Generative A.I. Designs Novel Synthetic Proteins

(Gerd Altmann, Pixabay. https://pixabay.com/illustrations/digitization-particles-smartphone-7261158/)

16 Nov. 2023. A biotechnology company demonstrated its technology for designing entirely new types of therapeutic proteins based on a generative artificial intelligence model. Generate:Biomedicines in Somerville, Massachusetts describes its Chroma generative A.I. model in yesterday’s issue of the journal Nature, and is making the model available as open source through GitHub.

Generate:Bio is a five year-old company that designs and develops therapeutic proteins from scratch to meet specific chemical and medicinal properties. The company says its Chroma platform is based on an A.I. neural network or machine learning model trained on amino acid sequences, three-dimensional structures, and statistical patterns of proteins in nature. The model next applies diffusion principles for generating new protein sequences and structures, starting with simple data that then adds more complexity with each following step to achieve the desired functions. In addition, says Generate:Bio, the Chroma technology is programmable, in that the model can be steered to design novel protein structures and sequences that meet chemical and therapeutic objectives specified by the user.

The company says its process can design protein molecules from short peptides to more intricate enzymes, antibodies, and gene therapies. Generate:Bio is developing several of its own protein therapeutics for immunology, cancer, and infectious diseases, with a drug addressing SARS-CoV-2, the virus responsible for Covid-19 infections, in an early-stage clinical trial. The company is also collaborating with M.D. Anderson and Roswell Park cancer centers on therapeutic protein discovery and drug maker Amgen on a number of undisclosed disease targets.

Programming protein output with natural language prompts

In the Nature paper, a team led Generate:Bio co-founder and chief technology officer Gevorg Grigoryan analyzed 120,000 novel protein designs generated by the Chroma model, both single-chain and more complex proteins, but not programmed for specific therapeutic functions. The analysis shows the novel protein designs match many of the properties of natural proteins in the Protein Data Bank, a training source of the company’s algorithms. The researchers also demonstrated programmable structure, folding, sizing, and shaping features of Chroma, even creating proteins shaped like Roman alphabet and Arabic number characters. In addition, the team displayed the ability of programming Chroma output with natural language prompts.

The researchers validated the Chroma model by generating 310 novel designs and testing the proteins they produced on features such as solubility, melting temperature, and expression in E. coli bacteria, a commonly used model species in biology. And the team also produced proteins meeting specified chain lengths, from 100 to 950 amino acids. The authors say their tests show Chroma can produce very large and complex protein designs with standard A.I. graphic processing chips in a few minutes.

Grigoryan — also a professor of computer science and biology at Dartmouth College — says in a Generate:Biomedicines statement, “we have demonstrated that fundamentally new molecular concepts can now be generated at will and used to create novel proteins with a diverse set of desired properties.” Plus, the company is making the Chroma model code available as open source on the GitHub web site. “By making the code open source,” adds Grigoryan, “we’re proud to share this breakthrough in generative biology with the world to help advance our shared goals of fundamentally changing the way protein therapies are created.”

More from Science & Enterprise:

• Challenge Seeks Transformative Human/A.I. Research Tech
• Genomics Company, Amazon Partner on Generative A.I.
• Fujitsu, Linux Foundation Partner on Open-Source A.I.
• Biotech, A.I. Companies Model Animal Studies/Clinical Trials
• Darpa Seeking On-Demand, Distributed Production of Proteins

We designed Science & Enterprise for busy readers including investors, researchers, entrepreneurs, and students. Except for a narrow cookies and privacy strip for first-time visitors, we have no pop-ups blocking the entire page, nor distracting animated GIF graphics. If you want to subscribe for daily email alerts, you can do that here, or find the link in the upper left-hand corner of the desktop page. The site is free, with no paywall. But, of course, donations are gratefully accepted.

*     *     *