Genomics immunology ML = personalized cancer immunotherapy | pirl.unc.edu | alexr.bsky.social
Joined June 2007
- Tweets 22,363
- Following 4,855
- Followers 6,951
- Likes 100,179
1,063 Photos and videos
Pinned Tweet
Apr 9
The thing I feel most about coding agents is tremendous gratitude. Every half-finished project that has haunted me for a decade gets to actually come to life
15
2,024
salmon 2.0 (the rust re-write) is now live! It's also the version you get if you install via bioconda.
This is a from-scratch rewrite in rust, and the base for future development. I'll post a more detailed thread tomorrow, but you can read the docs here combine-lab.github.io/salmon…
2
9
447
I’ve been obsessed with KLK2 as a prostate cancer target
KLK2 was known to be a great diagnostic marker for 2 decades (~1990 - 2010s) …
but its therapeutic potential was ignored until @JNJNews started building a KLK2 empire in the last decade
open.substack.com/pub/katkaj…
4
6
57
9,302
alex rubinsteyn retweeted
Jun 11
New paper! How do RNAs "know" where to go inside a cell? We dug into the sequence elements that route RNAs to the right place. It turns out that, in mammals, they're surprisingly massive (>200 nt), multipartite, and wonderfully complicated. 🧵
15
125
575
60,072
alex rubinsteyn retweeted
Jun 11
as an example of how LLMs have fundamentally altered technical essays: all three of these failed trials have *extremely* sparse coverage outside of singular academic papers or SEC filings. if you are not a cancer-specific investor or curious oncologist, it is unlikely you’d ever be able to discover the existence of them. without the models, i certainly never would’ve found them!
and yet, their existence should dramatically alter any conclusions you have on neoantigen cancer vaccines; leaving them out would be missing a lot of the story
my point is that the amount of color you can inject into writing is now almost unlimited. the full record of human scientific inquiry is at your fingertips, and you now have the capacity to contextualize fields in a matter of days to weeks rather than years. use it!
Jun 8
How to build a cancer vaccine, and whether they will work this time
owlposting.com/p/how-to-buil…
for the past forty years, the medical field has attempted to create a 'cancer vaccine'. by and large, this effort has not been particularly successful. but things seem different this time around. there is, if you pay attention, a dizzying amount of optimism in the air regarding these potential miracle drugs. is it true? have we arrived? are cancer vaccines just over the horizon?
maybe! in this 8.3k word essay, i walk through the immunological theory behind cancer vaccines, the variants they come in, the (many) confusing aspects behind their few clinical successes, how machine-learning plays a role in their creation, and what lies ahead in their development
it's a very complicated subject, and i am deeply grateful to @iskander and @BenjaminGVincen for patiently answering my questions over the subject!
3
16
133
11,604
alex rubinsteyn retweeted
Jun 10
Best of luck @iskander!
ICYMI, listen to our episode on Geron here:
Jun 10
Starting a newco around unlocking immortality and curing all cancer by modulating the Hayflick limit.
Raising at $2b, send term sheets to DMs
(yes, you should listen to Approved pod)
1
1
1
742
Jun 10
Starting a newco around unlocking immortality and curing all cancer by modulating the Hayflick limit.
Raising at $2b, send term sheets to DMs
(yes, you should listen to Approved pod)
4
2
35
4,032
Jun 10
I asked Fable "how is babby formed" and you'll never guess what happened next
2
12
2,203
Jun 10
This model release is like one of those Jane Street recruiting challenges but the goal is to get Fable to answer any question at all
9
418
Jun 9
MHCflurry 2.3.0rc1 is on PyPI!
It’s a very big change set that moves as much work as possible out of NumPy / Pandas and into PyTorch. It also adds multi-GPU support for both training and inference. Training time on a 8xA100 machine is about an hour (down from a week)
1
3
27
2,003
Jun 9
No new models yet but getting consistently better evals in experiments. For now it will still work with previously released weights.
Thanks to @nvidia / @AllThingsApx for letting me burn a lot of Brev credits.
And if you’re feeling bold, go find some bugs
8
389
Jun 9
Asked Fable a simple question about signal peptides and got:
"Fable 5's safety measures flagged this message for cybersecurity or biology topics"
Not looking very useful so far
2
1
22
975
alex rubinsteyn retweeted
The Gospel of Wealth
Andrew Carnegie
carnegie.org/publication/the…
5 West 51st Street - New York City • 1889
1
2
3
253
alex rubinsteyn retweeted
Jun 8
How to build a cancer vaccine, and whether they will work this time
owlposting.com/p/how-to-buil…
for the past forty years, the medical field has attempted to create a 'cancer vaccine'. by and large, this effort has not been particularly successful. but things seem different this time around. there is, if you pay attention, a dizzying amount of optimism in the air regarding these potential miracle drugs. is it true? have we arrived? are cancer vaccines just over the horizon?
maybe! in this 8.3k word essay, i walk through the immunological theory behind cancer vaccines, the variants they come in, the (many) confusing aspects behind their few clinical successes, how machine-learning plays a role in their creation, and what lies ahead in their development
it's a very complicated subject, and i am deeply grateful to @iskander and @BenjaminGVincen for patiently answering my questions over the subject!
8
32
165
35,326
Jun 4
I've been hand-waving a lot about how under-appreciated cancer-testis antigens are as an alt target space for personalized immunotherapies.
Starting to put numbers on potential impact, eg how many patients have high expression of 1 CTA.
first attempt: ~1/3 of all cancers
3
7
39
4,662
Jun 4
TMB vs % patients with high expression of at least one CTA
They're nice extra antigens in eg melanoma and lung cancer, but might be the only source of TCR targets in eg basal breast cancer
2
533
Jun 4
TGCT = _ tumor
100%
testicular germ cell
0%
tenosynovial giant cell
0%
wut
3 votes • Final results
1
399
Jun 4
FATE1 is a worse CTA than I expected, XAGE1 is much more common than I realized, and PRAME really rules the roost -- highly expressed across many patients in many cancer types.
Besides HPV, if you had to point a prophylactic cancer vaccine at any target I'd pick PRAME
3
2
17
2,173
Jun 4
Red = fraction of patients within a cancer type with PRAME expression >= 25TPM
This is actually an under-count since if a patient has multiple CTAs I count them towards whichever is most common in that cancer type (so eg LUAD patients often XAGE1 and PRAME, but XAGE1 wins)
3
4
873
Jun 4
If you filter more stringently (50TPM) you still get a lot of PRAME and similar cancer types w/ >=75% CTA patients:
synovial sarcoma, melanoma, uterine carcinosarcoma, germ cell, rhabdomyosarcoma, uterine/endometrial/ovarian, lung {AD|SC}, &c
1
604