After abandoning its metaverse dream, Meta has laid off its Protein folding team that built the revolutionary ESMFold or Evolutionary Scale Modelling for protein structure prediction, exactly two years ago. The 12-member team also created a comprehensive database of over 600 million protein structures. 

The move indicates Meta shifting its focus from life sciences to generating revenue from commercial AI. During the second quarter of 2023, the company achieved its highest level of profitability since 2021. On the advertising front, there was a notable 34% annual growth in the number of ad impressions displayed across their suite of applications. No wonder the Facebook owner is giving up on protein-folding ambitions. 

As per a report by 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 in order to gain deeper insights into the development of advanced intelligence with potential business applications, as opposed to focusing solely on “curiosity projects”.

In a major restructuring bid, Meta held its final round of layoffs — part of the plan to cut 10,000 jobs announced in May. The letting go of ESMFold is also a part of this. Previously, Meta laid off over 11,000 employees, bringing its workforce back to its mid-2021 level after a period of rapid expansion since 2020. The layoffs that have reduced the workforce from nearly 90,000 (mid-nov 2022) to about 65,000 employees. 

For years, Meta’s managers have been promoted based on the size of the teams they build. However, in Mark Zuckerberg’s “year of efficiency”, promotions will now become less frequent. The “efficiency” is resulting in cost-cutting, organisational restructuring, and mass layoffs. 

Meanwhile, Meta is going all-in for generative AI for sure.

Although its new decentralised social media platform Threads became an instant success, as AIM reported earlier, Zuckerberg may be possibly using the data to train an LLM more powerful than Llama 2. Meta is also planning to unveil a range of AI-powered chatbots featuring unique personalities for platforms like Instagram and Facebook to enhance user interaction and facilitate human-like conversations. 

Read more: Protein Wars: It’s ESMFold vs AlphaFold

ESMFold Vs AlphaFold

AlphaFold, launched in 2018 by Google DeepMind, published its second version in 2020 and released an open-source version of its deep-learning neural network Alphafold 2 last year. Besides AlphaFold and ESMFold, Chinese biotech firm Helixon developed OmegaFold, Generate Biomedicines brought Chroma, Baker Lab brought in RoseTTAFold and RoseTTAFoldDiffusion, and the list goes on. 

Meta’s ESMFold leverages a large-scale language model, focusing on evolutionary scale modelling while AlphaFold is built on the neural network-based model, achieving high accuracy for predicted multimeric interfaces and intra-chain accuracy. Although ESMFold has a 60 times faster inference, enabling analysis of metagenomic protein structural spaces, particularly for sequences from natural environments,  it is proven to be less accurate than AlphaFold. 

ESMFold generates predictions using a single sequence input, leveraging the language model’s internal representations, while AlphaFold and others employ multiple sequence alignments and templates. It is good at working with atomic-level predictions, especially for low-perplexity sequences, with prediction speed being a major advantage. 

Google Taking the Lead

Besides being the most cited paper of 2022, AlphaFold 2 won the CASP14 in 2020 and is regarded as the best protein-folding model. The collaboration with EMBL-EBI predicted the structure of a 200 times bigger protein database. Meanwhile, AlphaFold is finding several real-life applications including the prediction of protein structures of the COVID-19 outbreak—SARS-CoV-2 — advancing the development of drugs for malaria, neglected diseases like Leishmaniasis, delivering gene therapy, treating antibiotic resistance, fighting climate pollution and much more. However, there is barely any public information on ESMFold’s use cases. 

Besides AlphaFold, when it comes to using generative AI in healthcare, it is safe to say Google is leading. 

Their Med-PaLM 2, built upon a proprietary language model, PaLM 2, is a medical chatbot that answers medical questions and has been a fan favourite since its launch. It is being tested in healthcare institutions like the Mayo Clinic research hospital. Google Health, Google DeepMind and Google AI have unveiled Med-PaLM M, a large 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 a wide range of tasks, all using the same set of model weights. 

Nevertheless, this is not the first time that Meta disbanded a project by laying off the division. Earlier this year, the big tech announced 10,000 job cuts in its Metaverse division after suffering a $13.7 billion loss in 2022, turning their metaverse dreams into a nightmare. It is still blurry why Zuck decided to give up on the ESMFold team, especially considering the immense potential of protein folding models in touching human lives. 

Read more: From Humble Beginnings to Scientific Stardom: Meet the Protein Prodigy from Bengal

The post Meta’s Double Helix & Farewell to Protein Folding appeared first on Analytics India Magazine.