07.04.2026

James J. Collins is one of the leading scientists at the interface between biology, engineering and artificial intelligence at MIT. In an interview with Bayern Innovativ, he talks about his current research and provides insights into how generative AI could contribute to the development of new antibiotics in the fight against antibiotic resistance.

Professor Collins, can you briefly introduce yourself and outline your current research focus for those who don't know you yet?
James Collins: Yes, with pleasure. My name is James Collins. I am Termeer Professor of Medical Engineering and Science and Professor of Bioengineering at MIT. I am also a founding member of the Wyss Core Faculty at Harvard University and a member of the Broad Institute, a biomedical research center focused on genomics, run in partnership with MIT, Harvard University, and several Harvard-affiliated teaching hospitals.
Our laboratory focuses on two areas. One is synthetic biology, which is the application of engineering principles to develop synthetic gene circuits and other molecular components. These are used to specifically control and reprogram living cells and cell-free systems in order to create new functions for various applications. Secondly, we focus on the use of AI to discover and develop antibiotics to counter the growing global threat of antimicrobial resistance.

Why has antibiotic resistance become such a major health problem worldwide?
James Collins: Antibiotic resistance is a serious and unfortunately growing problem, partly due to two factors. One is the fact that we have overused and misused antibiotics, because by using antibiotics in situations where they are not needed, we are encouraging bacterial pathogens that may be present to adapt and become more resistant. They develop acquired mutations that make them insensitive to the effects of antibiotics. Such mutations can also intensify and lead to multi-drug resistance (MDR) and extensive drug resistance (XDR). Added to this is the use of antibiotics by the agricultural industry as a growth additive and preventative measure.
Unfortunately, these challenges are accompanied by a comparatively weak development pipeline for antibiotics, with very few new active substances being discovered and developed. As a result, we cannot sufficiently replenish our antibacterial arsenal, which is weakened by increasing resistance, with new, effective antibiotics.

Your team uses generative AI in the active ingredient design of potential new antibiotics. What exactly does the AI generate, i.e. what kind of "output" does it deliver?
James Collins : The fact that we are working with generative AI is still relatively new. A few months ago, we published a paper in the journal Cell, which was led by Aarti Krishnan, a postdoc in our lab. In this paper, we used a few generative approaches. One approach involved a chemically plausible mutation procedure in which a starting molecule was specifically modified to generate variant structures that exhibit higher antibacterial efficacy. Another approach used Variational Autoencoders (VAE) to specifically modify compounds or fragments and extend them with additional chemical substructures, which can lead to potentially more effective antibiotics.
Our work has shown that our approaches have identified seven new antibacterial compounds, two of which have been further developed in mechanistic and animal studies.

Can you tell us about your most recent study, in which you used generative AI to develop new antibiotic candidates, and explain the process in simple steps?
James Collins: Yes, of course. We usually start with a training phase where we apply a compound library - in our case about 39,000 compounds - to selected pathogens and measure their antibacterial activity based on growth inhibition.
We then assign the compounds the characteristic "antibacterial" or "non-antibacterial" based on a threshold value for growth inhibition and then train a graph neural network (GNN) that learns to link structural properties of the compounds with their antibacterial effect. In this way, we can, for example, screen fragment libraries, i.e. substructures in the range from 8 to 17 or from 11 to 17 atoms, to identify those with significant antibacterial activity. We then use these as starting points and apply a generative approach to generate many millions of possibilities.
We then evaluate these possibilities based on their predicted antibacterial activity, potential toxicity, manufacturability and other drug-like properties. We also calculate how similar the generative compounds are to existing antibiotics, with the aim of only developing compounds that have no structural similarity to existing antibiotics to ensure structural novelty.
We then work with a contract research organization for chemical synthesis to produce a selection of these compounds. We then carry out experimental in vitro tests with these synthesized compounds in our laboratory to see which of them show antibacterial activity. Compounds with proven activity are characterized in detail: by determining the minimum inhibitory concentration, by studying the development of resistance and analyzing the mechanisms of action, and by identifying potential targets. The most promising of these compounds are then followed up.

What do you think is the biggest advantage of using AI in antibiotic discovery and what limitations should be considered?
James Collins: I think probably the most exciting aspect of AI-assisted discovery and AI-assisted drug design is that these approaches allow us to explore a much larger chemical space, opening up completely new opportunities to find novel antibiotics.
I think one of the biggest challenges is that many very interesting compounds are often not easy to synthesize. That's a critical gap. But I think we can get better at developing chemical synthesis routes. This is probably the biggest limiting factor for generative AI in antibiotic discovery and design at the moment.

Have AI-generated antibiotics already been used outside the lab, i.e. in the real world? If not, what important steps are still needed to make real-world use possible?
James Collins: We have tested many of the compounds in our lab against bacterial pathogens and in mouse models.
We are also working with Phare Bio, a non-profit social venture that we helped to found. Phare Bio takes the most promising molecules and develops them towards clinical application together with biotech and pharmaceutical companies and with the help of funding from the Advanced Research Projects Agency for Health (ARPA-H). Together with Phare Bio, we are in the process of advancing 15 novel AI-generated compounds to IND-enabling studies (studies that enable the filing of an Investigational New Drug (IND) application).

Thank you very much for the interview, Professor Collins!