🤖 machine learning

Genetically Aligned Patient Representations Improve Hematological Diagnosis

This study introduces a two-stage framework that aligns single white blood cell images with chromosomal and somatic genetic data to create genetically informed patient representations, which significantly outperform existing histopathology foundation models in hematological diagnosis and enable effective retrieval of diseases and genetic alterations.

Original authors: Muhammed Furkan Dasdelen, Fatih Ozlugedik, Ilaria Looser, Rao Muhammad Umer, Christian Pohlkamp, Carsten Marr

Published 2026-05-29

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammed Furkan Dasdelen, Fatih Ozlugedik, Ilaria Looser, Rao Muhammad Umer, Christian Pohlkamp, Carsten Marr

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a complex mystery: identifying the specific type of blood cancer a patient has. Traditionally, doctors act like two different detectives working in separate rooms. One detective looks at photos of blood cells under a microscope (the visual clues), while the other detective reads a genetic report listing mutations and chromosomal errors (the molecular clues). Usually, these two detectives work independently, and only later do they compare notes to make a final diagnosis.

The paper you shared introduces a new AI system called GenBloom that acts like a "super-detective" who can read both the photos and the genetic reports simultaneously, learning how they fit together to solve the mystery faster and more accurately.

Here is a simple breakdown of how they built it and what they found:

1. The Problem: Two Languages, One Patient

In blood cancer diagnosis, visual clues (what the cells look like) and genetic clues (what the DNA says) are tightly linked. However, most AI models today only speak one language. They are great at looking at pictures, but they don't understand the genetic code, or vice versa. This paper wanted to build a model that speaks both languages fluently.

2. The Solution: A Two-Stage Training Camp

The researchers trained GenBloom in two distinct phases, like a student first learning to read, then learning to write.

Stage 1: The Visual Boot Camp (Pretraining)
First, they fed the AI over 700,000 images of single blood cells from more than 1,500 patients. The AI wasn't told what disease these patients had yet. Instead, it was just asked to look at the cells and learn their shapes, sizes, and textures.
- Analogy: Think of this as a student studying thousands of photos of birds to learn the difference between a sparrow, an eagle, and a hawk, without yet knowing the scientific names of the species.
- Result: The AI became very good at recognizing visual patterns in blood cells.
Stage 2: The Genetic Alignment (Connecting the Dots)
Next, they took a smaller group of patients (146 people with a specific type of leukemia) who had both cell images and genetic test results. They taught the AI to link the visual "photos" it had already learned with the specific "genetic codes" (like specific mutations or chromosomal errors).
- Analogy: Now, the student is shown a photo of a bird and told, "This is a Red-tailed Hawk because of this specific genetic marker." The AI learns to say, "Ah, when I see this specific shape in the photo, it usually means this specific genetic error is present."
- Result: The AI created a single "patient ID card" that contains both the visual look and the genetic makeup of the patient.

3. The Results: Why It Matters

The researchers tested this new "super-detective" against other powerful AI models that only looked at pictures.

Better Diagnosis: When asked to identify specific types of leukemia, GenBloom was significantly more accurate than the other models. It outperformed them even though it was trained on much less data.
- Analogy: It's like a detective who solved a case with fewer clues because they knew exactly which clues mattered, rather than a detective who had to sift through a massive pile of irrelevant evidence.
Cross-Modal Retrieval (The "Search Engine" Feature): Because the AI learned to link images and genes, it can now act like a search engine.
- You can show it a picture of a blood cell, and it can find the matching genetic profile.
- You can give it a genetic mutation (like "NPM1 mutation"), and it can find the matching blood cell images.
- Analogy: It's like having a library where you can find a book by its cover art, or find the cover art by reading the title. Before, you could only search by one or the other.

4. What They Did Not Claim

It is important to stick to what the paper actually says:

They did not say this AI is currently being used in hospitals to diagnose real patients today.
They did not claim it replaces doctors.
They did not predict future cures.
They simply proved that by teaching an AI to look at both pictures and genetic data together, it creates a much better "representation" (a digital summary) of the patient, which leads to more accurate classification of the disease in a computer test.

The Bottom Line

GenBloom is a new tool that bridges the gap between what blood cells look like and what their DNA says. By training the AI to understand both languages at once, it creates a more accurate and efficient way to categorize blood diseases, offering a new way to search for and match patients based on both their visual and genetic fingerprints.

Technical Summary of GenBloom: Genetically Aligned Patient Representations for Hematological Diagnosis

Problem Statement
Hematological diagnosis requires the tight integration of microscopic morphology (visual evaluation of blood smears) with cytogenetic and molecular genetic profiling to achieve precise tumor classification, risk stratification, and treatment selection. While deep learning foundation models trained on whole-slide images (WSIs) have shown promise in general computational pathology, existing hematological models predominantly analyze morphology alone, failing to capture underlying molecular heterogeneity. Furthermore, the scarcity of large-scale paired datasets (images linked to specific genetic alterations) has restricted the development of models that can jointly model these modalities. There is a critical need for a framework that aligns single-cell visual features with chromosomal aberrations (karyotypes) and somatic mutations to create comprehensive patient representations.

Methodology
The authors present GenBloom, the first genetically aligned, slide-level blood model tailored to hematology. The framework employs a two-stage training paradigm:

Image Pretraining (Self-Supervised):
- Data: A cohort of 1,634 patients (>794,000 single-cell images) spanning various hematological diseases (e.g., AML, MDS, MPN, lymphoma) and healthy controls.
- Architecture: A patient-level transformer aggregator (Vision Transformer, ViT) with 6 layers and 12 heads. To handle unordered sets of cells, the model removes standard patchification and positional encodings, replacing them with an MLP to enforce permutation invariance.
- Training Strategy: The model uses a DINOv2/iBOT-style self-supervised approach adapted for "patient bags." It operates in embedding space using a frozen DinoBloom-B encoder to extract single-cell representations. The training involves multi-crop subsampling (global and local bags) and a student-teacher objective combining DINO-style [CLS] alignment and iBOT masked embedding prediction.
Genetic Alignment (Supervised Contrastive Learning):
- Data: The AML-Hehr dataset comprising 189 AML patients with paired peripheral blood smear images, somatic mutations (25 gene-level binary features), and cytogenetic profiles (1,104-dimensional binary vectors derived from CytoGPS processing of ISCN karyotypes).
- Alignment Mechanism: The model projects slide embeddings, karyotype vectors, and mutation vectors into a shared 128-dimensional $\ell_2$ -normalized space using modality-specific MLP heads.
- Loss Function: A cross-modal supervised contrastive loss aligns the modalities by pulling together embeddings of patients sharing the same class label (e.g., specific fusion or mutation) while pushing apart different classes.
- Regularization: To prevent representational collapse and preserve modality-specific information, lightweight decoders are attached to reconstruct the original binary genetic vectors, supervised by a binary cross-entropy loss.

Key Contributions

GenBloom Framework: Introduces a novel architecture that integrates single white blood cell images with cytogenetic abnormalities and somatic mutations to learn comprehensive patient embeddings.
Two-Stage Training: Demonstrates a successful paradigm of large-scale self-supervised pretraining on cellular morphology followed by supervised multimodal alignment to anchor patient embeddings within the genetic space.
Cross-Modal Retrieval: Enables "off-the-shelf" retrieval capabilities where a slide can retrieve relevant genetic profiles (and vice versa), and slides can retrieve other slides with the same disease or genetic alteration.
Efficiency: Achieves state-of-the-art performance with substantially fewer parameters and less training data compared to large-scale histopathology foundation models (e.g., Gigapath, PRISM, TITAN).

Results

Classification Performance: On the AML-Hehr genetic subtyping task (5 classes), GenBloom-G (genetically aligned) achieved the highest balanced accuracy, outperforming the second-best model (TITAN) by 38% in k-NN and 5% in logistic regression. On the APL-AML dataset, it outperformed PRISM by 15% (k-NN) and 45% (logistic regression).
Retrieval Performance: GenBloom-G significantly outperformed random baselines in cross-modal retrieval. For slide-to-karyotype retrieval, it achieved a top-5 accuracy of 0.62 (vs. 0.23 for random), representing a 2.7x improvement. It also demonstrated strong performance in out-of-domain retrieval tasks on the cAItomorph cohort for specific genes (e.g., NPM1, JAK2, ASXL1).
Ablation Studies: The study confirmed that initializing the transformer with pretrained weights (GenBloom-V) is superior to random initialization. Furthermore, band-level karyotype encoding provided better alignment signals than coarser arm-level representations, and the inclusion of the reconstruction loss significantly improved retrieval metrics (MRR) and classification accuracy.

Significance
The paper claims that incorporating genetic data into patient encoders significantly increases the quality of patient representations. GenBloom demonstrates that aligning morphology with karyotype and mutation profiles creates a shared latent space that captures clinically meaningful relationships between cellular appearance and molecular disease drivers. This approach aligns with clinical diagnostic workflows where visual and genetic data are routinely paired. The authors posit that this framework paves the way for future multimodal, hematology-specific AI, potentially supporting faster triage, prioritization for confirmatory testing, and more informed treatment decisions. The code and model weights are made available to facilitate further research in this domain.