How generative AI and E. coli are speeding up new drug discovery

For many, hearing the word E. coli is often a reason to be concerned, as the bacteria can lead to incidents of food poisoning in humans.

As it turns out, E. coli might well be the panacea that enables a new form of generative AI for healthcare that could help enable researchers to generate new antibodies. Generative AI in recent years has captured popular imagination by enabling users to generate text or images on demand, but its uses go much deeper, too. Generative models that provide large machine learning (ML) models that can create new things is an emerging area in science helping to accelerate discovery.

Sean McClain, founder and CEO of Absci, came up with the idea of engineering E. coli to produce antibodies that have the potential to improve human health. Absci has been able to build out a generative AI model using data collected from testing with E. coli. Today, the company announced that the generative AI model has been able to create an entirely new (de novo in scientific terms) antibody in software, thanks in no small part to the often-maligned E. coli bacteria.

“A lot of people think of [E. coli] as a complete negative, but it’s an organism that has turned out to be the hero for healthcare,” McClain told VentureBeat.

The science of using E. coli to build antibodies

Generative AI is typically built using some form of large language model (LLM) that has been trained on a large number of parameters.

The initial set of data to train a generative AI model is critically important. A challenge with developing antibodies is that they often need to be made in a living organism, such as mammalian cells, which McClain said can have scalability limitations on the number of antibodies that can be produced. With E. coli, he said that an order of magnitude more antibodies can be produced, which enabled the development of a large dataset to train the generative AI model.

The promise of the generative AI model for antibody development is that it can dramatically accelerate the path to new discoveries. McClain said that on average it takes five and a half years for researchers to be able to get a new antibody into clinical testing. Once those drugs make it into testing, only approximately 4% are likely to be successful.

McClain said that his company’s new breakthrough is that it is now able to use its generative AI model to build an antibody that can bind to a specific target. The process creates an entirely new antibody and is done all on the AI system.

“This is a huge paradigm shift within the industry and, ultimately, it’s going to drive to getting drugs into the clinic in 18 to 24 months, instead of five and a half years, and it’s going to increase that 4% success rate,” McClain said.

Taking the prompt approach to generative AI

Among those helping to lead the generative AI efforts at Absci is chief AI officer Joshua Meier, whose career has included stints working as a research engineer at Facebook, where he was part of a group working on generative biology.

The Absci model was trained using a combination of supervised and unsupervised learning. Meier said that data is fed into the model and it learns how different proteins interact with each other.

With a typical generative AI model, a user will provide a prompt — that is, a description of a desired output to get a result. With the Absci model, a scientist will prompt the model with a protein to target in order to generate an antibody. The prompts can become very specific in order to generate very unique and specific antibodies.

In terms of the hardware that enables the Absci system, the company has built its own in-house supercomputer for its generative AI model, benefiting from a partnership with Nvidia.

“We formed a partnership with Nvidia and we’ve been working with them with a focus on model scaling,” Meier said.

Scale is one particular area that Absci has been able to excel. McClain said that his company is currently able to validate approximately 2.8 million AI-generated antibody drug candidates per week. Overall, McClain is hopeful that the generative AI approach will lead to a new era for medicine.

“This type of technology is going to enable personalized medicine,” McClain said. “Being able to take a patient sample, find a target that’s relevant for a disease, and then instantaneously be able to design a drug or an antibody that’s going to cure that particular disease — and all at a click of a button.”