GenBloom: Hematology Multimodal Representation

• GenBloom is a hematology-specific, slide-level representation learning framework that aggregates single-cell images into patient embeddings aligned with cytogenetic and mutation profiles.
• It employs a two-stage training process, starting with self-supervised vision pretraining followed by supervised genetic alignment, to fuse morphology with molecular data.
• The framework outperforms traditional methods in both classification and cross-modal retrieval tasks, demonstrating its effectiveness in integrating visual, cytogenetic, and genomic information.

Searching arXiv for GenBloom and closely related papers mentioned in the provided data. GenBloom is a hematology-specific, slide-level representation learning framework that takes as input sets of single white blood cell images from peripheral blood smears, aggregates them into a patient-level embedding, and aligns that embedding with the patient’s karyotype and somatic mutation profile using supervised contrastive learning. It is trained in two stages: self-supervised, vision-only pretraining of a patient-level transformer aggregator, followed by genetic alignment in a shared latent space. The framework is designed for the clinical setting in which visual single-cell evaluation is routinely interpreted together with cytogenetics and molecular genetics, and it provides both downstream diagnostic representations and cross-modal retrieval capabilities (Dasdelen et al., 28 May 2026).

1. Definition and scope

GenBloom operates at single-cell resolution but produces a patient-level representation. The input is an unordered set of single white blood cell images, represented through cell embeddings, and the output is a slide embedding $s_p \in \mathbb{R}^{768}$ together with a genetically aligned embedding $z^s_p \in \mathbb{R}^{128}$. Its stated objective is to align morphology with chromosomal aberrations and somatic mutations so that patients with similar genetic drivers cluster in embedding space, while preserving utility for classification, retrieval, and linear probing (Dasdelen et al., 28 May 2026).

The framework differs from unimodal morphology-only models and from whole-slide-image foundation models trained mainly on solid-tumor histopathology. In the reported formulation, GenBloom is hematology-specific, uses a hematology cell encoder, treats patient-level inputs as sets rather than spatial grids, and explicitly aligns morphology with karyotype and mutation vectors. This yields a multimodal representation that is consistent with routine hematological diagnosis, where smear morphology is interpreted jointly with cytogenetics and molecular genetics (Dasdelen et al., 28 May 2026).

A recurring source of confusion is nomenclature. In the supplied material, “GenBloom” refers to the hematology framework described above, whereas “BloomGML” denotes a separate bilevel graph machine learning framework, “BloomGML: Graph Machine Learning through the Lens of Bilevel Optimization” (Zheng et al., 2024). The two are unrelated in domain and methodology.

2. Modalities, cohorts, and encodings

The image modality consists of peripheral blood smear single-cell images from the Munich Leukemia Laboratory. For image-only pretraining, the cohort contains 1,634 patients and 794,527 single-cell images spanning acute leukemias, MDS, MPN, overlap syndromes, lymphoma, multiple myeloma, reactive changes, and healthy controls. Each cell image is resized to $224 \times 224$, normalized with ImageNet mean and standard deviation, and embedded by a frozen DinoBloom-B encoder. For the patient-level aggregator, each patient is represented as a set of up to 500 cell embeddings, and the model explicitly treats this input as an unordered set (Dasdelen et al., 28 May 2026).

The cytogenetic modality is derived from the AML-Hehr dataset, which contains 189 AML patients plus healthy controls and 52 distinct karyotypes. Karyotypes are documented in ISCN and converted via CytoGPS into binary indicators for loss, gain, and fusion per cytoband. Because there are 368 cytobands and three indicators per band, the resulting karyotype vector is $y^k_p \in \{0,1\}^{1104}$. The mutation modality is based on targeted NGS panels at diagnosis, with pathogenic variants aggregated to gene-level binary indicators. After feature selection requiring measurements for at least 30 patients and both positive and negative labels, the final mutation encoding is $y^m_p \in \{0,1\}^{25}$ (Dasdelen et al., 28 May 2026).

The framework uses separate datasets for pretraining, alignment, and external evaluation.

Dataset	Content	Use
Pretraining dataset	1,634 patients; 794,527 cells	Stage 1 self-supervised pretraining
AML-Hehr	189 AML patients + healthy cohort; paired cytology, cytogenetics, molecular genetics	Stage 2 alignment and in-domain evaluation
APL-AML	APL vs other AML	Out-of-domain evaluation
AMH	AML vs healthy	Out-of-domain evaluation

Within AML-Hehr, genetic alignment uses 146 patients for training and 43 held-out patients for testing. For the 5-class genetic subtyping task on the test set, the classes are PML::RARA fusion, CBFB::MYH11 fusion, NPM1 mutation, RUNX1::RUNX1T1 fusion, and controls, with train/test counts $18/6$, $28/9$, $28/8$, $25/7$, and $47/13$, respectively. In downstream experiments, all tasks use 5-fold cross-validation with fixed test sets (Dasdelen et al., 28 May 2026).

3. Architecture and representation space

GenBloom consists of a frozen single-cell encoder, a patient-level transformer aggregator, and modality-specific projection heads. The single-cell encoder is DinoBloom-B, a previously published hematology foundation model that takes $z^s_p \in \mathbb{R}^{128}$0 RGB cell images and outputs fixed-dimensional embeddings. In GenBloom, DinoBloom-B is not further trained; it supplies the tokens for the patient-level model (Dasdelen et al., 28 May 2026).

The patient-level aggregator is a small Vision Transformer adapted from the ViT architecture. It has $z^s_p \in \mathbb{R}^{128}$1 layers, $z^s_p \in \mathbb{R}^{128}$2 self-attention heads per layer, embedding dimension $z^s_p \in \mathbb{R}^{128}$3, and MLP hidden dimension 3072. Patchification is removed: each cell embedding is treated as a token. Positional encodings are also removed to enforce permutation invariance over cells. The input comprises a learnable [CLS] token and up to 500 cell embeddings, and the final [CLS] token is the patient-level slide embedding

$z^s_p \in \mathbb{R}^{128}$4

Because there is no spatial grid, the [CLS] token functions as a learned pooling mechanism over an unordered set of cells (Dasdelen et al., 28 May 2026).

Three modality-specific MLP projection heads map the slide embedding, karyotype vector, and mutation vector into a shared 128-dimensional latent space: $z^s_p \in \mathbb{R}^{128}$5 The $z^s_p \in \mathbb{R}^{128}$6-normalization makes dot products equivalent to cosine similarity. This shared space is the genetically aligned latent space used for k-NN, logistic regression, and cross-modal retrieval (Dasdelen et al., 28 May 2026).

4. Two-stage training procedure

Stage 1 is self-supervised vision pretraining and yields GenBloom-V. The input space is not raw images but frozen DinoBloom embeddings. Training uses multi-crop subsampling over cells: for each patient, the model samples $z^s_p \in \mathbb{R}^{128}$7 global bags containing 70% of cells, approximately 350 cells, and $z^s_p \in \mathbb{R}^{128}$8 local bags containing 20% of cells, approximately 100 cells. Each bag is processed independently by the aggregator. The training scheme follows a teacher-student setup in which the teacher network is an exponential moving average of the student parameters (Dasdelen et al., 28 May 2026).

Two losses are used in Stage 1. The first is a DINO-style alignment loss on [CLS] tokens across multi-crop views,

$z^s_p \in \mathbb{R}^{128}$9

and the second is an iBOT-style masked token prediction loss on masked cell embeddings,

$224 \times 224$0

The combined objective is

$224 \times 224$1

Reported training details for this stage are 100 epochs, batch size 64 patients, single NVIDIA H100 80GB hardware, and estimated emissions of approximately $224 \times 224$2 for all experiments (Dasdelen et al., 28 May 2026).

Stage 2 performs genetic alignment and yields GenBloom-G. The transformer aggregator is initialized from GenBloom-V and fine-tuned with a cross-modal supervised contrastive objective defined over subtype labels $224 \times 224$3. For an anchor modality $224 \times 224$4 and target modality $224 \times 224$5, the unidirectional loss is

$224 \times 224$6

with bidirectional loss

$224 \times 224$7

This is applied for slide-to-karyotype and slide-to-mutation alignment, producing $224 \times 224$8 and $224 \times 224$9 (Dasdelen et al., 28 May 2026).

To retain detailed genetic information and avoid degeneracy, GenBloom adds decoders that reconstruct the original karyotype and mutation vectors from the bottleneck representations. The reconstruction term is

$y^k_p \in \{0,1\}^{1104}$0

and the total alignment objective is

$y^k_p \in \{0,1\}^{1104}$1

The reported ablations vary aggregator initialization, karyotype resolution at band level versus arm level, and reconstruction weight $y^k_p \in \{0,1\}^{1104}$2 (Dasdelen et al., 28 May 2026).

5. Empirical behavior, downstream tasks, and ablations

The reported downstream tasks are patient-level classification and retrieval. Classification is evaluated on AML-Hehr as a 5-class genetic subtyping task, on APL-AML as a 2-class APL-versus-other-AML task, and on AMH as a 2-class AML-versus-healthy task. The primary metric is balanced accuracy, with both $y^k_p \in \{0,1\}^{1104}$3-NN ($y^k_p \in \{0,1\}^{1104}$4) and logistic regression used as evaluation protocols. Baselines include GenBloom-V, GenBloom-G, GigaPath, PRISM, TITAN, and mean pooling of DinoBloom cell embeddings (Dasdelen et al., 28 May 2026).

On AML-Hehr 5-class classification, GenBloom-G achieved the highest balanced accuracy and outperformed TITAN, the second-best method, by $y^k_p \in \{0,1\}^{1104}$5 balanced accuracy in $y^k_p \in \{0,1\}^{1104}$6-NN and $y^k_p \in \{0,1\}^{1104}$7 balanced accuracy in logistic regression. On APL-AML, GenBloom-G again ranked best, outperforming PRISM by $y^k_p \in \{0,1\}^{1104}$8 balanced accuracy for $y^k_p \in \{0,1\}^{1104}$9-NN and $y^m_p \in \{0,1\}^{25}$0 balanced accuracy for logistic regression. On AMH, GenBloom-V was slightly better than TITAN in $y^m_p \in \{0,1\}^{25}$1-NN and similar in logistic regression, while genetic alignment did not substantially change performance, which the paper reports as consistent with AMH being a simpler morphology-dominated task. Across datasets and both evaluation protocols, GenBloom-G ranked significantly higher than all baselines according to a Friedman test with $y^m_p \in \{0,1\}^{25}$2 (Dasdelen et al., 28 May 2026).

Retrieval experiments evaluate both slide-to-slide and cross-modal settings. For slide-to-slide retrieval, the metric is mean Average Precision at 3. On AML-Hehr and APL-AML, GenBloom-G achieved the highest mAP@3, followed by GenBloom-V, and outperformed TITAN by about $y^m_p \in \{0,1\}^{25}$3. On AMH, GenBloom-V was best, followed by TITAN, mirroring the classification results (Dasdelen et al., 28 May 2026).

Cross-modal retrieval uses cosine similarity in the shared latent space,

$y^m_p \in \{0,1\}^{25}$4

The paper evaluates genomics-to-slide, slide-to-genomics, and slide-to-slide retrieval. On AML-Hehr, selected in-domain results include slide-to-karyotype top-5 accuracy $y^m_p \in \{0,1\}^{25}$5 versus random $y^m_p \in \{0,1\}^{25}$6, with MRR $y^m_p \in \{0,1\}^{25}$7 versus $y^m_p \in \{0,1\}^{25}$8; karyotype-to-slide top-5 accuracy $y^m_p \in \{0,1\}^{25}$9 versus $18/6$0, with MRR $18/6$1 versus $18/6$2; slide-to-mutation top-5 accuracy $18/6$3 versus $18/6$4; and mutation-to-slide top-5 accuracy $18/6$5 versus $18/6$6. All improvements over the random baseline are statistically significant with $18/6$7 (Dasdelen et al., 28 May 2026).

The ablations identify the transformer aggregator and the reconstruction term as important. On the AML-Hehr test set, a transformer with GenBloom-V initialization achieved logistic-regression balanced accuracy $18/6$8, slide-to-karyotype MRR $18/6$9, and karyotype-to-slide MRR $28/9$0, compared with $28/9$1, $28/9$2, and $28/9$3 for mean pooling. Band-level karyotype encoding slightly outperformed arm-level encoding, with $28/9$4 versus $28/9$5 in logistic-regression balanced accuracy. Reconstruction weight $28/9$6 yielded the best reported values, whereas reducing or removing reconstruction lowered both retrieval and classification performance (Dasdelen et al., 28 May 2026).

The paper also reports out-of-domain per-gene retrieval on cAItomorph. For JAK2 with $28/9$7, slide-to-mutation F1 was $28/9$8 versus random $28/9$9, and mutation-to-slide F1 was $28/8$0 versus $28/8$1. For NPM1 with $28/8$2, the corresponding values were $28/8$3 versus $28/8$4 and $28/8$5 versus $28/8$6. For NRAS with $28/8$7, they were $28/8$8 versus $28/8$9 and $25/7$0 versus $25/7$1. In 13 of 14 gene-direction combinations, GenBloom significantly outperformed random with $25/7$2, which the authors interpret as evidence that the embeddings encode gene-specific morphological signals (Dasdelen et al., 28 May 2026).

6. Clinical positioning, interpretability, and limitations

GenBloom is explicitly positioned to mirror routine hematology workflows. Peripheral blood and bone marrow smears are examined first, and cytogenetics and molecular profiling then refine diagnosis, risk, and treatment. The framework uses smear morphology as the primary input but constrains the representation space by genetics. The reported use cases include fast triage, subtype classification for entities such as PML::RARA, NPM1, and RUNX1::RUNX1T1, and cross-modal retrieval of similar patients or genetic profiles. A plausible implication is that the method is intended not as a replacement for cytogenetics or molecular testing, but as a genetics-aware representation layer for integrated decision support (Dasdelen et al., 28 May 2026).

Interpretability in the paper is indirect rather than cell-explanatory. The shared latent space clusters patients by recurrent genetic subtype in UMAP visualizations, and the transformer architecture implicitly weights cells through attention. However, the paper does not provide explicit attention visualization or a detailed analysis of which morphological features per cell drive alignment. It therefore emphasizes representation quality and retrieval more than mechanistic interpretability (Dasdelen et al., 28 May 2026).

The limitations are clearly stated. The primary alignment cohort is relatively small, with 189 patients, and comes from a single center. Rare alterations have few examples, the cohort is largely AML, and generalizability to non-AML hematologic malignancies is not fully characterized. The data are cross-sectional at diagnosis, only targeted gene panel mutations and karyotypes are included, and cross-institutional robustness remains to be tested. Future directions mentioned or implied include integrating additional omics such as transcriptomics and whole-genome sequencing, adding modalities such as flow cytometry and clinical laboratory values, scaling to larger multi-center pretraining, adapting the framework to other hematologic diseases, and jointly fine-tuning the cell encoder with the aggregator (Dasdelen et al., 28 May 2026).

Within the broader literature, GenBloom sits at the intersection of computational pathology foundation models and multimodal alignment with molecular data. Its distinctive contribution is the combination of single-cell hematology cytology, a hematology-specific cell foundation model, a transformer aggregator over unordered cell sets, and explicit alignment with both karyotype and targeted mutation profiles. The reported result is that genetically aligned patient representations built from hematology smear images outperform general whole-slide foundation models on hematology-specific tasks while supporting slide-to-slide and slide-to-genetics retrieval in a unified latent space (Dasdelen et al., 28 May 2026).