Thanks for the concern! I was thinking that too, but I found a session that I think aligns well with our project (screenshot of it attached for anyone wondering!). However, if you (or anyone!) know of any conferences that are more aligned with space biology/life support, I’d be really interested in learning more; I know of AbSciCon, but that seems to be taking place next year/already happened this year.
3 Likes
Oh wow, this is awesome to learn about! It’s great that AGU is more interested in human spaceflight and biology. Interesting, there was VERY little interest by AGU a few years ago when we attended. Thanks for sharing! Cheers 
3 Likes
NB06: NB06: Light vs Dark - Gravity's Series v1 | Kaggle is officially live on the repository hub. I have expanded the pipeline into a multi-factor interaction series basic capstone to explicitly isolate the environmental delta you highlighted.
Instead of simple pairwise comparisons, the engine now models the light × spaceflight interaction term natively using a 2x2 contrast design matrix in PyDESeq2. The code handles probabilistic RSEM expected count quantization, suppresses outlier variance across all 28,187 genes via Cook’s distance mitigation, and isolates high-confidence targets.
Key Analytical Takeaways:
- The Interaction Cloud: Out of the entire transcriptomic landscape, the pipeline isolated 35 specific genes that demonstrate a significant hardware/gravity-specific interaction ($p_{adj_ISS} < 0.05$, $|\text{Interaction LFC}| > 1.5$).
- Phenotypic Variance: The vast majority of the plant’s genome handles light treatments uniformly across environments (forming the baseline cloud). However, these 35 target outliers (such as top hits
AT1G21528andAT5G20150) have their standard terrestrial operational logic completely overwritten when lighting stimuli intersect with microgravity.
The final visual canvas has been built as an interactive 4-quadrant scatter plot inside an isolated HTML iframe backdrop to ensure seamless browser rendering. You can hover directly over individual coordinates to extract exact gene IDs and Delta-LFC values for downstream functional analysis.
Next Phase: Hosting an Open Benchmark Suite (30+ Models)
To scale the utility of this framework for the AWG, my next step is transitioning this 6-notebook pipeline into a standardized evaluation benchmark across 30+ state-of-the-art computational and foundational models.
To ensure this directly serves the group’s operational goals rather than being a purely academic exercise, the benchmark will evaluate the models on two core criteria:
- Data Engineering Resilience: Testing how robustly these 30+ models can autonomously ingest raw, uncurated space-biology assets, self-diagnose technical pipeline artifacts (like the RSEM decimal parsing shifts and whitespace indexing anomalies we solved here), and execute clean alignment without manual human cleaning.
- Autonomous Signal Convergence: Measuring their ability to independently navigate high-dimensional genomic noise ($p \gg n$) and accurately converge on the exact same high-confidence interaction networks and biological pathways isolated by our baseline framework.
This transitions the “Gravity’s Series” from a static walkthrough into a functional validation suite. The goal is to establish a blueprint that helps the subgroup automate and standardize the processing of incoming spaceflight data packages as they hit the repository.
The capstone notebook and the full workflow history are open for exploration on the OSD-120 Code Hub. Feel free to fork, adjust the LFC thresholds, or build out the functional pathway tracking where the science needs it next.
Thx!
2 Likes
@amanda - can you share the link, or the screenshot with me for this whole session (P003) – trying to get a sense of who is organizing it, and whether we urge AWGers to participate. Also I can only find AGU2025 with poster titles on this, so double checking this is for AGU2026
Thanks!
@rtscott2001 Sure thing! Here’s the link: https://agu.confex.com/agu/agu26/prelim.cgi/Session/279867
1 Like
Thank you so much!
@PlantAWG We’ll be having our 5th meeting on 7/16 (Thursday) at 7pm PST. We’ll work on our AGU abstract as well as provide updates for our current data analysis pipeline. Feel free to reach out if you have any questions. Hope to see you there!
Here is the meeting link: PGC Subgroup Meeting #5
1 Like
Hi everyone,
I’ve been a bit quiet here, but I’ve been working on this in the background. Here’s a new Kaggle benchmark I’ve put together on NASA GeneLab’s OSD-120 Arabidopsis RNA-seq dataset. It’s called “OSD-120 Resilience Evaluation” and is designed to probe how well large language models can both plan and execute a realistic GLM-style RNA-seq workflow on spaceflight vs ground-control data.
The benchmark has two tasks:
- a hard execution task, where models must autonomously write and run Python code that ingests the OSD-120 AnnData object, fits an interaction model, and returns aligned outputs and interaction scores;
- an easier JSON specification task, where models describe a complete OSD-120 resilience-analysis plan in structured JSON without running any code.
By comparing these tasks, you can see where frontier models move from “nice biological language” to actually preserving biologically meaningful signal (including a curated set of 35 canonical genes) through the pipeline. The benchmark is backed by the NB01–NB06 notebook series that walks through the original OSD-120 analysis from root signal to pathway-level interpretation.
kaggle
Benchmark leaderboard and task descriptions:
Writeup (motivation, design choices, and how the rubric evolved):
I’d be very interested in any feedback from the plant AWG on which aspects of the OSD-120 study you’d most like to see stressed in future versions (e.g., additional covariates, alternative contrasts, or different biological endpoints).
I wanted to share this here before it circulates on social, since this group is closest to the underlying biology. I also have my ORCID tied to this R&D focus via a dedicated URL that should be live tomorrow.
Do reach out if you have any questions.
Thx!
Gaston D. @PlantAWG
1 Like
Could you join the next Plant AWG mtg and present this?
1 Like
I’ll try my best, my schedule is a bit unorthdox at the moment.
However, Thursdays time frame should be fine.
Worse case scenario, I plan to make a project walkthrough of what was going on my end during the testing period. It was very interesting data that I gathered in general whether it was input/output tokens, to the testing/exam questions, and tailoring the test/exam for the models due to consistent failures of actually understanding the source material I provided through the dataset and the NB01-NB06 series.
Fascinating stuff all round. I’ll try to make it, I’ve been a night owl as of late, so Thursday should be fine.
1 Like
Morning,
Just one more thing, I figured I’d share the OSD-120 Resilience Evaluation materials ahead of Thursday in case scheduling gets tight on my end.
The project site has the benchmark overview, tasks, scoring rubric, and NB01–NB06 notebook lineage, and the slides walk through the main ideas and results.
- Project site: https://osd120-resilience-evaluation-site.vercel.app
- Slides (walkthrough): OSD-120 Resilience Evaluation - Google Slides
If I can’t join live, feel free to review these and send any questions, I’m happy to follow up async.
There’s also a project navigator at the bottom which captures areas of the whole project and relays them back to you.
Enjoy the Tuesday & rest of the week ahead, thx!
You may want to share this also with @anna.lewkowicz & @james.casaletto in the Causal Inference subgroup of the @AIMLawg
They are doing work on that specific dataset, iirc
1 Like
@PlantAWG Reminder that our meeting is at 7pm (PST) tomorrow (7/16)!