Meta dismisses protein folding team, shifts focus to commercial AI.

2023-08-11

After abandoning its metaverse dreams, Meta has laid off its protein folding team, which had developed the revolutionary ESMFold or Evolutionary Scale Modeling for protein structure prediction two years ago. The 12-member team also created a comprehensive database with over 600 million protein structures.

This move indicates that Meta is shifting its focus from life sciences to revenue generation from commercial artificial intelligence. In the second quarter of 2023, the company achieved its highest level of profitability since 2021. In terms of advertising, the year-on-year growth rate of ad impressions displayed in their application suite reached 34%. No wonder the owner of Facebook gave up on protein folding ambitions.

According to a report by the Financial Times, Yaniv Shmueli, a former research scientist and engineering manager at Meta AI who worked on ESMFold, mentioned that Meta had attempted to adjust its research approach to gain a deeper understanding of the development of advanced intelligence with potential commercial applications, rather than focusing solely on "curiosity projects".

ESMFold vs. AlphaFold

AlphaFold was introduced by Google DeepMind in 2018 and released its second version in 2020, followed by an open-source version of AlphaFold 2 based on deep learning neural networks last year. In addition to AlphaFold and ESMFold, Chinese biotech company Helixon developed OmegaFold, Generate Biomedicines launched Chroma, and Baker Lab introduced RoseTTAFold and RoseTTAFoldDiffusion, among others.

Meta's ESMFold utilizes large-scale language models with a focus on evolutionary scale modeling, while AlphaFold is based on neural network models and achieves high accuracy in predicting both multimeric interfaces and single-chain accuracy. Although ESMFold has an inference speed 60 times faster than AlphaFold and can analyze the protein structure space of macrogenomes, especially sequences from natural environments, its accuracy is not as high as AlphaFold's.

ESMFold uses a single sequence input to generate predictions, leveraging the internal representation of the language model, while AlphaFold and others use multiple sequence alignments and templates. ESMFold performs well in atomic-level predictions, particularly for low-complexity sequences, making its prediction speed a major advantage.

Google Takes the Lead

In addition to being the most cited paper in 2022, AlphaFold 2 won CASP14 in 2020, considered the best protein folding model, and predicted the structure of a protein database 200 times larger in collaboration with EMBL-EBI. Meanwhile, AlphaFold is making an impact in various practical applications, including predicting protein structures for outbreaks like COVID-19 (SARS-CoV-2), advancing drug development for diseases like malaria, conducting gene therapy, combating antibiotic resistance, and addressing climate pollution, among others. However, there is almost no publicly available information about the use cases of ESMFold.

When it comes to using generative artificial intelligence in the healthcare field, Google can be said to be leading the way. Their Med-PaLM 2, built on their proprietary language model PaLM 2, is a medical chatbot that answers medical questions and has been popular since its launch. It is being tested in medical institutions such as the Mayo Clinic. Google Health, Google DeepMind, and Google AI have introduced Med-PaLM M, a large-scale multimodal generative model that flexibly encodes and interprets biomedical data. It can handle various types of medical data, including clinical language, medical images, and genomics, and performs well on various tasks using the same model weights.

Meta Continues with Layoffs

In a major restructuring, Meta has carried out its final round of layoffs, which is part of the 10,000 job cuts announced in May. The layoff of ESMFold is also part of this. Previously, Meta had laid off over 11,000 employees, bringing its workforce back to mid-2021 levels after experiencing rapid expansion since 2020. The layoffs reduced the number of employees from nearly 90,000 (as of mid-November 2022) to approximately 65,000.

Over the years, Meta's management has been promoted based on the scale of the teams they built. However, in Mark Zuckerberg's "year of efficiency," promotions will become less frequent. This "efficiency" has led to cost-cutting, organizational restructuring, and large-scale layoffs.

Meanwhile, Meta is certainly fully committed to supporting generative artificial intelligence.

As reported earlier by AIM, their new decentralized social media platform Threads has been successful, but Zuckerberg may be using this data to train a more powerful LLM than Llama 2. Meta also plans to launch a series of AI chatbots with unique personalities to enhance user interaction on platforms like Instagram and Facebook and facilitate human-like conversations.

This is not the first time Meta has disbanded a project by laying off a team. Earlier this year, the tech giant announced the layoff of 10,000 positions after suffering a $13.7 billion loss in its metaverse division, turning their metaverse dreams into a nightmare. It is currently unclear why Zuckerberg decided to abandon the ESMFold team, especially considering the enormous potential of protein folding models in improving human life.

RECENT AI NEWS

RECENT AI TOOLS