Alex Rubinsteyn
@alexr.bsky.social
personalized cancer immunotherapy = genomics + immunology + machine learning + oncology
(pirl.unc.edu)
(pirl.unc.edu)
That’s how I’m thinking about it. A lot of decent SWEs now just yell at Codex all day
I would love to have something vaguely similar with a cloud lab in the loop, even if it’s for a very restricted subset of experiments
I would love to have something vaguely similar with a cloud lab in the loop, even if it’s for a very restricted subset of experiments
November 25, 2025 at 3:37 PM
That’s how I’m thinking about it. A lot of decent SWEs now just yell at Codex all day
I would love to have something vaguely similar with a cloud lab in the loop, even if it’s for a very restricted subset of experiments
I would love to have something vaguely similar with a cloud lab in the loop, even if it’s for a very restricted subset of experiments
Reposted by Alex Rubinsteyn
Ok; mim (github.com/COMBINE-lab/...) preprint submitted! Excited for folks to see it and share thoughts. The key takeaway; mim allows the quick, one-time, building of a small auxiliary index that then allows scaling gzipped FASTQ parsing linearly in # of threads. 1/2
GitHub - COMBINE-lab/mim: A small, auxiliary index to massively improve parallel fastq parsing
A small, auxiliary index to massively improve parallel fastq parsing - COMBINE-lab/mim
github.com
November 25, 2025 at 2:13 PM
Ok; mim (github.com/COMBINE-lab/...) preprint submitted! Excited for folks to see it and share thoughts. The key takeaway; mim allows the quick, one-time, building of a small auxiliary index that then allows scaling gzipped FASTQ parsing linearly in # of threads. 1/2
The bottlenecks are still experimental but getting to a starting point quickly does have a cumulative effect across rounds of experiments
Lots of “find that sequence in a supplement or patent” type tasks just never happen when you’re already constrained by too few people for too many experiments
Lots of “find that sequence in a supplement or patent” type tasks just never happen when you’re already constrained by too few people for too many experiments
November 25, 2025 at 2:27 PM
The bottlenecks are still experimental but getting to a starting point quickly does have a cumulative effect across rounds of experiments
Lots of “find that sequence in a supplement or patent” type tasks just never happen when you’re already constrained by too few people for too many experiments
Lots of “find that sequence in a supplement or patent” type tasks just never happen when you’re already constrained by too few people for too many experiments
Not Anthropic’s models but OpenAI has had a big impact on how quickly our lab can move into new terrain and find default sequence elements to use in constructs, positive/negative controls &c
It’s not accelerating towards singularity but it’s not nothing either
It’s not accelerating towards singularity but it’s not nothing either
November 25, 2025 at 2:25 PM
Not Anthropic’s models but OpenAI has had a big impact on how quickly our lab can move into new terrain and find default sequence elements to use in constructs, positive/negative controls &c
It’s not accelerating towards singularity but it’s not nothing either
It’s not accelerating towards singularity but it’s not nothing either
I must be bad at this, once I’m stuck I seem to stay stuck
(Though for simpler toy / one line “bad” questions I did find translating to Swahili used to work)
(Though for simpler toy / one line “bad” questions I did find translating to Swahili used to work)
November 25, 2025 at 2:59 AM
I must be bad at this, once I’m stuck I seem to stay stuck
(Though for simpler toy / one line “bad” questions I did find translating to Swahili used to work)
(Though for simpler toy / one line “bad” questions I did find translating to Swahili used to work)
It was pretty much SAT redux but was nice to show I was better at math the second time
November 25, 2025 at 2:24 AM
It was pretty much SAT redux but was nice to show I was better at math the second time
Impressions of first real bio deep dive with Gemini 3 — it seems almost on par with ChatGPT 5 Pro, can’t tell if it’s a little worse or I’m not used to it — but definitely grabbing right eg lipid ratios from correct papers but! The suggested next steps are way worse. Weird
November 25, 2025 at 2:10 AM
Impressions of first real bio deep dive with Gemini 3 — it seems almost on par with ChatGPT 5 Pro, can’t tell if it’s a little worse or I’m not used to it — but definitely grabbing right eg lipid ratios from correct papers but! The suggested next steps are way worse. Weird
GRE + CS Subject test helped me get into grad school despite very low undergrad grades, so I’m biased to think they’re great
November 25, 2025 at 2:09 AM
GRE + CS Subject test helped me get into grad school despite very low undergrad grades, so I’m biased to think they’re great
This feels like the worst of all worlds: OK fine, actually measuring preparedness for college is OK for elite institutions but state flagships should let in people who can’t add. Seems like a sure way to massively concentrate prestige and later career success in a few private institutions
November 25, 2025 at 1:45 AM
This feels like the worst of all worlds: OK fine, actually measuring preparedness for college is OK for elite institutions but state flagships should let in people who can’t add. Seems like a sure way to massively concentrate prestige and later career success in a few private institutions
What was your success rate in expressing those pairs?
November 25, 2025 at 1:41 AM
What was your success rate in expressing those pairs?
Yep: this is usually a problem with Claude but almost never encounter it with ChatGPT 5 Pro, it’s actually an excellent aid for lit search (combining disparate corners of literature that would be hard to find by keywords, following complex biological constraints), but today it’s been refusing
November 25, 2025 at 1:40 AM
Yep: this is usually a problem with Claude but almost never encounter it with ChatGPT 5 Pro, it’s actually an excellent aid for lit search (combining disparate corners of literature that would be hard to find by keywords, following complex biological constraints), but today it’s been refusing
Got stuck on a few different experiment design queries with ChatGPT 5 Pro
Finally got the Gemini app and tried Gemini 3 and…it perfectly answered my question with every relevant experimental detail across a few different papers/patents
Finally got the Gemini app and tried Gemini 3 and…it perfectly answered my question with every relevant experimental detail across a few different papers/patents
November 25, 2025 at 12:54 AM
Got stuck on a few different experiment design queries with ChatGPT 5 Pro
Finally got the Gemini app and tried Gemini 3 and…it perfectly answered my question with every relevant experimental detail across a few different papers/patents
Finally got the Gemini app and tried Gemini 3 and…it perfectly answered my question with every relevant experimental detail across a few different papers/patents
But unfortunately:
“Thank you for your interest in the AI-coscientist Trust Test Program. Due to our current capacity, we have now closed the application form. We will scale our capacity over time to accommodate more applicants in the future.
“
“Thank you for your interest in the AI-coscientist Trust Test Program. Due to our current capacity, we have now closed the application form. We will scale our capacity over time to accommodate more applicants in the future.
“
November 24, 2025 at 12:42 PM
But unfortunately:
“Thank you for your interest in the AI-coscientist Trust Test Program. Due to our current capacity, we have now closed the application form. We will scale our capacity over time to accommodate more applicants in the future.
“
“Thank you for your interest in the AI-coscientist Trust Test Program. Due to our current capacity, we have now closed the application form. We will scale our capacity over time to accommodate more applicants in the future.
“
Ironically this reads like GPT 3.5
November 24, 2025 at 12:40 PM
Ironically this reads like GPT 3.5
Oh I haven’t tried or heard of Co-Scientist but it’s the kind of thing I’m looking to test
November 24, 2025 at 12:40 PM
Oh I haven’t tried or heard of Co-Scientist but it’s the kind of thing I’m looking to test
Which AI tools have you tried for science (e.g. lit review or experiment brainstorming)?
Which did you find useful?
Which did you find useful?
November 23, 2025 at 11:45 PM
Which AI tools have you tried for science (e.g. lit review or experiment brainstorming)?
Which did you find useful?
Which did you find useful?
Thought of a decent eval for “AI scientist” / “AI science assistant” tools
Other than Kosmos, ChatGPT {Thinking, Pro}, Claude/Opus, and Gemini;
what else is worth trying for an experimental design question that requires integrating across a few different corners of literature?
Other than Kosmos, ChatGPT {Thinking, Pro}, Claude/Opus, and Gemini;
what else is worth trying for an experimental design question that requires integrating across a few different corners of literature?
November 23, 2025 at 3:41 PM
Thought of a decent eval for “AI scientist” / “AI science assistant” tools
Other than Kosmos, ChatGPT {Thinking, Pro}, Claude/Opus, and Gemini;
what else is worth trying for an experimental design question that requires integrating across a few different corners of literature?
Other than Kosmos, ChatGPT {Thinking, Pro}, Claude/Opus, and Gemini;
what else is worth trying for an experimental design question that requires integrating across a few different corners of literature?
ChatGPT is such a delightful superpower for niche bio research
Needed a constrained kind of binder for a constrained subset of surface proteins on a certain subset of lymphocytes and just asking the sky gods while putting my kids to sleep surfaced a dozen good papers
Needed a constrained kind of binder for a constrained subset of surface proteins on a certain subset of lymphocytes and just asking the sky gods while putting my kids to sleep surfaced a dozen good papers
November 20, 2025 at 11:25 PM
ChatGPT is such a delightful superpower for niche bio research
Needed a constrained kind of binder for a constrained subset of surface proteins on a certain subset of lymphocytes and just asking the sky gods while putting my kids to sleep surfaced a dozen good papers
Needed a constrained kind of binder for a constrained subset of surface proteins on a certain subset of lymphocytes and just asking the sky gods while putting my kids to sleep surfaced a dozen good papers
Schools and hospitals scrambling to keep marauding kidnappers at bay, dark shameful stuff going on
November 20, 2025 at 2:22 PM
Schools and hospitals scrambling to keep marauding kidnappers at bay, dark shameful stuff going on
Met Adam Grippen (first author) yesterday and I think his new analysis fixes a lot of these problems and still points to a real but smaller effect
I've been asked about this paper (www.nature.com/articles/s41...) enough times that I finally had to read it.
The basic claim is that mRNA vaccines had huge synergistic benefit with checkpoint blockade against lung cancer melanoma.
tl;dr it's glam-slop & everyone involved should be embarrassed
The basic claim is that mRNA vaccines had huge synergistic benefit with checkpoint blockade against lung cancer melanoma.
tl;dr it's glam-slop & everyone involved should be embarrassed
SARS-CoV-2 mRNA vaccines sensitize tumours to immune checkpoint blockade - Nature
mRNA vaccines targeting SARS-CoV-2 also sensitize tumours to immune checkpoint inhibitors.
www.nature.com
November 19, 2025 at 12:00 AM
Met Adam Grippen (first author) yesterday and I think his new analysis fixes a lot of these problems and still points to a real but smaller effect
Hard to reconcile the claimed high ex vivo CD8+ T-cell responses of FixVac and its imperceptible clinical activity compared with single epitope TCR-Ts
Either immunogenicity data has some huge catch or there’s something important to understand in the discrepancy
Either immunogenicity data has some huge catch or there’s something important to understand in the discrepancy
November 18, 2025 at 11:55 PM
Hard to reconcile the claimed high ex vivo CD8+ T-cell responses of FixVac and its imperceptible clinical activity compared with single epitope TCR-Ts
Either immunogenicity data has some huge catch or there’s something important to understand in the discrepancy
Either immunogenicity data has some huge catch or there’s something important to understand in the discrepancy
Sure, as long as you accept some failure rate. Maybe 80% success for 25mer peptides from a good manufacturer (e.g. JPT, SciTide, &c). For DNA it probably varies by technology, not sure what % of sequences are high purity manufactured as eg mbDNA. Not as sure about RNA but haven’t an RUO failure yet
November 7, 2025 at 8:55 PM
Sure, as long as you accept some failure rate. Maybe 80% success for 25mer peptides from a good manufacturer (e.g. JPT, SciTide, &c). For DNA it probably varies by technology, not sure what % of sequences are high purity manufactured as eg mbDNA. Not as sure about RNA but haven’t an RUO failure yet
Other than neoantigen vaccines, has any other process been in trials which chooses therapeutic sequences computationally & then uses quick commoditized manufacturing (DNA/RNA/peptides)?
What about personalization through an assay (eg neoAg TCR-Ts, bacteriophage bank, n=1 ASOs)?
What about personalization through an assay (eg neoAg TCR-Ts, bacteriophage bank, n=1 ASOs)?
Automated medicine =
molecular profiling as input data for:
computational target selection ->
computational therapeutic design ->
computational safety screening ->
on demand manufacturing using commoditized modalities (DNA, RNA, peptides)
molecular profiling as input data for:
computational target selection ->
computational therapeutic design ->
computational safety screening ->
on demand manufacturing using commoditized modalities (DNA, RNA, peptides)
November 7, 2025 at 8:07 PM
Other than neoantigen vaccines, has any other process been in trials which chooses therapeutic sequences computationally & then uses quick commoditized manufacturing (DNA/RNA/peptides)?
What about personalization through an assay (eg neoAg TCR-Ts, bacteriophage bank, n=1 ASOs)?
What about personalization through an assay (eg neoAg TCR-Ts, bacteriophage bank, n=1 ASOs)?