All news

Open protein models move value from prediction to lab data

A sculpture of a white paper ribbon folded into twisting loops on a graphite table, one short segment in cobalt

DeepMind published AlphaFold's method and code, and a university team released a rival the same day. Structure prediction is becoming a free input.

About one GPU minute per model now gives the structure of a protein of 384 amino acids, work that took months or years of lab time per protein. On 15 July DeepMind, the Alphabet research lab, published the full method behind AlphaFold, its protein structure prediction system, in a peer-reviewed paper and released the source code under an open-source license, with the trained parameters free for non-commercial use.

Publication turns the prediction itself into a free input. Once the leading lab releases its method and code, the model stops being a moat, the same pressure open models already put on paid access to language models. Value moves to what public data cannot supply: proprietary measurements of how proteins bind and behave, and the speed of the loop between design and experiment. In science, AI-first companies will hold their edge in data and workflow, not in the published model.

The quality leaves little room for doubt. In last year's blind assessment AlphaFold's predictions had a median backbone error of 0.96 ångström against 2.8 for the next best method, competitive with laboratory experiments in most cases. The same day a University of Washington team published RoseTTAFold, a similar deep learning system with public code that computes a structure in as little as ten minutes on a single gaming computer.

The gap between known sequences and known shapes shows how much ground has opened. Experiments have determined about 100,000 unique protein structures, a sliver of the billions of known sequences. Training is within reach as well: the paper describes about eleven days on 128 TPU cores, a budget a funded company can spend to train its own weights for commercial use.

Structure prediction is about to become a standard input for every biology team, and that sets the bar for founders. Companies that pair these models with their own lab, generate data no one else has and design molecules instead of only predicting them can build defensible businesses. An open tooling layer sits beside them, from running the models at scale to turning predictions into decisions for drug and enzyme programs. The license matters on day one: the released AlphaFold weights bar commercial use, so plan to retrain or to build on code with open terms.

One lever still sits with DeepMind. The parameters stay closed to commercial use, and while they do, the owner of the best model keeps a toll on its predictions. Opening the license would remove it, and so would the step its companion work on the human proteome describes, predicted structures for whole proteomes, handing every team the output without running a model. Adoption settles the rest: how fast drug developers build these predictions into discovery programs, and how many new machine learning biology companies raise capital on the back of them.

Open research compresses the value of any single model and rewards whoever owns the data and the customer. Fanaura's venture thesis weighs proprietary data above model architecture for exactly that reason, and the same logic argues for a long horizon on science companies whose results compound with every experiment.

For founders

Building a machine learning biology company with its own lab data? Send us the project.

Fanaura invests its own capital and the capital of its limited partners (LPs), usually through a separate SPV for each deal. If we see a fit, we get in touch.

Apply in five minutes
Subscribe

Get the news by email

One short note each weekday. No marketing, and you can unsubscribe at any time.