ConnectoFM: New Foundation Model Advances Neural Circuit Reconstruction from Brain Imaging
Researchers have developed ConnectoFM, a foundation model pretrained on 1.7 million electron microscopy images to improve the automated reconstruction of neural circuits from brain tissue. The model combines masked image modeling and contrastive learning to create robust visual representations across diverse brain samples from six species. This advance could significantly accelerate connectomics research by reducing manual annotation work and improving accuracy in identifying synapses and other critical neural structures.
ConnectoFM represents the first foundation model specifically designed for connectomics, the field focused on mapping neural circuits at high resolution. Trained on a diverse corpus of 1.7 million unlabeled electron microscopy images from six species and 25 brain subdomains, the model learns to recognize biologically meaningful patterns in brain tissue images. When applied to three key downstream tasks—binary segmentation, multiclass cell typing, and instance segmentation—ConnectoFM outperformed existing electron microscopy foundation models and task-specific methods across 29 diverse datasets. Notably, the model achieved superior performance with only 10% labeled data compared to baseline methods trained on 100% of available annotations, demonstrating its data efficiency. The improvements were particularly pronounced for challenging structures like membranes, mitochondria, vesicles, synapses, and post-synaptic densities, which are critical for understanding neural connectivity.
Limitations & open questions
The preprint does not discuss potential limitations of the model, such as generalization to brain tissue from other organisms beyond the six species studied, computational requirements for deployment, or comparison of performance metrics with human expert annotation accuracy. Additionally, the timeline for potential clinical or research applications and any planned open-source release are not addressed.
What different sources said
- bioRxivCenter
ConnectoFM: A Foundation Model for Learning the Language of the Connectome
Related
Fluorescence Correlation Spectroscopy Used to Monitor Chlamydia Protein Production in Cell-Free System
Researchers used fluorescence correlation spectroscopy (FCS) to track the real-time production of a Chlamydia outer protein (CopB) fused with a fluorescent marker in a cell-free protein synthesis system. The technique allowed them to measure protein concentration, size, aggregation, and maturation rates without requiring protein purification. This approach could streamline the characterization of proteins during synthesis for research and biotechnology applications.
DNA Origami Nanoparticles with Oligolysine Coating Show Promise for Retinal Cell Uptake
Researchers found that coating DNA origami nanoparticles with PEG5K-K10, a cationic polymer, significantly improves their uptake into retinoblastoma cells while maintaining safety and mobility in the eye. The coating was essential for cell internalization and did not impair the particles' ability to diffuse through the vitreous humor. These findings suggest DNA origami nanoparticles could be viable carriers for delivering therapeutic agents to treat eye diseases.
Spatial Clustering of Adhesion-Deficient Cells Controls Epithelial Tissue Rigidity, Study Shows
A computational study using vertex modeling demonstrates that how adhesion-deficient cells are spatially arranged in epithelial tissue—not just their number—determines whether the tissue loses mechanical rigidity. Clustered mutant cells cause sustained mechanical changes through sequential boundary removal, while randomly distributed ones are rapidly eliminated with minimal lasting effects. The findings suggest spatial organization is a key factor in early-stage cancer progression and epithelial-mesenchymal transition.