Ben Kuznets-Speck
@biophysben.bsky.social
Biophysicist working at the edge of single cell biology, machine learning, and statistical-mechanics. Postdoc in Yogesh Goyal's lab at Northwestern; PhD in biophysics, UC Berkeley; BS in math and physics, CWRU.
Big thanks to @goyallab.bsky.social, @vaikuntsuri.bsky.social, and Ekta Prashnani, as well as @czbiohub.bsky.social, @nitmb.bsky.social, @nufeinbergmed.bsky.social, @nucsb.bsky.social
December 2, 2025 at 1:37 AM
Big thanks to @goyallab.bsky.social, @vaikuntsuri.bsky.social, and Ekta Prashnani, as well as @czbiohub.bsky.social, @nitmb.bsky.social, @nufeinbergmed.bsky.social, @nucsb.bsky.social
Why?
Diffusion models are great generators—but their potential for classification & regression has barely been tapped.
And current deep classifiers?
• Hard to interpret
• Poor at uncertainty
• Don’t understand distribution geometry
We help address all three.
Diffusion models are great generators—but their potential for classification & regression has barely been tapped.
And current deep classifiers?
• Hard to interpret
• Poor at uncertainty
• Don’t understand distribution geometry
We help address all three.
December 2, 2025 at 1:33 AM
Why?
Diffusion models are great generators—but their potential for classification & regression has barely been tapped.
And current deep classifiers?
• Hard to interpret
• Poor at uncertainty
• Don’t understand distribution geometry
We help address all three.
Diffusion models are great generators—but their potential for classification & regression has barely been tapped.
And current deep classifiers?
• Hard to interpret
• Poor at uncertainty
• Don’t understand distribution geometry
We help address all three.
Why?
Diffusion models are great generators—but their potential for classification & regression has barely been tapped.
And current deep classifiers?
• Hard to interpret
• Poor at uncertainty
• Don’t understand distribution geometry
We help fix all three.
Diffusion models are great generators—but their potential for classification & regression has barely been tapped.
And current deep classifiers?
• Hard to interpret
• Poor at uncertainty
• Don’t understand distribution geometry
We help fix all three.
December 2, 2025 at 1:29 AM
Why?
Diffusion models are great generators—but their potential for classification & regression has barely been tapped.
And current deep classifiers?
• Hard to interpret
• Poor at uncertainty
• Don’t understand distribution geometry
We help fix all three.
Diffusion models are great generators—but their potential for classification & regression has barely been tapped.
And current deep classifiers?
• Hard to interpret
• Poor at uncertainty
• Don’t understand distribution geometry
We help fix all three.
Researchers usually target genes that are highly over or under expressed in variant (disease) cells relative to baseline. We go beyond such differential expression analysis to probabilistically match perturbation-induced transcriptional shifts to measured changes from healthy and diseased tissue.
September 23, 2025 at 2:12 PM
Researchers usually target genes that are highly over or under expressed in variant (disease) cells relative to baseline. We go beyond such differential expression analysis to probabilistically match perturbation-induced transcriptional shifts to measured changes from healthy and diseased tissue.
Reposted by Ben Kuznets-Speck
This is not just HHMI. All Chan Zuckerberg Initiative funding calls are CLOSED. The Gates Foundation has no Grand Challenges funding opportunities. I've never seen anything like this. Private foundations have just decided to stop supporting biomedical research.
August 9, 2025 at 4:50 PM
This is not just HHMI. All Chan Zuckerberg Initiative funding calls are CLOSED. The Gates Foundation has no Grand Challenges funding opportunities. I've never seen anything like this. Private foundations have just decided to stop supporting biomedical research.
That's an interesting question! At this stage we focused on RNAseq data but it would be a cool next step to incorporate information about chromatin state!
July 12, 2025 at 5:20 PM
That's an interesting question! At this stage we focused on RNAseq data but it would be a cool next step to incorporate information about chromatin state!