Yong is a scientist at The Hospital for Sick Children in Toronto, working on molecular biology, studying cell volume, and targeted kinase degradation. His institution's default AI tool is a general enterprise assistant, and he had worked with three widely used general purpose AI tools alongside it. He came to EMET, BenchSci's agentic workbench for preclinical R&D, with a real basis for comparison.
Every question he put to EMET came from work already on his bench: an uncharacterized gene set, a pan kinase inhibitor he was tracking, and a PROTAC he wanted to design but lacked the chemistry background to build.
"If this works well, that will be a huge benefit to scientists like me, because I have no idea how to design a PROTAC."
A PROTAC prototype, from a scientist who is not a chemist
Yong wanted to use a kinase inhibitor that binds the ATP pocket as the warhead in a PROTAC. He described the objective and asked EMET for a prototype. It returned three candidate designs, each with its SMILES string.
Then he tested it. He ran the SMILES strings through a docking simulation against a predicted structure. The compounds bound the ATP pocket.
"I'm pretty much a novice at generating this kind of chemical. Honestly, I cannot do it without any help from AI."
The conventional route means commissioning analog generation from an outside chemistry group, which he put at several weeks and a significant budget line. EMET produced three, and the evidence to check them.
"This is awesome."
The task where the gap was clearest
Yong had put the same PROTAC prompt to one of the general purpose AI tools he uses. It came back with an explanation of the method and a list of software he might try.
"It did nothing. It did nothing. It showed me how to design a PROTAC and which tools I could leverage. But honestly, that's not the information I wanted to see."
EMET, on the same prompt, built them.
"The thing is, it actually implements what I prompt about. That's a good example of how it outperforms [the other tools] for research purposes."
"That's probably the most astonishing point to me, because the designing part is not trivial."
Evidence with the methods attached
Asked about a physical parameter across a panel of cell lines, Yong wanted more than a list of papers. He wanted the methods buried inside them.
"We need deeper information. Which cell lines were actually used, which resources were used in those papers, how they measured it, which software tool they used to analyze it."
That is what EMET returned, drawn from BenchSci's knowledge graph. "Your KG really worked well," he said. It showed the same instinct in his gene set, running GO term analysis unprompted. "Surprisingly, it does GO term analysis automatically. That's very awesome."
He checked its work
Yong also probed EMET, where he already knew the answer: clinical trial activity for a pan-kinase inhibitor he follows. Trials underway, no results. EMET said the same.
He did the same with the PROTAC candidates, taking the SMILES strings out of EMET and docking them himself rather than accepting the designs on trust. That is the posture the workbench is built for. Yong sets the objective, EMET does the work, and he checks it before he acts on it.
The story is narrower and more useful than "AI accelerates research." A scientist with no chemistry background reached three credible starting compounds and the means to evaluate them in an afternoon.
If you are an academic scientist looking for access to EMET be sure to sign up using the following link: https://emet.benchsci.com/
Want to explore what’s possible?
Try EMETRelated articles
“As an enabler and as a time saver for me, it has transformed my work day”: a TB drug discovery leader with EMET
A third of his time back, and a protein he can finally detect
