After training as a mathematician and physicist, Felix Wong, PhD, developed a keen interest in biology. He found the perfect fit in the lab of James J. Collins, PhD, at MIT, where he started using machine learning models to find new antibiotics and has since been a co-author on several research papers describing machine learning models to discover one-off hits.

But what Wong has been interested in is how to advance this approach to identify not only individual compounds but also entire classes of small-molecule drugs.

In a new study published in Nature, Wong and colleagues describe a step forward in their drug discovery approach: they down-sample chemical space with AI to identify interesting chemical scaffolds beyond individual small-molecule hits. For this work, the researchers pointed their strategy towards identifying broad classes of antibiotics, such as methicillin-resistant Staphylococcus aureus (MRSA) and other gram-positive bacteria. Wong and colleagues determined the antibiotic activities and human cell cytotoxicity profiles of 39,312 compounds and applied ensembles of graph neural networks to predict antibiotic activity and cytotoxicity for 12,076,365 compounds.

Wong predicts that Phare Bio, a non-profit social enterprise Collins co-founded that uses cutting-edge AI and deep learning to address the most urgent global threats, would use the research article’s findings.

But the real success of this research, Wong believes, is that this approach is widely applicable for many drug classes, including small molecules that selectively kill senescent cells, which have been implicated in cancer and aging. This concept inspired Wong to co-found Integrated Biosciences. This start-up is harnessing this approach to drug discovery to create a pipeline of clinical candidates targeting age-related diseases.

This study was a collaboration between Wong’s Integrated Biosciences and Collins’ labs at Wyss, Broad, and MIT.

Felix Wong Integrated Biosciences
Felix Wong, PhD, is the co-founder of Integrated Biosciences.

Out of the black box 

The Nature article demonstrates how large-scale phenotypic screening applies to drug discovery, a key component of research in Collins’ lab at MIT and Integrated Biosciences.

“It’s not just a one-disease, one-target approach, which seems to be pretty much a consensus approach nowadays,” said Wong. “If you have phenotypic screens, that allows you to integrate information across many diverse targets, which is quite important. It’s kind of how antibiotics were discovered traditionally. Still, it’s also important for things like age-related diseases, where things are so complicated that you don’t know what is going on or which targets to drug.”

A key innovation of the paper is that the first instance of an explainable AI approach to drug discovery leads to the structural classification of small molecules, in this case for antimicrobials.

Employing explainable AI-driven drug discovery can be powerful in providing a computationally efficient way to down-sample chemical space. Instead of discovering one-off chemical hits, explainable AI can discover entire structural classes of compounds that have efficacy for things like antibiotic or senolytic activity.

Explainable AI is a set of tools and frameworks to help the user understand and interpret natively predictions made by machine learning models. With it, users can debug and improve model performance and help others understand the models’ behavior.

Many AI models typically used not just in drug discovery but also in things like ChatGPT are complicated models that are black boxes—they’re trained on a bunch of data and then output something without an explanation.

Explainable AI has an added feature that also outputs a method for its madness. For example, this would be like ChatGPT saying how it was deconstructing a prompt to output each kind of paragraph that it does in its output. For drug discovery, what this entails is, at least for small molecules, identifying chemical structure substructures that are important for the predicted biological activity.

“This really is a paradigm shift in a way, in the sense that we’re not really just looking for one-off hits,” said Wong. “Now we’re kind of looking for salient predictions of entire chemical substructure motifs that have activity.”

Shine a light

In the middle of 2022, Wong co-founded Integrated Biosciences with Max Wilson, PhD, a professor at the University of California, Santa Barbara, to discover drugs that target age-related stress responses.

“My postdoc in Jim’s lab was coming to a close, and we also had these articles on antibiotic discovery using machine learning coming out,” said Wong. “So, it was quite immediate to [Max and I] that we could apply the same methods that we were developing… to found a company that focuses on using small molecules to modulate these aging pathways, with the idea that synthetic biology allows us to get very specific and clean data that we can now use to train these machine learning models.”

At Integrated Biosciences, Wong and Wilson combined their experience with explainable AI-driven drug discovery and optogenetics to control aging cells using light to find new small molecules that target these age-related stressors and other pathways.

“What really excites us about optogenetics is that we can now non-invasively modulate aging and age-related pathways in cells,” said Wong. “We can do this in a way that’s spatially or temporally modulated, where we can make certain cells age faster just by shining light on them. We can also turn the light on and off to make cells age in pulses, which is quite interesting. It allows for new types of dynamic screening.”

Earlier this year, Wong and Wilson published a paper describing their novel optogenetic technique to study the integrated stress response in live cells without physical or chemical damage. The paper demonstrates that cells make intricate computations to determine their stress response based on their past experiences. Combining this optogenetics platform with explainable AI is the core of how Integrated Biosciences identifies candidate anti-aging drugs.

Infinite space

Wong said that Integrated Biosciences was lucky to have investor interest and to be relatively well funded, with a financial runway that should support their small team of six people. “We’re not swimming in cash,” said Wong. “We’re putting in a lot of work per capita and trying to deliver value.”

In addition to using small molecule decks to screen for candidates, Wong said that Integrated Biosciences has teamed up with Dave MacMillan, PhD, to use his Nobel Prize-winning organocatalysis technology.

“We can use organic catalysis to generate really funky compounds that are not in anything that one can run even for big pharma,” said Wong. “We combine chemical information sources and run all the screens ourselves. We have very targeted synthetic biology constructs to get the information we want.”

Combining these screens with training best-in-class models, Integrated Biosciences is classifying drug chemical space, which isn’t just big—it’s essentially infinite.

“The main idea is that we can use information to have more accurate and predictive machine learning models when you’re mining this infinite tube of chemical space,” said Wong.

Wong insists Integrated Biosciences must derive value from moving assets through the clinic.

“We’re a bunch of scientists who know how to do the science,” said Wong. “That’s one of the things that we want to focus on for our series A and beyond, which is that we still want to do basic research and innovate in the way we present in the Nature paper. We’ve been making a lot of progress, not just in terms of paper but also in things that are unannounced, like the development of lead assets and working out partnerships with other larger companies. And we hope to have more leads in 2024.”

Previous articleMolecular Jackhammers Destroy Cancer Cells with Vibration
Next articleAdvancing Automation for Process Development