Quick Jump
Spatial Discovery
AtoMx SIP 2.2 brings several new features and enhancements to users in the beta release of its brand new visual-first experience. This includes pre-processing and pre-calculations of many of “upstream” modules like quality, cluster, etc. In addition to visual enhancements, this new experience brings new analysis modules like Novae domains and Spatial Discovery.
In this “vlog” post, I quickly explain what the Spatial Discovery is and demonstrate and example of how I use it in combination with LLMs to accelerate my discoveries. Click the chapters to the right of the video to begin.
Study Insights
When downloading the Spatial Discovery Study Insights you will get a zip file. Within that zip file is a series of text files containing pertinent results from the study. One of those files is the readme markdown file which explains what each file contains, the column names, the data types, and other useful information that an LLM can use for added context.
Example Prompts
One of the most exciting things about this interactive approach is that you can really ask any number of questions to the prompt. I typically start by confirming if it can understand the files I uploaded and some questions defining my spatial domains and cell types before asking about biological processes but it’s entirely open-ended. Here’s some example prompts that you can try.
- Can you read and understand the contents of these files related to my colon cancer dataset? If so, I would like to have a conversation about the data with you.
- Tell me the most likely cell types from my Leiden clusters. From the data, can you provide me a table with the Leiden cluster, the most likely cell type, 1-3 marker genes that helped you determine this cell type, and a brief biological justification?
- From the Leiden cluster and domain information, can you provide a contingency table showing the counts? For example, in domain1 there are X number of Leiden 1 cells.
- Given the composition of cell types (Leiden) and domains, can you assign a biological name to each domain for me?
- What are the top 5 biologically relevant pathways in each domain?
- For <insert pathway>, show me the relationship between mean intensity and spatial autocorrelation for each cell type (Leiden cluster). Make this an interactive xy scatter plot for me.
- Tell me something interesting about this dataset.
- Give me three interesting pathway paterns that you found that I should check out spatially in AtoMx SIP.
Tips when working with LLMs
- We have tested this on several models (chatGPT, copilot, Gemini) and premium tiers tend to perform best as they can hold more content than the free tiers. Free tiers also have a shorter conversaton limit and more restricting data upload size limits.
- If working with many samples and/or conditions, using “temporary” chats can reduce the LLMs from “pulling in” results that are not related to the current study.
- Taking screenshots of the sample from AtoMx SIP and pasting them in the chat can help provide the LLM more context. LLMs thrive on plain text and verification is recommended.
If you discover other tips when working an analyzing your CosMx SMI results in this format, I would love to hear from you.