←
Pavle Jeremić754 posts
Pavle Jeremić avatar

Pavle Jeremić

@SynBioMars

Building the post scarcity future

◷ Joined February 2014▧ 6 Photos and videos
Pavle Jeremić avatar
Pavle Jeremić@SynBioMars·Nov 16

Anthropic lists Thursday. It bought a biotech and runs a wet lab. We published a result today: the protein designs the structure models were most confident in were the ones that didn’t work. 84% of them didn’t bind anything. Structure confidence does not predict function. A model trained on measured function does. I’m on @MTSLive later today to get into it.

After 6 years in stealth, Aether's first paper is out. We designed and measured hundreds of proteins. The ones that leading models were most confident in were least likely to work: 84% didn't bind. A confident predicted shape is not evidence of function.

Pavle Jeremić avatar
Pavle Jeremić@SynBioMars·Nov 16

After 6 years in stealth, Aether's first paper is out. We designed and measured hundreds of proteins. The ones that leading models were most confident in were least likely to work: 84% didn't bind. A confident predicted shape is not evidence of function.

Jasmin Mujanović avatar
⟳ Pavle Jeremić retweeted
Jasmin Mujanović@JasminMuj·Oct 25

I do not at foreclose the possibility of various sorts of political shenanigans even after Vucic’s concession but this extraordinary result in Serbia is what happens when people show up: in the streets, at the ballot box. That’s the lesson. If you care about democracy, show up.

Pavle Jeremić avatar
Pavle Jeremić@SynBioMars·Aug 6

Biological systems are much more like machines than text, they data that describes them needs to have a lot higher dimensionality

This is a very poor one line summary of the article by @DaphneKoller IMO. The point is not just data volume. The right kind of data scaled in the right axis with the right experimental designs for well defined tasks + using the models to discover biology (not eval maxxing)

Crémieux avatar
⟳ Pavle Jeremić retweeted
Crémieux@cremieuxrecueil·Jul 26

Yes! These people have made building in San Francisco so restrictive that OVER HALF of the homes there now are illegal to build today For SF to get in compliance, 130,860 homes would have to be destroyed This self-inflicted housing shortage is the city's biggest problem—by far

Post attachment

when rent is $8k for a one bedroom, please remember that the fault is in the people who, for 50 years, showed up to argue against nearly every unit of housing proposed these people are no san franciscans

Gleb Kuznetsov avatar
⟳ Pavle Jeremić retweeted
Gleb Kuznetsov@glebkuz·Apr 13

Multiplexing can be thought of as "GPU-ification" of a complex biological system Serial measurement becomes massively parallel To turn biology digital, you need data at *scale* and *relevance* The power of multiplexing is you get both.

Multiplexed in vivo screening is the future of drug development. @ManifoldBio is multiplexing protein therapies in vivo. @GordianBio is multiplexing gene therapies in vivo. @waypointbio is multiplexing cell therapies in vivo. GT Bio is multiplexing LNPs in vivo. 50Y portcos all

Post attachmentPost attachmentPost attachmentPost attachment
Circe avatar
⟳ Pavle Jeremić retweeted
Circe@vocalcry·Apr 11

My HOA is two weeks away from developing a nuclear weapon

Pavle Jeremić avatar
Pavle Jeremić@SynBioMars·Apr 9

Co-design is the future! Any effort to design de-novo functions must include both the protein design and the actual molecular function/target in model, awesome work here further validating that strategy

What if AI could invent enzymes that nature hasn’t seen? 👩‍🔬🧑‍🔬 Introducing 🪩 DISCO: Diffusion for Sequence-structure CO-design 14 rounds of directed evolution and over a year of wet lab work. That's what it took to engineer an enzyme for selective C(sp³)–H insertion, one of the most challenging transformations in organic chemistry. DISCO surpasses this with a single plate. No pre-specified catalytic residues, no template, no theozyme, no inverse folding, just joint diffusion over protein sequence and structure. 📝 Blog: disco-design.github.io/ 📄 Paper: arxiv.org/abs/2604.05181 💻 Code: github.com/DISCO-design/DISC…

0:15
Kane 謝凱堯 avatar
⟳ Pavle Jeremić retweeted
Kane 謝凱堯@kane·Apr 6

San Francisco has failed at infrastructure for so long that reporters will look at a 25 story building and call it "inconceivable" and "sci-fi"

Post attachment

Even the sci-fi film "Big Hero 6" could not envision a San Francisco with tall buildings in the Marina. In the real world, one is likely coming to the Marina Safeway site. No matter what politicians or people calling for a boycott want. From @esksf missionlocal.org/2026/04/mar…

Pavle Jeremić avatar
Pavle Jeremić@SynBioMars·Apr 6

Protein function models are the future! Requires massive multi-functional datasets to generalize not just to “natural” functions but also to push well beyond what nature evolved.

Progress in AI modeling of proteins leaves major gaps affecting most proteins and especially functional analysis. The opportunities to transcend them beacon: AI models can now predict static protein structures with high accuracy. This achievement is rightly celebrated. It is equally important to recognize what remains unresolved and why those gaps matter hugely for biology. 1. Modeling intrinsically disordered regions (IDRs) is a central limitation. Roughly 30–40% of amino acid residues in the human proteome fall into this category, and ~70% of proteins contain substantial disordered segments. These regions do not adopt a single stable structure; instead, they exist as dynamic ensembles that often become structured only upon binding or under specific cellular conditions. Current AI models -- trained on static structures -- do not predict these ensembles. Instead, they either assign low confidence or produce arbitrary conformations. This is not a minor edge case; it is a large and functionally critical fraction of proteome space, deeply involved in signaling, regulation, and disease. 2. A second key limitation concerns protein function. Biology ultimately depends on changes in conformation, interactions, and state. Many key biological processes arise from shifts between multiple conformations or from subtle perturbations induced by amino acid substitutions, post-translational modifications, or binding partners. Current models are optimized to predict a single, most likely structure. They are not designed to capture how that structure changes under perturbation, nor how populations of states shift. As a result, predicting function -- arguably the central goal -- remains a weakness in many cases. Outlook These two challenges point to a deeper issue: proteins are not static objects but dynamic systems governed by energy landscapes. What is needed next is not just better structure prediction, but models that can capturing ensembles, relative state populations, and the effects of perturbations on those distributions. This will likely require accurate and scalable measurements of proteins, integrating generative models, explicit or learned energetics, and dynamic sampling into a unified framework. In this sense, the field is entering a new phase. Predicting “the structure” was a milestone. Understanding how proteins move, adapt, and function -- especially in the large, disordered fraction of the proteome -- remains the frontier.

Post attachment
Daniel avatar
⟳ Pavle Jeremić retweeted
Daniel@growing_daniel·Mar 8

Make San Francisco a shitty place to be a drug addict

Noah Smith 🐇🇺🇸🇺🇦🇹🇼 avatar
⟳ Pavle Jeremić retweeted
Noah Smith 🐇🇺🇸🇺🇦🇹🇼@Noahpinion·Feb 25

Star Wars depicts a future where cybersecurity just doesn't work. They have AGI but they keep it bottled up in droids; they don't network anything. As soon as R2-D2 gets access to an actual network he successfully hacks the Death Star.