alex rubinsteyn

Genomics + immunology + ML = personalized cancer immunotherapy | |

[alex rubinsteyn @iskander]

Slides from a talk I gave in September at the Lunenfeld-Tanenbaum Research Institute in Toronto.

Machines of Loving Hype: Let's get specific about "curing cancer with AI"

(secretly, personalized TCR therapies for NUT carcinoma and beyond)

[Citrini @citrini]

There’s a really pronounced divide between people who are on the AI side of things and people who are on the biology side of things. Even among people working in AI, the divide persists based on where they’re focused.

The AI side, adherents of scaling laws and the bitter lesson, believe that AI getting better means AI being more capable of novel discoveries in medicine (that go beyond discovering new drugs in well understood targets or speeding up dry lab work etc).

The biology people think the idea of a bitter lesson in biology is kind of ridiculous and that regardless of AI’s capabilities it’s going to be at least 5-10 years until you could hope to reliably simulate human toxicology in silico (right now AI/ML can’t even reliably simulate in vivo studies like binding affinities or functional assays).

Even the most AI pilled of them, who will list all the areas that AI can reduce the time spent on various areas, don’t think the concept of AI discovering novel targets or generating new understanding of causal biology is realistic in the near term. Most, in fact, will readily characterize the assumption that even an “ASI” level model could find cures hidden in our current understanding (without expanding that understanding with significantly more biological data) as patently false.

This was a big part of why we wanted to experience it ourselves by using AI for a (very) small portion of what it can do in biology. It was both mind blowing and revealed some poignant insights into how different the problem of making AI better at something like coding is to the problems of biology.

Even if it unlocked the ability to generate new molecules at a stunning pace, its simulation (in silico) of binding only had 49% accuracy. And if we wanted to improve that, we’d need to feed it back the data from the lab - data that took about 30x as long to get as it took to generate the molecules. And these were tests in a test tube (turn of phrase), imagine how long it would take if we wanted to improve the model with data that required recruiting humans and monitoring their reactions.

Human biology is very complex, and adding a drug - especially a novel drug and doubly so for a novel drug acting on a novel mechanism - means mapping interactions that have nth-order-effects that we don’t have the data to predict. Virtual cells have made advancements enabled by AI, but the perturbations equate to a small fraction of the possible.

This is before you even get into curing disease, especially uniquely human diseases. Brain function involves cellular identities that are unique to humans, meaning animal studies are of little use.

How long does it take to understand what the side effects of chronic administration of a drug in humans are? AI can speed up analyzing the data, or designing the trial or even getting it approved…but it can’t make a 6 month trial complete in 3 months.

I think its important to point out that significant advancements can be made in AI that contribute to biology, while also being realistic about how much it speeds up something as complex as curing a specific cancer or diseases of the brain.

[Jake Wintermute 🧬/acc @SynBio1]

Chloroplast genomes are the natural next step for generative DNA design

- 200kb is challenging yet possible for DNA assembly
- Less of a PR headache than bacteriophage
- Path to high-impact improvements in photosynthesis

https://t.co/M3RBsy3sjw

[alex rubinsteyn @iskander]

Surreal to explain landlines to my kids in the same register as dinosaurs and the Roman Empire

[Shriram Krishnamurthi (primary: Bluesky) @ShriramKMurthi]

We should maximize the amount of proof we can get out of AI while the money cannon is on. It's like early days of Uber, except the benefits are 100x more tangible. A fully-verified hw+sw stack? That's like gold everyone can share w/out having to divide. Generational opportunity.

[alex rubinsteyn @iskander]

@ac_palmer @5_utr So seems like an interesting and kinda useful biomarker? Follow-up should be trial on continued therapy (maybe switch agent?) for ctDNA+?

[alex rubinsteyn @iskander]

@ac_palmer @5_utr Do any patients stop therapy bc of progression?

[The ODIN @TheODINInc]

This is for everyone who has ever looked at a DNA test and wanted the actual damn data.

Order yours:
https://t.co/aQArNIkte5

US shipping only for now.

[zeta @zeta_globin]

okay I’m just going to rapid fire this because I’ve gotta sleep and also this is just like talking as if you were a close friend:
-like half of bio is ~wrong, like the problem is us not AI- misleading in vitro models, misleading murine models and then all of our “rules” are highly context dependent and not applicable in at least 20% of in vivo contexts
-also everything is baked with people and politics which leads to certain narratives/models in the paradigm sense winning when they maybe shouldn’t- publishing is beyond corrupt, holding people’s livelihood ransom based on whether they generate positive results and hide the negative is just a bad system
-lit search synthesis by AI is therefore not at all close to god, just sort of a condensation of issues
-but it’s still good and bad- like astra max got thymic involution at peak human adolescence wrong on monday but I’ll have grad students use it for questions I don’t want to be bugged with
-protein folding tells you v little abt function- like timing, location, ptms, nearby molecular species, idrs, local environment etc
-no one is building a model that integrates different echelons of regulation for actual in vivo tissues
-everyone needs to undiscover the cell/discover tissues and organ systems and overall interacting physiology
-human data should be step 1 in everything
-idk as far as experimental stuff/basic research gaps- automation and scaling aren’t issues it’s just like if your goal is solving human health what are you starting at? I would love for RNA biology to be solved first- I’m still biased towards building an actual causative model of life
-I think this shit could all be way more variable based on the individual than anyone is ready to deal with but everyone violently disagrees with me on that- but like my baseline basophil count is in the CML range and I’m am biased as a weirdo
-I hate non-durable single target based therapies in oncology for solid tumors with a passion and in my schizophrenic head I’d rather have biological understandings that allowed us to develop actual curative treatments
-to be positive ML based DNA encoded libraries are absolutely sick and I have new molecular glues in like the picomolar range which means tox will probably be an issue but yolo

[alex rubinsteyn @iskander]

@TensorTwerker Having some real "open the pod bay doors" moments over here.

[alex rubinsteyn @iskander]

Making the agents nervous by downloading panels of viral proteomes

[alex rubinsteyn @iskander]

At every biorisk filter I downgrade a model, and went all the way from Astra down to 5.6 Luna. Not much further to fall.

[Rohan Arun @RohanArun]

BREAKING: We broke the wall for Integer Multiplication!

Conditional κ = 0.00302 (+80%)

For days we were stuck in incremental improvements, but today zxv changed the game.

zxv’s insight: stop throwing away work. Helper registers that finish holding a completed transform of their own input now count as output, instead of being undone.

He brought us to 0.0023(+37.5%), and immediately after 1337isnot1337 pushed that to 0.00302.

The game is afoot :)

[Ramith Hettiarachchi @ramith__]

did a group presentation today with @3blue1brown's manim style transitions. Sharing the template for anyone to build upon

Brought to life with @NarekMaloyan's manim-web + reveal.js + claude + codex

view template: https://t.co/LOSKmrtlXa
view repo: https://t.co/gTb9xxYp7F

[owl @owl_posting]

https://t.co/XlNv16ayVS

the owl posting podcast portfolio is incredible. congratulations to coherence, insanely talented team

[Cheshire @Thatsregrettab1]

https://t.co/S5T4icjMcz

From James Heathers

[Dr. Anthony Letai @NCIDirector]

If your Author Accepted Manuscript was supported by @NIH funding, it must be posted publicly to PMC without an embargo. There's no cost and it does not require buying expensive open access plans. Take steps to upload your Author Accepted Manuscript here: https://t.co/mrmZO46sy4

[Miles Cranmer @MilesCranmer]

Two years of work, and PySR v2 is finally out!

If you haven't used it before: PySR is a machine learning library that finds short equations that fit your data.

I've always wanted a PyTorch for symbolic learning. PySR v2 is a big step towards that:

[Minhua Chu @chuminhua432]

🇨🇳 China IND Tracker (October 12–18)

16 novel drugs received their first IND clearances in China during the week, according to China’s Center for Drug Evaluation (CDE) disclosures.

Notable INDs:

1. $ANL Adlai Nortye (阿诺生物) — AN4035 Injection
MOA: CEACAM5-targeting ADC conjugated with a highly potent pan-RAS(ON) inhibitor
Indication: Advanced or metastatic solid tumors with RAS mutations and enriched CEACAM5 expression, including colorectal cancer, pancreatic ductal adenocarcinoma, and non-small cell lung cancer

2. Yuekang Pharma (悦康药业) — YKYY036 Injection
MOA: INHBE-targeting siRNA conjugated with GalNAc ligand
Indication: Obesity or overweight

3. Precision Biotech (精准生物) — TC-13-183 Cell Therapy
MOA: Cell therapy
Indication: CEA-positive advanced peritoneal metastatic colorectal cancer

4. CSPC (石药集团) — SYS6088 Injection
MOA: PD-1/VEGFA/CTLA-4 trispecific antibody
Indication: Advanced or metastatic solid tumors

5. Qilu Pharma (齐鲁制药) — QLS2401 Injection
MOA: PSMA/STEAP1/CD3 trispecific antibody
Indication: Advanced prostate cancer

6. Institute of Biophysics, Chinese Academy of Sciences / Sinovac Biotech (中科院生物物理所/科兴集团) — SNA01 Injection
MOA: Long-acting anti-RSV mab for intramuscular administration
Indication: Prevention of RSV lower respiratory tract infections in neonates and infants

7. Essight Biotech (灵识质源) — ES502 Injection
MOA: RAS G12V-targeting TCR-TCE
Indication: RAS G12V positive advanced solid tumors

8. Chipscreen (微芯生物) — NW001 for Injection
MOA: Epigenetic immune-modulating antibody-drug conjugate (EIMAC), comprising a humanized anti-PD-L1 mab conjugated to an engineered HDACi payload
Indication: Advanced malignancies, including solid tumors and lymphomas

9. Innostellar Biotherapeutics (朗昇生物) — LX111 Injection
MOA: Gene therapy
Indication: RVO-ME and DME

10. Metis TechBio (剂泰科技) — MTS109 Injection
MOA: mRNA encoding a CD3×CD19×BCMA trispecific T-cell engager (TCE), delivered via LNPs
Indication: SLE

[Paolo Tarantino @PTarantinoMD]

We now have prospective data with TOPO1 ADC sequencing from:
- SATEEN (#ESMOBreast26)
- HER3-DXd (#ESMOBreast26)
- SWITCH (upcoming #ESMO26)

All validate prior retrospective data, showing that TOPO1 ADC resistance is not overcome by target switch.

New payloads urgently needed.