BRENNEN

Industries / Health / Pharmaceuticals / Biotech

Protein design

This work makes proteins to order: growing natural ones in cells, evolving them in the lab, and now drawing new ones on a computer before anyone makes them.

7

Milestones achieved, from 1975 to 2023.

3

To watch for: announced, not yet confirmed outside.

3

Open deliverables. Nobody has shipped them.

The road

AchievedWatch forOpen

Achieved

  1. 1975

    Antibody factories that never die

    Georges Köhler and César Milstein fused a mouse antibody-making spleen cell to a tumour cell. The result, a hybridoma, grows forever and makes one single antibody against a chosen target. The method became the basis of most antibody drugs and many diagnostic tests, and the two shared the 1984 Nobel Prize in Physiology or Medicine with Niels Jerne.

    bioc.cam.ac.uk

    Credit

    Georges Köhler and César Milstein, MRC Laboratory of Molecular Biology, Cambridge

  2. 1978

    Human insulin grown in bacteria

    A Genentech team, with Arthur Riggs and Keiichi Itakura at City of Hope, made the two chains of human insulin in E. coli from genes built from chemicals, first reported in January 1979. Eli Lilly scaled up production, and the FDA approved Humulin in October 1982. A review calls it the first protein medicine made by recombinant DNA to be approved for use in people.

    Lessonthe hard part was not the gene but the yield and the manufacturing.

    pmc.ncbi.nlm.nih.govnsf.gov

    Credit

    David Goeddel, Dennis Kleid and Herbert Boyer (Genentech); Riggs and Itakura (City of Hope); Eli Lilly

  3. 1985

    Evolution moved into the lab

    George Smith showed that a bacteriophage, a virus that infects bacteria, can carry a protein on its coat, so that the ones that stick to a target can be fished out. Frances Arnold ran the first directed evolution of enzymes in 1993, making random changes and keeping the best performers. Gregory Winter applied the phage method to antibodies, and the first drug from it, adalimumab (Humira), was approved in 2002. The three shared the 2018 Nobel Prize in Chemistry.

    statnews.com

    Credit

    George Smith (University of Missouri), Frances Arnold (Caltech), Gregory Winter (MRC LMB)

  4. 2003

    A shape nature never made

    The Rosetta software designed Top7, a protein of 93 amino acids with a fold not found in the Protein Data Bank. The crystal structure, solved at the Advanced Light Source in Berkeley, matched the computer model to 1.17 ångström.

    Lessonshape could be designed long before function could.

    als.lbl.govsciencedaily.com

    Credit

    Brian Kuhlman, Gautam Dantas and David Baker, with the University of Washington, UNC Chapel Hill and Fred Hutchinson Cancer Research Center

  5. 2020

    A Nobel laureate retracts a paper

    Frances Arnold withdrew a 2019 Science paper from her Caltech group on making lactams with enzymes after the results could not be repeated, saying publicly "I was a bit busy when this was submitted, and did not do my job well."

    Lessona published result in protein engineering means little until another bench repeats it.

    chemistryworld.com

    Credit

    Frances Arnold, Caltech

  6. 2022

    A designed protein as medicine

    South Korea's Ministry of Food and Drug Safety authorised SKYCovione, a COVID-19 vaccine built on a two-part particle that assembles itself and carries 60 copies of part of the virus spike. Its phase 3 trial enrolled 4,037 adults, and UK authorisation and a World Health Organization emergency listing followed. The University of Washington calls it the world's first computer-designed protein medicine.

    ipd.uw.edu

    Credit

    the Neil King and David Veesler labs at the Institute for Protein Design, with SK Bioscience and GSK

  7. 2023

    Design by diffusion

    RFdiffusion starts from random noise and cleans it up into a protein backbone, the way image generators work. It produced binders to five targets, with 19% of designs working in the lab, and an electron-microscope structure of one influenza binder matched the design to 0.63 ångström. The amino-acid sequences for those backbones came from ProteinMPNN, published in Science in September 2022, which runs in about a second and was reported by its authors as more than 200 times faster than the previous best software.

    pmc.ncbi.nlm.nih.govsciencedaily.com

    Credit

    Joseph Watson, Nathaniel Bennett, Brian Trippe, David Juergens and David Baker (University of Washington); ProteinMPNN by Justas Dauparas and colleagues

Watch for

  1. Whole antibodies drawn from scratch

    The Baker lab reports in Nature that RFantibody, an antibody-specific version of RFdiffusion paired with yeast screening, designs complete antibody binding regions against a chosen patch on a target. Antibodies against a Clostridioides difficile toxin bound at mid-nanomolar strength, and electron microscopy of five designs matched the computed model for four. No overall success rate is given.

    Achieved whenan unaffiliated group runs the public RFantibody code and reports comparable success rates and structures on its own targets.

    ipd.uw.edu

    Claimed by

    Institute for Protein Design, University of Washington (Nathaniel Bennett, Joseph Watson, David Baker), November 2025

  2. An AI agent runs a design campaign

    Anthropic reports that its models Claude Opus 4.8 and Mythos Preview, given one written protocol and a cloud GPU budget but no human design decisions, researched 16 targets, chose binding sites, ran open-source design tools and returned 30 ranked candidates for each. Two contract labs, Adaptyv Bio and Twist Bioscience, made every design: 354 of 1,320 bound (27%), rising to 49% among each campaign's top-ranked design, and the two labs agreed on 1,099 of 1,235 shared designs (89%).

    Achieved whenan outside laboratory reruns the released prompts on new targets and reports similar success rates.

    assets.claude.com

    Claimed by

    Anthropic (Amir Shanehsazzadeh and colleagues), August 2026

  3. AI-designed antibodies reaching patients

    Several firms report antibodies from machine design at or near first trials. Absci's ABS-101 is in a clinical trial, which the company says it reached in about two years against an industry average of five and a half; Generate:Biomedicines has three phase 1 trials of a design it optimised rather than drew from scratch; LabGenius expects a regulatory filing in 2026; Nabla Bio says 1 to 10 of every 100 antibodies its model designs hit the target.

    Achieved whena regulator approves one, or independent trial results are published. The article states that no AI-designed antibody has regulatory approval.

    pharmaceutical-journal.com

    Claimed by

    Absci, Generate:Biomedicines, LabGenius and Nabla Bio, reported August 2025

Open

  1. 01

    One rule for counting a hit

    A public script and table that recomputes the success rate of every open binder-design competition under a single written rule, saying what counts as made, what counts as a binder and what counts as designed from scratch, so that numbers from different events can be compared.

    The brief
    Credit · unclaimed
  2. 02

    Guessing which targets will fail

    A model that predicts, before any protein is made, how hard a given target will be. It does not predict which design will bind, but whether a campaign against that target will produce anything at all.

    The brief
    Credit · unclaimed
  3. 03

    Designs that were never made

    A public test set for the earlier failure that design papers rarely discuss: a protein that cannot be produced at all. Collect every design with a recorded production outcome, then train and score predictors of that failure.

    The brief
    Credit · unclaimed

Open 01 · the brief

One rule for counting a hit

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public script and table that recomputes the success rate of every open binder-design competition under a single written rule, saying what counts as made, what counts as a binder and what counts as designed from scratch, so that numbers from different events can be compared.

Why it is still open

The Adaptyv community preprint names "the lack of standardized benchmarks, experimental design targets, and robust computational metrics for method comparison" as the problem, and the Anthropic report says competitions cannot be compared directly because they differ in number of entries, design budget per team and permitted formats. ProteinBase publishes the raw results but applies each organiser's own tally.

biorxiv.orgassets.claude.com

Built with what exists

The public ProteinBase collections; the Adaptyv EGFR competition data on GitHub; the Anthropic dataset on Hugging Face (CC BY 4.0 data, MIT-licensed scripts); Python and pandas.

  1. First step

    Pull the two EGFR competition releases and the Anthropic release and reproduce their published counts exactly, before changing any definition.

  2. Done looks like

    A versioned repository covering at least six competitions that reproduces each organiser's published number and gives the recomputed one beside it, with every difference explained in a line of text.

Open 02 · the brief

Guessing which targets will fail

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A model that predicts, before any protein is made, how hard a given target will be. It does not predict which design will bind, but whether a campaign against that target will produce anything at all.

Why it is still open

The Anthropic report states that confidence scores separated binders from non-binders within a target on 12 of 13 targets, but that across targets the scores "rose with hit rate, but not steeply enough to have flagged the failures in advance". Published analyses so far score individual designs, not targets.

assets.claude.combiorxiv.org

Built with what exists

The Anthropic per-design dataset (15 targets with binder and non-binder results); the Overath analysis of 3,766 measured binders across 15 targets; ProteinBase collections; open structure predictors; scikit-learn.

  1. First step

    Assemble one table of target, number tested and number bound from the three public sources, and check how well median confidence alone orders those targets.

  2. Done looks like

    A published model that ranks held-out targets by measured success rate better than the median confidence score does, reported as a rank correlation with leave-one-target-out testing, with the code and the held-out split fixed in advance.

Open 03 · the brief

Designs that were never made

Credit · unclaimed

The first to publish is credited on this line, and the deliverable moves from Open to Achieved.

What to build

A public test set for the earlier failure that design papers rarely discuss: a protein that cannot be produced at all. Collect every design with a recorded production outcome, then train and score predictors of that failure.

Why it is still open

The published work scores binding instead. Adaptyv's second EGFR round reports that 378 of 400 designs were produced and the rest were not, and Anthropic's release covers 1,440 designs of which 120 gave no usable measurement. Neither the Adaptyv community preprint nor the Overath analysis sets production as the task; the latter is built from designs that were measured.

adaptyvbio.substack.combiorxiv.org

Built with what exists

The Adaptyv GitHub releases for both EGFR rounds; ProteinBase collections; the Anthropic Hugging Face release; open sequence models and standard classifiers.

  1. First step

    Write the loader that pulls the production column from the two EGFR releases and reproduces the published 378-of-400 figure.

  2. Done looks like

    A dataset card listing at least 1,000 designs labelled produced or not, a train and test split fixed by competition round, and a reported area-under-curve for at least two baseline predictors on the held-out round.