Alex Rives
Related people
Episodes
Notes
No Priors: Artificial Intelligence | Technology | Startups · Biohub: The Future of Biology is Open-Source with Co-Founders Mark…
-
Why Does Biohub Study Systems Instead of One Disease?
Mark Zuckerberg says Biohub studies systems such as inflammation and the immune system, rather than one disease, so other groups can use the tools on specific therapies.
Inflammation is a focus of the Chicago Biohub because it connects to many illnesses. The immune system is a useful case for the cellular engineering work. The stack runs from proteins to cells to a whole system, and immune cells can travel through the body. How to make the immune system function better, and how to connect that last step to a treatment, is work he says biotech companies and individual academics are better suited to do. Biohub's part is the shared tool set that speeds those groups up.
-
Why Does Biohub Release Biology Models as Open Source?
Mark Zuckerberg says Biohub releases its biology models as open source so more scientists get the tools sooner, which he expects will move the field faster than keeping the work inside a company.
The effort could have been a business. Biology models stay smaller than language models because there is less data, and that data is not something a lab can buy off a shelf. New methods have to be invented first, including cellular engineering in New York and measurement devices in Chicago. Zuckerberg puts the horizon for tools this complicated at 10 to 15 years. Skipping the question of how each model will make money keeps the strategy simple. The aim is still not for Biohub to cure the diseases.
-
How Would Biohub's Virtual Cell Link Genes to a Phenotype?
Alex Rives says Biohub's virtual cell would model each level of biological complexity and connect the proteomic layer, the genetic layer, and the transcriptomic layer to the phenotype. Views on the exact input and output differ, and the aim is enough generality to ask about a new intervention in a context the model has not been trained on and get an answer.
The gap the field still has to close is making the predictions generalize. Closing the gap will require an enormous effort to generate data. What to do next is ordinary constraint management. Every lab in every field across the world probably feels compute constrained, and the same limit probably holds for the work.
Ongoing debates cove
-
Can a Model Predict Where a Drug Will Bind?
Alex Rives says a model that can simulate where a drug distributes and binds across the body is the beginning of a solution to toxicity.
The step he is describing is the move from ESM Fold 2 to a model that can design a molecule and send it toward the clinic with a much better hit rate. Figuring out how to optimize a drug, how to get it through preclinical work, and how to do the early safety check are problems people could spend an entire career on. A new scientific paradigm simplifies questions that were once hard, and he expects many of those problems to be solved in an emergent way. Accurate representations at the molecular level are what he says will start that rapid progress.
-
Why Was Baby KJ's Therapy Aimed at the Liver?
Priscilla Chan says baby KJ's disease was chosen because the therapy had to reach his liver, where a product could be delivered.
The team at CHOP delivered a CRISPR therapy to edit a mutation that would have led to significant neurotoxicity and changed his life. Chan treats that delivery choice as the judgment that unlocks a first application. She also says a human test can still show renal toxicity when a kidney cell carries a receptor the drug was not supposed to hit. A single-cell atlas, including cell types that had not been predicted before the modeling, can show which cells carry that receptor before a human trial. A transcriptomic model is one way to see how different cells react to
-
How Does Biohub's Protein Model Design Antibodies?
Alex Rives says Biohub's new ESM fold model designs single-chain antibodies on a computer, and a lab test of 96 proteins found nanomolar binders, the level needed for therapeutic activity, even though the team did not build the model for antibodies.
About a week before the conversation, the team announced ESM fold as an open system for scientific discovery in protein biology. It is a world model with a language-model base, trained on billions of protein sequences, and it learns emergent representations of protein biology. It predicts atomic-resolution protein structure fast enough to illustrate a Pareto optimal frontier of speed and accuracy. The team folded over 1.1 billion proteins, predi
-
How Does Biohub Model Biology From Proteins Up?
Mark Zuckerberg says each layer of biology ends up qualitatively different, so Biohub builds models from proteins upward instead of jumping straight to cells.
Protein interactions have to be understood before cells can be understood. How cells work has to be understood before a larger system, such as the immune system, can be understood. A very high level of abstraction might simulate a system, but understanding how it will work means a simulation at each level. Biohub starts with the protein building blocks. The data to collect differs by layer, and the modeling techniques keep advancing. The strategy is still to build the stack hierarchically. AI work and wet-lab work are a single effort,
-
Why Does Biohub Build Tools Instead of Cures?
Priscilla Chan says Biohub's goal is to build tools that accelerate whole scientific fields so the field can collectively cure disease, and Biohub is not the group that will cure the diseases.
The work started 10 years ago as a way to give back. Mark Zuckerberg wanted an organization that could cure, prevent, and manage all disease by the end of the century. Famous Nobel Prize winning scientists laughed and treated the end of the century as a stretch. She now treats that timeline as too conservative. The scientists said labs work in silos, published information stays locked up for long periods, and labs lack tooling. A great tool built by one postdoc lives on one computer and is gone when t