Data Challenge Results 

Entrants were tasked with 1) creating an algorithm that will accurately match fluorescent images of genetic tools to the most likely spatial transcriptomic cell types and/or 2) presenting such an algorithm as part of a user-friendly tool. The two categories were judged separately and resulted in two winning teams. The judging rubric and the challenge answer key have been posted on the GitHub page. 

The winning teams will present their solutions in a webinar on August 27th, from 10-11am (Pacific).

Register for the webinar here. 

Congratulations to our algorithm accuracy winners!

(left to right)
Daniel Ryskamp Rijsketic, PhD (Senior Data Scientist)
Austen Casey, PhD (Postdoctoral Scholar)
Boris Heifets, MD, PhD (Principal Investigator; Associate Professor of Anesthesiology, Perioperative and Pain Medicine) 

Affiliation: Stanford University 

Tool/algorithm name: UNRAVEL (UN-biased high-Resolution Analysis and Validation of Ensembles using Light sheet images) 

Description of tool/algorithm: UNRAVEL is a Python library and command-line tool for mapping and quantifying fluorescent labels across the whole mouse brain from 3D images acquired through light sheet microscopy or serial two-photon tomography. It supports region-wise and voxel-wise analysis, provides hierarchical anatomical descriptions, and integrates data from the Allen Brain Cell Atlas to estimate cell-type composition and gene expression at every ontological level within any region of interest. 

Links: https://github.com/b-heifets/UNRAVEL  
https://b-heifets.github.io/UNRAVEL/  

Congratulations to our user-friendly tool winners! 

(left to right)
Nikhil Karthik, PhD (Editor)
Chaitali Khan, PhD (Postdoctoral Fellow) 

Affiliations:
Karthik: Physical Review Letters, American Physical Society.
Khan: National Heart, Lung, and Blood Institute (NIH) 

Tool/algorithm name: Genetic Tool to Cell Type Bayesian Mapper (GT2CT Bayesian Mapper) 

Description of tool/algorithm:  The GT2CT Bayesian Mapper algorithm uses a Multilayer Perceptron to learn a multi-celltype 3D spatial probability distribution from MERFISH cell data. Then, it fits the set of K-subclass-cell types (where K represents the typical cell type specificity of Genetic Tools) that best describes the intensity-based construction of the spatial probability distribution of GFP/RFP cells. Finally, the algorithm finds the fractions of the K subclass-cell types in the GFP/RFP populations by marginalizing the learnt probability distributions over VISp. The algorithm can be generalized to non-VISp areas, and to classification at class, subclass, supercluster, or cluster hierarchical levels.  

Link: https://github.com/astronikil/GT2CT 

Thank you to the other participating teams for entering this challenge, entrants are listed below alphabetically. 

Nooshin Bahador, PhD (Postdoctoral Fellow) 

Affiliation: University Health Network (Canada) 

Tool/algorithm name: Cell Type Analyzer 

Description of tool/algorithm (2-3 sentences preferred): GitHub repository for a user-friendly tool and supporting codes for cell type inference from 2D fluorescent brain slice images. 

Link:  https://github.com/nbahador/cell_type_analyzer.git 

Inkar Kapen (Data Analyst I) 

Affiliation: Allen Institute for Brain Science 

Tool/algorithm name: ResNet18 3D from torchvision video models 

Description of tool/algorithm: ResNet-18-based 3D convolutional neural network designed for video or volumetric data trained using MERFISH and Fluorescent scan data. 

Links: https://mapmysections-bappbnxmnxoresuethu8hcm.streamlit.app/ (tool),  https://docs.pytorch.org/vision/main/models/generated/torchvision.models.resnet18.html and https://github.com/inkarkapen/MapMySections/tree/master 

Taha Razzaq (Machine Learning Researcher) and Asim Iqbal, PhD (Co-founder & CEO)
Affiliation: Tibbling Technologies 

Tool/algorithm name: Tibbling NeuroMapper  

Description of tool/algorithm: A web-based tool that automatically registers brain slice images to the Allen CCF atlas and quantifies the distribution of cell types across brain substructures. The tool provides intuitive visualizations of substructure composition, enabling seamless integration of custom samples with standardized atlas references.  

Link: https://huggingface.co/spaces/TibbtechUser/MapMySections 

Thomas Topilko, PhD (Postdoctoral Fellow) and Silas Dalum Larsen (PhD Fellow) 

Affiliation: University of Copenhagen (Denmark) 

Tool/algorithm name: CellMap2D 

Description of tool/algorithm: CellMap2D is a Python-based pipeline developed for the Allen Institute’s MapMySections challenge, aiming to infer the transcriptional identity of cells in high-resolution 2D brain sections. Built by Thomas Topilko and Silas Dalum Larsen, the tool processes raw histological images through cell segmentation, anchors them in 3D brain space using DeepSlice, and performs 2D registration with ANTs for precise alignment. The registered cells are then compared to spatial transcriptomic reference maps from the ABC Atlas using kernel density estimation (KDE), allowing the pipeline to rank Candidate cell subclasses based on spatial similarity and thus predict likely cell identities. 

Link: https://github.com/Tom-top/CellMap2D  

About This Challenge

Transgenic lines and viral tools provide highly valuable resources for targeting subpopulations of cells in the brain of mouse (and other species). The recent identification of >5,000 mouse brain cell populations with distinct gene expression patterns and spatial profiles provides opportunities not only to create new cell type-specific genetic tools, but also to better characterize the cell type specificity of existing genetic tools widely used in neuroscience or newly generated in a lab. Accurately defining this connection between genetic tools and known cell types represents a critical step in interpreting the results of experiments using these tools, from functional assays to potential gene therapies. A detailed cell type characterization of the labeled cells can be achieved through a combination of single cell RNA-sequencing and cell sorting; however, such methods take time, are costly, and are prone to bias. A method for directly inferring cell types from fluorescent images without the need for additional experiments would be immediately applicable to thousands of existing genetic tools, greatly improving their utility and interpretability.

Entrants are tasked with 1) creating an algorithm that will accurately match fluorescent images of genetic tools to the most likely spatial transcriptomic cell types and/or 2) presenting such an algorithm as part of a user-friendly tool like MapMyCells. Although images span the whole brain, this challenge is focused on defining cell types in primary visual cortex (VISp)

This challenge includes images and associated cell type specificity for anonymized genetics tools from the Genetics Tools Atlas, a searchable web resource representing information and data on enhancer-adeno-associated viruses (enhancer AAVs) and mouse transgenes. Multiple modalities for summarizing data are included as part of the atlas, but only coronal sections collected using Serial Two-Photon Tomography (STPT) sections will be directly included as part of the challenge. Cell type specificity is assessed by applying single cell RNA-sequencing (SMART-Seq v4) on fluorescently labeled cells from each genetic tool, and then mapping these cells to the published taxonomy of cell types in whole mouse brain (Yao et al, 2023), which are available in the Allen Brain Cell Atlas. Qualitative assessments of cell labeling patterns are also provided for many of the genetic tools.

This data challenge is supported by the National Institute Of Neurological Disorders And Stroke of the National Institutes of Health under Award Number U24NS133077. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.