ESMFold2: Zuckerberg open source model surpasses AlphaFold and predicts 1.1 billion proteins

Biohub, a research institute founded by Mark Zuckerberg, has just published in Nature a model of artificial intelligence that promises to redefine computer biology. ESMFold2 is capable of predicting 1.1 billion protein structures, surpassing Google's AlphaFold in scale and accessibility. And unlike the competitor, the model is completely open source.

How does ESMFold2 work and predict proteins?

ESMFold2 is a language model (LLM) applied to biology. He was trained with millions of sequences of proteins and learned how to predict their three -dimensional structure — a task that previously required expensive and time - consuming laboratory experiments. The differential is on the scale: while Google's AlphaFold covered about 200 million proteins, ESMFold2 reaches 1.1 billion, almost five times more. The search, published in Nature, details the use of an optimized transformer architecture that processes long sequences without losing accuracy.

Why does ESMFold2 beat AlphaFold?

The main advantage of ESMFold2 is not only in the amount of planned structures, but in computational efficiency. The model uses less hardware resources to make predictions, which makes it accessible to smaller laboratories and open source projects. In addition, it incorporates self-supervised learning techniques that allow better generalization for rare or artificial proteins. The result is a tool that can be adapted by any researcher, without relying on centralized servers.

The impact of ESMFold2 on computational biology

With 1.1 billion available structures, scientists can study from protein evolution to the development of new drugs. The discovery also accelerates areas such as synthetic biology and enzyme engineering for industry. The fact that the model is open means that the community can retrain it with specific data, creating specialized versions for tropical diseases or plant proteins, for example.

How does the open source of ESMFold2 accelerate innovation?

The open source release of ESMFold2 follows a trend that IBM and Red Hat have already signaled: high impact AI models need to be collaborative. The ESMFold2 code is available at GitHub, and researchers have already started to create variations for areas such as antibiotic resistance and therapeutic protein design.

Frequently Asked Questions about ESMFold2

How does ESMFold2 compare to AlphaFold in terms of performance?

ESMFold2 covers 1.1 billion structures, surpassing the AlphaFold 200 million, and requires less computational power, making it more accessible to the scientific community.

Where can I access the EMFold2 code?

The code and trained models are publicly available in the official Biohub repository in GitHub, without restrictions of use.

Can this advance help in the development of new medicines?

Yes, predicting the protein structure is a fundamental step towards understanding diseases and designing molecules that interact with them, accelerating drug discovery.

---

==References== original publication #36Kr

See also:* More news about AI