Will AI Graduate from Tool to Lab Member? A Q&A with John Tsang

Tsang: When using AI in a classical way, you have one agent, and you have an interactive discussion with it. With a multi-agent approach, as a few research groups have now described, you ask different independent bots to take different roles. One bot could be a computational biologist, another one could be a biochemist, and yet another one could be someone who knows a lot about antibodies, for example. And then there's also a human who's involved in the team that can give the bots a high-level task and then let them talk to each other.

What is very interesting is that in examples of this type of approach so far, the majority of the interactions happen among the AI bots. The human doesn't necessarily go in and tell them what do. And researchers have found that the bots are capable of coming up with a work plan and executing it. Even when the human, while part of the team, is not doing the major coordination.

So that is quite interesting, suggesting that by having interactions among these bots—and I think the key is that they take on different roles—they are able to start to generate something that's quite emergent, not just straight answering. They can engage a multi-step plan and make decisions based on the data. That is, I think, quite intriguing.

One striking example is “The Virtual Lab,” a system developed by Biohub colleagues in which an AI principal investigator led a team of AI specialists—an immunologist, a computational biologist, and a machine learning expert—to design nanobodies (small antibody-like molecules) against a new SARS-CoV-2 variant. The team produced 92 candidate designs, and two were experimentally confirmed to bind the new variant. Notably, more than 98% of the words exchanged during the project came from the AI agents themselves; the human researcher set the goal and stayed in the loop, but the bulk of the back-and-forth largely happened among AI agents.

Tsang: I think the answer lies on two levels. First, I think it's going to be helpful in proposing and hypothesizing things that we haven't thought of yet. However, that will be based on existing data, and we still have limited data on lots of things in immunology.

For example, one project that I co-lead called the Human Immunome Project aims to capture human immunology variation across different people and across different populations. Now that kind of data simply doesn't exist yet. If you ask a chat bot or even multi-agents and say, "I know this vaccine works in the U.S. with this level of efficacy, what happens if I bring it to Tanzania?" it would have no idea because it simply doesn't have enough information and knowledge yet.

So that's where a different kind of AI, models like quantitative models of the human immune system, come in. That's another area that we are actively working on, how to develop that kind of model. Of course, once such models become available and useful, then AI agents can use them to query and answer the questions we ask.

Additionally, AI agents, I think, could form a team with us to work together on that problem, where we need to design experiments. We need to think about, given a certain number of resources and budget, what's the best first experiment, for example, what's the best first data collection exercise? Those are the kind of things that we are thinking about a lot.