HASSL

Single-cell microscopy images carry biologically meaningful hierarchical structure: fine-grained cellular subtypes nest inside broader morphological and functional groups. Self-supervised vision models learn useful embeddings from such images without manual labels, but their representations are frequently dominated by coarse, biologically uninteresting factors — most notably which imaging modality or dataset a cell came from. When these confounders take over the latent space, the subtle morphological distinctions that matter for biology are flattened out, and downstream tasks that depend on fine structure suffer.

HASSL (Hierarchy-Aware Self-Supervised Learning) is a representation-learning framework that directly targets this failure mode. Rather than treating all images as a flat set of instances, it is designed to preserve the nested organization of cell morphology across multiple levels of granularity, so that related-but-distinct subtypes remain separable in the embedding space instead of collapsing together under modality-driven variation.

HASSL was developed by a Munich-based collaboration spanning the student-run TUM.ai initiative, the Technical University of Munich, LMU Munich, and Helmholtz Munich, including Carsten Marr's group. It sits alongside other self-supervised microscopy foundation models such as SubCell and SpatialDINO, distinguished by its explicit focus on hierarchical structure.

#Key Features

• Hierarchy preservation: The framework is built to retain the multi-level structure of single-cell morphology, keeping fine subtypes distinct rather than letting coarse factors such as imaging modality dominate the learned representation.
• Segmentation-guided distillation teacher: A distillation setup uses a segmentation teacher to sharpen the model's morphological awareness, grounding the embeddings in cell shape and structure rather than background or acquisition artifacts.
• HDBSCAN hierarchy-aware contrastive loss: A contrastive objective built on HDBSCAN clustering sharpens boundaries between related but distinct subtypes at each level of the hierarchy, encoding structure that flat contrastive losses miss.
• Label-free downstream transfer: Learned embeddings support clustering and classification across diverse microscopy datasets without per-dataset retraining, in the frozen-embedding style typical of self-supervised microscopy models.

#Technical Details

HASSL builds on the DINO family of self-distillation vision transformers, extending them with a segmentation-guided teacher and an HDBSCAN-based hierarchy-aware contrastive loss. It was trained and evaluated on a curated corpus of 2.3 million single cells aggregated from 20 microscopy datasets spanning 208 cell classes, giving broad coverage of imaging conditions and morphologies. Evaluation spans downstream benchmarks including the Allen Cell drug-perturbation dataset, HPA single-cell classification, and the RxRx OpenPhenom collection, using metrics such as adjusted mutual information for clustering quality. Against standard self-supervised baselines including DINO, DINOv2, DINOv3, SimCLR, MoCo, and BYOL, HASSL reports a 2.8% improvement in top-K accuracy and a 7.8% gain in F1 score on drug classification, indicating that the hierarchy-aware components translate into measurable downstream gains.

#Applications

HASSL is aimed at cell biologists, high-content-imaging scientists, and computational researchers who need to organize and classify large collections of single-cell microscopy images without exhaustive manual annotation. Because its embeddings preserve fine morphological distinctions, they are well suited to phenotypic and morphological profiling — clustering cells by subtype, annotating cell classes, and separating treatment conditions in drug-perturbation screens where subtle shifts in cell appearance encode mechanism of action. A single pretrained backbone can be reused across heterogeneous datasets acquired on different instruments, lowering the analysis burden for labs assembling image collections from multiple sources.

#Impact

By showing that explicitly modeling hierarchy — through a segmentation-guided teacher and a clustering-based contrastive loss — recovers biologically meaningful morphological structure that flat self-supervised objectives suppress, HASSL contributes a concrete recipe for building microscopy foundation models whose embeddings respect the nested organization of cell types. The authors release both code and a curated single-cell benchmark, supporting reproduction and downstream reuse by the imaging community. As a preprint that has not yet undergone peer review, its reported gains await independent replication, and the breadth of transfer beyond the evaluated datasets and cell classes remains to be established.