Board (SCMCIRB-YG2024059). Informed written consents were obtained from parents for all patients. The St. Jude cohort2,13 was used for training the ALL lineage classification model (distinguishing AML, T-ALL, and B-ALL, Supplementary Data 2, 3). The B-ALL samples from this cohort were also utilized for validating the B-ALL model (which differentiates B-ALL subtypes). For B-ALL samples harboring multiple pathogenic variants, the final subtype assignment was prioritized based on the alteration associated with inferior prognosis. Next, we aggregated RNA-seq count data from 1606 patients aligned to the hg38 reference genome. Similarly, to mitigate pipeline-specific bias, 3212 gene expression profiles was quantified by performing both DESeq2 and FPKM methods. Single-cell RNA-seq data processing For each single-cell transcriptomic sequencing dataset, we obtained gene expression matrix and anchored cells to the human bone marrow reference using Azimuth36 for each cell type annotation. We subsequently isolated cells annotated as pro-B, pre-B, naïve B, transitional B, or memory B cells with cell type prediction scores >0.5, which were defined as B cells for downstream analysis. These cells were further categorized into five clusters (Cluster 1–5) based on their expression profiles. The non-zero mean expression of all genes across these selected cells was calculated to generate pseudobulk gene expression profiles. Noise-augmented sample amplification The rank-transformation converts gene expression values into discrete integer ranks within each sample. Leveraging this gene expression rank order as the input for BALL6, we applied a data augmentation technique that introduced controlled noise perturbation to generate simulated expression profiles. While these simulated expression profiles exhibited deliberate variations from the original data, they preserved the core disease-state information encoded in the rank structure. With the combined DESeq2 and FPKM expression values for each sample, our method operated as follows: the whole protein-coding gene set denoted N={g1,g2,...,gn}, is randomly split into several subsets with fixed length m, M={g1,g2,...,gm}, M∈N. Here, we defined Ranki,N as the expression rank of gi in set N. For one subset Mi, we have an intermediate variable r: Suppose there are j subsets Mi containing gene gi, the mean of r-value: The simulated gene expression rank was expressed as follows: We set m = 1000 and k = 0.1 to control for the jitter effect in simulation. Finally, each original expression profile generated one simulation, and these data were integrated as model input, a procedure that expanded all samples equally. Features selection and ranking We applied dimension reduction by pre-selecting genes which most informative in subtype prediction. Only protein-coding genes with expression levels (FPKM) greater than 1 in 90% of training samples were included. Informed by recent methodologies37,38, each gene was evaluated according the following 12 screening criteria using the expression level and true subtype of the sample: (a) differential gene expression analysis35 (absolute fold change >2 and p < 0.001) in both 1-vs-other and 1-vs-1 comparisons; (b) high expression across subtypes; (c) top 75% variation across all samples; (d) top 5000 kendall correlation values; (e) chi-square test to select the top 5000 significant genes based on independence from subtypes via the sklearn module; (f) ANOVA to select the top 5000 significant genes based on mean differences across subtypes via the sklearn module; (g) Max Information Coefficient (MIC) to select the top 5000 features ranked by importance; (h) Recursive Feature Elimination (RFE) with logistic regression as the base model, a maximum of 5000 iterations, and class weight balancing enabled (max_iter=5000, class_weight = ‘balanced’); (i) LASSO regression with L1 regularization and the saga solver, also with class balancing; (j) Light Gradient Boosting Machine (LightGBM) with 500 decision trees, a learning rate of 0.05, and 32 leaf nodes (n_estimators = 500, learning_rate = 0.05, num_leaves = 32); (k) Random Forest (RF) implemented with 100 trees and class weight balancing. For each method, a maximum of 5000 genes were selected. We scored each gene according to the number of criteria met. Genes scoring higher than 9 were designated high-confidence candidates. These genes were further supplemented with literature-supported and manually curated genes exhibiting pronounced phenotypic traits. A total of 476 genes (229 for the AL model and 292 for the B-ALL model) were selected (Supplementary Data 4) for the BALL6 modeling. The rank-transformation, denoted as Ranki,N, was performed by ordering the selected genes within each individual sample based on their raw expression values and using their resulting relative integer ranks as the input features. Subsequently, data originated from DESeq2 or FPKM were uniformly processed through rank transformation and served as input for the BALL6 model. Notably, although FPKM and TPM (Transcripts Per Million) employ distinct quantification methods for RNA-seq counts, their values become identical after rank transformation. Therefore, TPM is equally suitable for the modeling approach used here. Model architecture and hyperparameter fine-tuning The core model architecture comprised an embedding layer to process the ordered genes, followed by two bidirectional long short-term memory (LSTM) layers to capture complex features and learn appropriate weights. We utilized Keras Tuner with the hyperband algorithm for hyperparameter optimization, performing separate searches for the AL Model and the BALL Model given their different input data sizes and numbers of prediction classes. For each model, we conducted 50 hyperparameter search trials under our defined search space where the embedding dimension was searched in [128, 512] with a step size of 128; the first LSTM layer units in [128, 512] with step 64; the second LSTM layer units in [64, 256] with step 64; dropout in LSTM layers and the dropout layer in [0.1, 0.5] with step 0.1; optimizer among Adaptive Moment Estimation (Adam), Root Mean Square Propagation (RMSprop), and Stochastic Gradient Descent (SGD), and learning rate sampled logarithmically from 1e−4 to 1e−2. After the 50 trials, the optimal values for key hyperparameters were identified based on the highest validation accuracy. Ultimately, the embedding layer dimension was set to 384. The LSTM layer sizes in the AL model were determined to be 896 and 384, while those in the B-ALL model were set to 1536 and 768, respectively. The model employs the SoftMax activation function in its final layer, which outputs the predicted probabilities for each subtype and ensures that these probabilities sum to one. The subtype with the highest probability is then taken as the final predicted subtype. Overfitting prevention To mitigate overfitting, dropout layer and recurrent dropout were set to 0.4 and 0.2, respectively. The model employed RMSProp as optimizer with learning rates of 5.20 × 10⁻⁴ for the B-ALL model and 4.88 × 10⁻⁴ for the AL model, and was trained for a maximum of 50 epochs for the B-ALL model and 20 epochs for the AL model. During model fitting, an early stopping mechanism was implemented to monitor the validation loss, automatically restoring the best model weights if no improvement was observed. To further enhance generalization performance, class weights were explicitly specified to address the inherent imbalance among disease subtypes. To prevent data leakage, the datasets were rigorously partitioned into mutually exclusive training, validation, and independent testing cohorts (Supplementary Data 2). The development datasets were randomly split into 80% for training and 20% for validation, preserving the original subtype proportions in each subset. Crucially, the entire development pipeline, including feature selection and hyperparameter tuning, was strictly confined to the training and validation sets. The independent testing datasets (e.g., SCMC and Bld.2020 for the AL model; St. Jude Cloud and Bld.2020 for the B-ALL model) remained completely unseen to the models until the final evaluation stage. Supplementary information Supplementary Information Supplementary Data1-5 Acknowledgements This work was supported by the National Natural Science Foundation of China (U24A20678 to Y.L., 32400443 to H.W.), the Project of Shanghai Municipal Science and Technology Commission (25JS2840200 to Y.L.), the Major Scientific Research Program for Young and Middle-aged Health Professionals of Fujian Province (2022ZQNZD011 to Y.L.), JINYE ZHONGZI Project from SCMC-HN (2024JYZZ-FY-03 to Y.L.) and Shanghai Key Laboratory of Clinical Molecular Diagnostics for Pediatrics (20dz2260900 to Y.L.). Author contributions B.C., S.Z., R.F., J.R., W.W. X.C. and Y.W. performed the analysis. B.C., H.S. and G.Y. performed the modeling. Y.Z., J.L., Y. H., Y. H., Q.S., H.L., R.W. and S.S. collected and integrated the data. H.W. and Y.L. supervised the study. B.C., H.W. and Y.L. designed the study and wrote the manuscript. All authors reviewed the manuscript. Data availability The RNA-seq gene expression data for the St. Jude cohort were sourced from St. Jude Cloud (https://www.stjude.cloud). RNA-seq data for the Bld.2020 cohort are available in Genome Sequence Archive for Human (GSA-human) ofhina under accession numbers HRA000119 (https:/ngdc.cncb.ac.cn/gsa-human/browse/HRA000119). The processed data analyzed in this work are available to researchers upon request. Code availability BALL6 is freely available on the web at https://cccg.ronglian.com/#/analysis, with detailed usage instruction. The source code is deposited on GitHub at https://github.com/yuliulab/BALL6. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information Bowen Cui, Email: xcxiongmao@126.com. Han Wang, Email: wanghan@scmc.com.cn. Yu Liu, Email: yu.liu@sjtu.edu.cn. Supplementary information The online version contains supplementary material available at https://doi.org/10.1038/s41698-026-01556-1. References Associated Data Supplementary Materials Supplementary Information Supplementary Data1-5 Data Availability Statement The RNA-seq gene expression data for the St. Jude cohort were sourced from St. Jude Cloud (https://www.stjude.cloud). RNA-seq data for the Bld.2020 cohort are available in Genome Sequence Archive for Human (GSA-human) ofhina under accession numbers HRA000119 (https:/ngdc.cncb.ac.cn/gsa-human/browse/HRA000119). The processed data analyzed in this work are available to researchers upon request. BALL6 is freely available on the web at https://cccg.ronglian.com/#/analysis, with detailed usage instruction. The source code is deposited on GitHub at https://github.com/yuliulab/BALL6.