CHROMA (cytogenetics)

CHROMA is a self-supervised foundation model for clinical cytogenetics, trained on images of individual chromosomes from metaphase spreads. It is unrelated to Chroma, the protein design diffusion model of the same name; this CHROMA analyses karyotypes, not proteins. It was developed by King Abdullah University of Science and Technology with Xinhua Hospital of Shanghai Jiao Tong University School of Medicine and the industry partner Smiltec (Suzhou), posted as a preprint in May 2025 and published in npj Precision Oncology in March 2026.

Karyotyping remains a first-line diagnostic assay in medical genetics and oncology, capable of revealing trisomies, translocations, inversions, deletions, ring chromosomes, and dicentrics across the whole genome at once. It is also stubbornly manual: a cytogeneticist typically selects and reviews around twenty metaphase spreads per case, and the workload compounds in laboratories where staining quality is variable and chromosomes frequently overlap. Prior automation has been narrowly task-specific — a classifier for one abnormality type, trained on one dataset — because comprehensively annotated cytogenetic corpora are scarce and severely imbalanced, with the clinically important aberrations being the rarest.

CHROMA's approach is to learn a general representation of chromosome morphology first and specialise afterwards. Pretraining runs on roughly 4 million single-chromosome and metaphase images drawn from 84,471 specimens across multiple clinics, which spans the real range of preparation quality. Fine-tuned heads then handle numerical abnormalities, stability-related structural aberrations, and instability-related lesions within a single framework, and a risk-control layer abstains on low-confidence cases rather than forcing a call.

#Key Features

• Band-guided masked autoencoding: The masking strategy is tailored to chromosome banding topology rather than applied uniformly, directing the model toward the fine banding patterns that distinguish subtle structural aberrations.
• Noise injection and denoising: An auxiliary corruption-and-recovery objective on unmasked regions builds tolerance to the overlapping chromosomes, low resolution, and inconsistent staining that characterise routine clinical material.
• Risk-controlled predictions: The model issues a call only when confidence clears a threshold and otherwise routes the case to a human expert, a design aimed at limiting false positives when few metaphase cells are available per sample.
• Label efficiency: It reaches target performance with 35–45% fewer labelled samples than task-specific baselines, and degrades little as training sets become more imbalanced.
• One model across aberration classes: Numerical changes such as monosomy and trisomy, structural changes such as translocations and inversions, and instability lesions such as fragments and dicentrics are handled by adapting a single pretrained backbone.

#Technical Details

The encoder is a ViT-Large — 24 transformer blocks, 1024-dimensional embeddings, 16×16 patches on 224×224 inputs — paired with an eight-block ViT-Small decoder during masked autoencoder pretraining. Pretraining ran for 800 epochs with a 40-epoch warm-up at batch size 1536 across eight NVIDIA V100 GPUs, about 18.5 days of compute, and performance tracked a data scaling law with R² = 0.89. Three annotated downstream sets were built with a human-in-the-loop workflow: 830,000 chromosomes for numerical abnormalities, 50,000 for stability-related structural abnormalities, and 50,000 for instability-related events including 1,706 rare aberrations. On 24-class chromosome identification, CHROMA reaches 99.8% average specificity and 94.7% sensitivity, and cell-level detection of monosomy and trisomy reaches AUC 0.959 for chromosome 7 and 0.966 for chromosome 21, both significant against the strongest baseline. Comparators were direct supervised learning, contrastive self-supervised pretraining, and a standard masked autoencoder.

#Applications

The intended setting is the clinical cytogenetics laboratory: prenatal screening, reproductive medicine, constitutional genetic diagnosis, and hematologic malignancy workup, where karyotype interpretation drives counselling and therapy decisions. By triaging confident cases automatically and flagging uncertain ones, the system is positioned as decision support that reduces roughly 40% of expert annotation workload rather than as an autonomous diagnostic. The authors emphasise resource-limited laboratories, where staining quality is more variable and trained cytogeneticists are scarcer.

#Impact

CHROMA extends the foundation-model pattern established in histopathology to a diagnostic modality that had seen only bespoke classifiers, and its peer-reviewed publication in a Nature Portfolio journal gives the results external scrutiny. Source code and trained weights are released under AGPL-3.0 on GitHub and Hugging Face, alongside annotated benchmark datasets. The large internal pretraining corpus cannot be shared for privacy reasons and is available only under a data use agreement, which limits independent reproduction; evaluation is retrospective on held-out internal test sets, and prospective clinical validation has not been reported.