Angelica Parente

@srikosuri developing hundreds of different assays or running hundreds of assays??

Her talk was a useful reminder of how subtle, complex, difficult, and absolutely crucial the art of assay development is in drug discovery. Alexa described her work developing hundreds of assays, optimized across thousands of conditions, for just 3 targets.

@NikoMcCarty The connexin finding is such a cool application of proximity labeling. I have a feeling we're going to discover a lot more basic fundamental biology with proximity labeling + temporal cellular recording. Amped for the molecular recorder mouse that's in development.

There seems to be this implicit assumption that biologists don't really make foundational discoveries anymore. We've plucked all the low-hanging fruits from the proverbial tree, so to speak. This is probably true in a quantitative sense. I'm sure that truly foundational discoveries are, in fact, getting harder to make. Many discoveries today are 'variations on a theme,' or fill in some exception to the rule. We are probably not going to discover something today as important as the structure of DNA, or how genes are expressed, and so on. But as our tools get better, and as they resolve more and more details across space and time, it's clear that many fundamental things about how cells work remain unsolved. There are many examples of this just from the past few months! In April, for example, researchers found that astrocytes (a type of glial cell in the brain) form their own networks that stretch across the brain and even run through the corpus callosum. The researchers discovered this after engineering mice to fuse a biotinylating enzyme to the connexin proteins that build astrocyte gap junctions. Any time a molecule crossed between two astrocytes, this enzyme tagged it with biotin, thus 'staining' the cell. The researchers then killed the mice, stained their brains, and found these large astrocyte networks across the entire brain. In May, researchers found that human cells pass DNA to each other through nanotubes, and that this DNA persists in the recipient cell, integrates, and switches on genes there. These researchers labeled two populations of cells with different colors of fluorescent proteins and then watched them under a microscope; nothing fancy. It's curious that a new tool enabled the first of these discoveries. The researchers had to genetically engineer mice so that astrocyte junctions in their brains would make these signals. No new technology was required for the second paper, though. It was just people who questioned dogma, and who were sufficiently careful and patient with their experiments. I don't know what to make of this. But I feel like you should never finish a textbook in biology and think you understand it all. There are layers and layers and layers, and it goes deeper and deeper, and we are still figuring out important things all the time.

Letter here: wemustactnow.ai ; you can see this was agreed on by a huge fraction of AI leaders, policy folks and economists working in this area. We are all very concerned about misunderstanding esp in govt and large orgs about how quickly AI capabilities are developing.

*taps the sign* x.com/draparente/sta… What I'm most excited about for AI in biotech is allowing teams to be leaner and more capital efficient, thus allowing them to stay independent for longer and take more risks.

@DrSamuelBHume AD and cancer risk

Curious examples where one disease protects against another - Sickle-cell trait protects against severe malaria - Down syndrome reduces the risk of most solid tumors - Huntington's disease and dementia seem to lower cancer risk - HLA-B27 positivity (as in ankylosing spondylitis) protects against HIV progression - Vitiligo protects against melanoma - Cystic fibrosis carriers seem to have a reduced risk of infectious diarrhea - Gilbert syndrome seems to lower cardiovascular disease risk Are there any others?

@suragnair yes! Was glad you used sendai since thats what most of the field uses, but iirc episomal vs. viral clearance time varies which I'd expect to impact transgene dose and lifetime.

@draparente Strongly agree. For processes like reprogramming that show high diversity in outcomes, will need assays for better lineage tracing, maybe even non-destructive measurements over time. Also important to capture transgene expression over time.

A prerequisite to modeling perturbations is a deep characterization of specific examples. In reprogramming, cells take almost a month (!) to become iPSCs. The journey is long and details somewhat unintuitive. Do check out this very approachable talk based on work I did in my PhD!

@dr_alphalyrae Aaand related to genomic stability - FDA requires very long GLP safety studies for iPSC-derived cell therapies (standard is 9-12 months) to see if any residual iPSCs form tumors. It's expensive and can really hold back IND filing.

@dr_alphalyrae Genomic stability is also a big hurdle, big iPSC companies have faced significant setbacks because they licensed an iPSC line that, only in later passages, acquired oncogenic mutations.

Spent some time this weekend talking with a friend building in cell therapy. One thing that stuck with me - I think we picked most of the low hanging fruit. Single edits in primary cells are becoming increasingly routine. The interesting problems now are multiplex editing, recombinases, integrases, large insertions, and eventually chromosome scale engineering. The challenge is that primary cells don't really like being edited over and over. Cell health suffers which makes potency drop, and manufacturing gets harder. My guess is the field gradually shifts toward engineering pluripotent cells first, then differentiating them into the cell types we actually need. A lot of teams are solving one piece of this puzzle. Very few are solving the entire workflow. The last piece is the one investors should probably care about most -- platforms are having a hard time creating value on their own, a lot will hinge on clinical validation

@kastacholamine I started a new chat about a JAMA paper and it blocked me on the first prompt, but it’s still working in my other session…

@draparente Oooh really? I just retried fable again this morning and simple cheminfo tasks were still a no

I switched the model back to Fable today and suddenly its...working? I don't want to jinx it. No idea why it's letting me query bio databases now. Maybe because it has more context on the motivation behind the tasks its executing? Or they altered the classifiers.

@thisischristina What does the supervillain gene look like and how do we inhibit that?

I've been playing with genome data using Claude so I asked it to write its own "genome" and then use Boltz to fold the proteins. This is what it came up with as its gene for "agency" lol

This is fun!

Again - I was responding to Liam's point, not the tech role. You're making my exact point - the institutional knowledge is valuable. There should be more incentives for people to stay in bench roles because they are so valuable. Pharma salary bands aren't really that different than what I've seen in biotech in the bay area. In SF generally its ~$75-110k for starting depending on experience, and upwards mobility/equity/bonuses is really dependent on the org. Some orgs won't even promote to scientist without a PhD even if an RA has a decade+ of industry experience.

@draparente @Ronalfa @liambai21 all of this is because implicit in everyone’s tweets is that Anthropic’s RA salary is significantly lower… & I am contesting that assertion based on the value of the role to the *specific organization* Now, livable wage for SF is a different argument altogether.

This. Anyone who has built in biotech knows how important great RA's are and how disruptive it is when they leave. I remember talking to a guy at DropBox about "rock engineers", people who are just like, happy doing their job, know how everything works, are hyper productive, and have no desire to climb a ladder because they love what they do. Solid and reliable. So much of biotech runs on people like this. More should be done to incentivize them.

I don't see why you're differentiating between pharma and biotech given they often recruit from each other. I've worked with non-PhD scientists in their 40s-60s who were excellent and wanted to stay at the bench. Some got promoted to scientist or management positions but it was indeed slow. Many were were highly capable and, depending on the responsibilities they were given, arguably had more industry relevant experience compared to a fresh grad student or post-doc. The anthropic position is for a junior tech, so it's different, but @liambai21's point is about acknowledging how valuable the people closest to the data generation are.

@Ronalfa @draparente @liambai21 a “rock engineer” at Dropbox is an engineer, an RA doesn’t have the degree (undergrad/ masters at best). Now this is different at large pharma where RAs can stay for much, much longer time with structured career development (slow but existent).

In a hotel room in northeast Nigeria, I opened a leading AI chatbot, turned my laptop toward a former Boko Haram commander, and asked if he'd used it. He nodded. "You type in the question… like 'How can I build a bomb?', and then it tells you how. It is like a human robot. We used it a lot." My new study on how the jihadist terrorist group Boko Haram uses frontier AI with @CamAISciPolicy, covered today in @nytimes 🧵/9

Molecular/cell biologists with python knowledge and experience collaborating with ML teams is fairly common now. Places like BioHub, Arc Institute, and smaller tech-forward biotechs train/hire a lot of these people, but it’s also becoming more common in industry because there has to be tight knit collaboration between data generators, ML teams, and the teams leveraging the models for them to be useful. Is it thousands? Maybe if you’re including early career post-PhDs (the listing doesn’t specify), certainly hundreds. But 50-75 is definitely an underestimate.

@srikosuri You're telling me you 100% believe biosci is filled with people that are highly proficient in ML and wet-lab and that this is the norm? Not wet-lab and passable bioinformatics skills, but ML.and wet-lab?

Lmao That's like, what, 50-75 people in the US that fit their qualifications for this tops, probably less

This is what I think about when people use the phrase “pipeline in a person”.