GPTomics

Took this for a spin on both @claudeai Code and @OpenAI Codex both with and without BioSkills. Seeing both some excellent performance without the skills and some clear skills based improvement

@TheTuringPost @ylecun In parallel, we’ve seen quite a few applications of JEPA to biology also. There’s been EchoJEPA, Cell-JEPA, GeneJEPA, JEPA-DNA and others. We’re currently adapting JEPA-AC to cell perturbation and have positive early findings out for BioJEPA-AC. github.com/GPTomics/bioje…

14 most important and influential types of JEPA ▪️ JEPA / H-JEPA ▪️ I-JEPA ▪️ MC-JEPA ▪️ V-JEPA ▪️ Audio-JEPA ▪️ Point-JEPA ▪️ 3D-JEPA ▪️ ACT-JEPA ▪️ V-JEPA 2 ▪️ LeJEPA ▪️ Causal-JEPA ▪️ V-JEPA 2.1 ▪️ LeWorldModel ▪️ ThinkJEPA Save the list and check this out to explore these JEPA milestones as a map of AI progress: turingpost.com/p/jepamap

Created a new set of 33 tasks and a harness to benchmark agents on bioinformatics datasets. github.com/GPTomics/bioTa… Take it for a spin

Sneak peek at what coming this week. Doing full benchmarks on BioSkills

Started training the updated @gptomics BioJEPA-AC model. Just realized it's only 18.4M parameters. Should be great to see how much performance the latest updates can squeeze out of that.

For those curious why we are pushing forward with the next @gptomics BioJEPA so quickly, the main reason is we made a very common mistake. v0.6 treated missing genes in a dataset the same as non expressed genes, both for training and eval. We’re fixing this and a few more issues

It's great to get this version out. Despite the issues our model achieves a mean sample-level Pearson correlation of 0.898 on the top 20 DEGs, a mean R^2 of 0.860 on the top 50 DEGs, 79.3% directional accuracy on the top 50 DEGs, and a severity correlation of 0.793

v0.6 Technical report for BioJEPA-AC. While we found some issues with the data prep and evals that we're fixing in v0.7, we wanted to share this out so you all can follow along. github.com/GPTomics/bioje…

During some final validations we found we had a few data-prep issues that impacted our dataset enough that we feel we need to retrain BioJEPA. The next release we'll do is going to be v0.7

Have you been curious how we calculate training loss on each component of @gptomics BioJEPA-AC? well now there's an explainer notebook that will walk you through each major component's loss. github.com/GPTomics/bioje…

@Luyi_T A quick way to expand capabilities is adding in the bioSkills we've built. It should quickly grow the skill set to ~425 and it's also available on clawhub github.com/GPTomics/bioSk… i. .

Introducing OmicsClaw: a local-first multi-omics analysis framework for spatial transcriptomics, single-cell, genomics, proteomics, and metabolomics. 50+ analysis skills run entirely on your machine. Your genetic and spatial data never leaves your infrastructure.

Even though BioJEPA-AC doesn't use dosage in it's training or inference, we've been curious if there's dose response baked in (hint: there shouldn't be). If you're curious how we do this evaluation, we built an explainer github.com/GPTomics/bioje…

@fiddle @openclaw Take a look at some of the skills we built for common bioinformatics tasks. They’re also available on clawhub but given your background you may want to customize some to fit your knowledge and research preferences github.com/GPTomics/bioSk…

It was my birthday a couple of weeks ago so I bought myself a Mac mini and installed @openclaw Meet Tessera! If you’re wondering what a “normie” genetics professor is doing with it or want tips to get started, ask 👇 !

Modern perturbSeq datasets include combination perturbations. It's important to evaluate these independently and treat them as more than the sum of individual perturbations. For @GPTomics BioJEPA-AC we built an explainer on how we eval them github.com/GPTomics/bioje…

@Teknium @domenjemec github.com/GPTomics/bioSk…

@reedbndr @Teknium @AlecStapp @Teknium we built over 400 skills to help with use cases just like this. We found that without the skills the agents would miss nuances inherent to bioinformatic analysis. The skills help cover that and push the agents to do more thorough analysis.

How do we know if @gptomics BioJEPA-AC is learning more than just expression data? we do Mechanism-of-Action (MoA) Analysis to see if it's learning any known gene networks. This notebook deepdives into how we do it: github.com/GPTomics/bioje…

Our BioJEPA-AC model uniquely outputs uncertainty along with our latent cell representation. We evaluate that uncertainty through our uncertainty calibration. @domenjemec wroteup a detailed explanation if you're curious how we do it: github.com/GPTomics/bioje…

Wonder how we evaluate perturbation retrieval on @GPTomics BioJEPA-AC? Wonder no more, we have a whole explainer notebook github.com/GPTomics/bioje…

We found some issues in the @gptomics BioJEPA v0.6 run so we’re retraining and rerunning our eval. This will delay the final report but you should still see more explainers in the mean time.