Seyone Chithrananda

Really excited to finally share this work from my internship at @MSFTResearch in summer 2023! This project began when Judith and Kevin raised a simple question: can we develop better latent mappings of olfactory stimuli, such as molecules to odor perception? Summary (1/12):

@AnjneyMidha @CS153Systems @Josh_Gilligan spotted in the wild

one of the @CS153Systems students (grad student in bioengineering) and i were walking in the hallway today after class ‘hey anj, thx for doing this class. the optimism is pretty cool’ he sort or conceded felt good! hard to get emotions out of these genius kids sometimes

Make Canada so compelling that the best and brightest want to net-immigrate, not flee. Let's stick to good ideas, of which Pichette has plenty. Making Canada a cage is not one of them.

@melkuo this seems like addressing a symptom not a problem… I agree that subsidizing tuition for postgrad is counterproductive given the earning power new grads command in the US, but holding back people from going where they can do the best work in their field is 🤷‍♂️

YOU HEARD IT FROM MELKUO FIRST

The size of new DNA sequences that can be integrated into the human genome is a foundational constraint for engineering and enhancing human cells. In a new collaborative study, we’ve now almost doubled the maximum size of DNA sequences that can be efficiently inserted into primary human cells. biorxiv.org/content/10.648…

@heeney_luke Based

Everyone taking CS 153. Only 3 people in my Stanford functional analysis class today. Remember to eat your veggies.

@zachweinberg is 100% on the money here FDA also largely doesn’t inspect China cGMP manufacturing sites - experimental, branded, and generic drugs! But, at your US cGMP facility? F-up? Start praying

I am so excited to share our new paper in @Nature: the first programmable, site-specific integration of a large DNA payload into T cells in vivo. A single IV injection results in therapeutic levels of TRAC-targeted CAR T cells in multiple models. #Ack1" target="_blank" rel="nofollow noopener">nature.com/articles/s4158… a 🧵

Today we report single-cell APEX-seq (scAPEX-seq) — a new method for unbiased mapping of *subcellular* transcriptomes at single-cell resolution. This approach reveals cell states that are not detectable by standard scRNA-seq, and enabled us to identify regulators of CAR T function that improve solid tumor killing. biorxiv.org/cgi/content/sh…

Aging may feel gradual… but what if it’s not? In our paper out today, we tracked fish continuously from puberty until death. This gave us a unique view of how aging unfolds across the adult lifespan. 🧵

Very happy to see this work from Fabian Rehm, Jason Chin, and team on the development of a high error-rate orthogonal DNA replication system in E. coli, thus supporting continuous hypermutation and evolution of target genes in vivo. The system is based on a protein-primed linear plasmid replication mechanism, which has become a reliable way of realizing the orthogonal replication concept through which hypermutation is durably targeted to an orthogonal plasmid while sparing the genome. Our original orthogonal DNA replication (OrthoRep) system in yeast, Rongzhen Tian's and Jason Chin's BacORep and EcORep systems in bacteria, and now this promising new system all repurpose protein-primed replication to achieve orthogonality. I look forward to seeing how these systems continue to be applied to gene and biomolecular evolution at scale! nature.com/articles/s4158…

An orthogonal polymerase for hypermutation in E. coli. Often, we want to mutate genes really quickly to make better variants. Jason Chin’s group created a molecular tool to increase E. coli mutation rates by six orders of magnitude. First, they took two genes from bacteriophage Φ29, a virus that does not naturally infect E. coli. One gene primes the cell for DNA replication (called the “terminal protein”), and the other copies the DNA (called “polymerase”). They engineered the latter protein to make *more* mistakes than it naturally would. This polymerase, as it runs along a strand of DNA, has an error rate of about 10⁻⁴ per base per generation, meaning it makes one error every 1,000 bases of DNA, every ten cell divisions. (This mutation rate is about one-million-fold higher than the natural error rate in E. coli.) Importantly, this error-prone polymerase will only copy linear DNA that has been introduced into the cell by the scientist; it doesn’t touch the host cell’s genome. But why is this needed? Can’t we just mutate genes in a test tube instead? Yes, of course! But there are two big advantages with doing hypermutations inside *living cells*: 1. Scale: A small tube contains billions of cells, each running its own mutation experiments in parallel. 2. Selection: Living cells act as a sort of “auto-selection” mechanism. The very fact that they are “alive” tells us that a mutated gene is “acceptable.” If you evolve a gene that works really well *in vitro* and then try putting it into a living cell, it might kill the cell! But if *E. coli* makes mutations itself, and continues growing, then you already know that the mutation did not kill the cell. This is valuable info. Of course, this new paper isn’t the first to do orthogonal hypermutations in living cells. Chang Liu’s laboratory, at UC Irvine, made a similar tool, called OrthoRep, many years ago. But OrthoRep only works in yeast. And the nice thing about using E. coli is that they can grow to much higher densities in tubes and, thus, you can run more evolution experiments in parallel. Last year, Peter Schultz’s group also reported a continuous hypermutation tool for E. coli, using a T7 polymerase and replisome. Their system reportedly introduced mutations ~100,000-times faster than the normal mutation rate. But the problem with their approach is that it seems to leak with the genome; it kills cells from time to time. This new tool appears to be fully orthogonal, doesn’t damage the host genome, works in E. coli, has a very high mutation rate, and only requires two genes to function (so it’s lightweight and thus easy to engineer.) We can do so many things with this. We could evolve thousands of nanobodies in parallel, screening each for binding affinity and stability and then using the data to train predictive models. Or, we could evolve entire biosynthetic pathways to optimize chemical production. Etc etc.

@bilaltwovec so exciting to see you ended up at iso!

working towards solving all disease with some brilliant colleagues has been a very rewarding experience, the future looks very exciting (and healthy)! 🔬🚀

Very disappointing. The Science Center was a pillar for every kid When we destroy places for kids to enrich themselves, we are quite literally destroying our future

Today I'm sharing a preprint on conformal risk control for non-monotonic losses, a paper three years in the making. The key idea: validity of conformal can be reframed as a consequence of algorithmic stability. Therefore, any stable algorithm inherits a conformal guarantee. 🧵

Can we simulate realistic evolutionary trajectories and “replay the tape of life”? In this work, we propose a flexible, generalizable framework for modeling how the entire protein seq evolves over time while capturing complex interactions across sites. 1/n doi.org/10.64898/2026.…

A major work from our lab is out in Science! @ScienceMag, another key step towards a true demonstration of RNA self-replication. science.org/doi/10.1126/sc… (1/6)

AI is cool and all... but a new paper in @ScienceMagazine kind of figured out the origin of life? The paper reports the discovery of a simple 45-nucleotide RNA molecule that can perfectly copy itself.

Not sure why there’s push back here. Stockholm Syndrome? You don’t want bright people doing rote pipetting in the same way you don’t want computational biologists to rewrite RNA-seq pipelines. Working on the higher level abstractions is way more fun and intellectually rewarding.

Continual learning from natural language is data-hungry. Can we make it sample-efficient? SIEVE distills natural language context (instructions, feedback, rules, etc.) into model weights using as few as 3 examples only of queries—outperforming prior methods and even in-context learning baselines. (1/n)