Daniel Asarnow
@dabiophysicist.bsky.social
Postdoctoral scholar in Veesler Lab @ UWa.
Former UCSF & SFSU & UCSC student & Exploratorium Explainer. He/him.
I take tiny pictures, in 2 & 3 dimensions.
https://github.com/asarnow
Former UCSF & SFSU & UCSC student & Exploratorium Explainer. He/him.
I take tiny pictures, in 2 & 3 dimensions.
https://github.com/asarnow
Yes, definitely. Based on what you share here and the applications I've seen succeed and fail, I believe you would be helping students!
January 21, 2026 at 7:20 PM
Yes, definitely. Based on what you share here and the applications I've seen succeed and fail, I believe you would be helping students!
Btw that antibody (pembro) also worked as a single agent, but another that failed out actually proved equally effective in combination therapies (nivo). Which is to say, the key might be finding the right extra switches to flip, in the other 75% of patients
January 21, 2026 at 6:20 PM
Btw that antibody (pembro) also worked as a single agent, but another that failed out actually proved equally effective in combination therapies (nivo). Which is to say, the key might be finding the right extra switches to flip, in the other 75% of patients
In one of the first immunotherapy trials, for stage 2 non-small cell lung cancer, the complete response rate ("cures") went from ~2% with chemo to ~24% with chemo+immunotherapy. An amazing achievement, but one that's immediately tantalizing... we don't need another 10-fold improvement, just 3
January 21, 2026 at 6:20 PM
In one of the first immunotherapy trials, for stage 2 non-small cell lung cancer, the complete response rate ("cures") went from ~2% with chemo to ~24% with chemo+immunotherapy. An amazing achievement, but one that's immediately tantalizing... we don't need another 10-fold improvement, just 3
So that's how it got here! 1st pic was our social hour poster after the paper came out. I usually imagine proteins at the scale of buildings, but this one in particular really sells the impression
January 19, 2026 at 1:53 AM
So that's how it got here! 1st pic was our social hour poster after the paper came out. I usually imagine proteins at the scale of buildings, but this one in particular really sells the impression
That's how their group meetings felt!
January 19, 2026 at 1:46 AM
That's how their group meetings felt!
Google Scholar (Labs) has a LLM powered search mode that's quite decent if you haven't tried it. Not a chatbot, just takes an unstructured query and returns real documents from the database (plus ~one sentence why they were selected)
January 18, 2026 at 1:52 AM
Google Scholar (Labs) has a LLM powered search mode that's quite decent if you haven't tried it. Not a chatbot, just takes an unstructured query and returns real documents from the database (plus ~one sentence why they were selected)
Great thread, thanks!
Btw the speed of light is just about 1 foot per nanosecond, which makes both a little more relatable I think
Btw the speed of light is just about 1 foot per nanosecond, which makes both a little more relatable I think
January 17, 2026 at 7:20 PM
Great thread, thanks!
Btw the speed of light is just about 1 foot per nanosecond, which makes both a little more relatable I think
Btw the speed of light is just about 1 foot per nanosecond, which makes both a little more relatable I think
I understand there are certain limits 😇
January 17, 2026 at 8:49 AM
I understand there are certain limits 😇
It's *so* cool. Most prints are way too small, so they're like rocks. Yours makes it start to come alive
January 17, 2026 at 8:27 AM
It's *so* cool. Most prints are way too small, so they're like rocks. Yours makes it start to come alive
To be clear this study has nothing to do with LLM text generation. It's about eg classifying geological specimens with SVMs, segmenting micrograph backgrounds with a CNN, or using PCA on microarray data.
January 15, 2026 at 8:50 PM
To be clear this study has nothing to do with LLM text generation. It's about eg classifying geological specimens with SVMs, segmenting micrograph backgrounds with a CNN, or using PCA on microarray data.
The vast majority of "AI augmentation" would also have served to automate existing analysis, especially until recently (but probably still does)
January 15, 2026 at 8:26 PM
The vast majority of "AI augmentation" would also have served to automate existing analysis, especially until recently (but probably still does)
Sec 2.6 (tokenization) and 3.2 (atom attention) of the AF3 paper SI
static-content.springer.com
January 15, 2026 at 10:17 AM
Sec 2.6 (tokenization) and 3.2 (atom attention) of the AF3 paper SI
It's both, as the standard AA, RNA, and DNA residues have unique tokens, but the 'sequence-local atom attention' learns about local atom constellations without tokenization. Microscopically though the "insights" don't necessarily transfer even for AAs (read this paper if you haven't already).
Investigating whether deep learning models for co-folding learn the physics of protein-ligand interactions - Nature Communications
This study shows that deep-learning-based co-folding models often produce physically unrealistic structures and are unsusceptible in their predictions to significant binding site modifications, demons...
www.nature.com
January 15, 2026 at 10:16 AM
It's both, as the standard AA, RNA, and DNA residues have unique tokens, but the 'sequence-local atom attention' learns about local atom constellations without tokenization. Microscopically though the "insights" don't necessarily transfer even for AAs (read this paper if you haven't already).
The code is my friend. The interpreter is the enemy
January 15, 2026 at 4:09 AM
The code is my friend. The interpreter is the enemy
Same outlet few years back 🙄
New Boom
Baby boomers are no longer the biggest generation in America. Today, it's millennials — young people born between 1980 and 2000.
www.npr.org
January 13, 2026 at 9:24 PM
Same outlet few years back 🙄
But US population isn't shrinking
January 13, 2026 at 9:07 PM
But US population isn't shrinking
One of the best things to do with a GameShark was multiplayer in the single player maps. Sometimes you could even finish games before crashing the console!
January 13, 2026 at 3:13 AM
One of the best things to do with a GameShark was multiplayer in the single player maps. Sometimes you could even finish games before crashing the console!
And structural genomics will finally regain its rightful place!
January 12, 2026 at 11:36 PM
And structural genomics will finally regain its rightful place!
Tbh if I can't skip the video I use the JavaScript console to increase the playback speed
January 12, 2026 at 10:44 PM
Tbh if I can't skip the video I use the JavaScript console to increase the playback speed
At one point the Exploratorium had an exhibit that was a pole with little metal rods coming out horizontally. You put a drinking straw over a rod, then bite down on it, and suddenly hear the music
January 11, 2026 at 6:31 PM
At one point the Exploratorium had an exhibit that was a pole with little metal rods coming out horizontally. You put a drinking straw over a rod, then bite down on it, and suddenly hear the music