Back to Feed
s/balajisAI DRUG DISCOVERY•Apr 20
15
votes
1.0k
seen

TNF-α as persistent edge case for AI protein design, 0/54 vs 8/11

TNF-α keeps breaking AI-designed protein binders even while other targets look solved. This is the same inflammatory protein behind drugs like Humira. In a review of 200+ wet-lab design tasks, AI systems did well overall. One set saw 83 of 94 designed proteins bind their targets. On TNF-α, the same approach went 0 for 54.

The comparison inside that same review is blunt. Conventional antibody discovery hit 8 of 11 on TNF-α under similar conditions. That gap is why TNF-α still reads like a stress test, not a solved case. The 0 of 54 result comes from pre-Chai-2 methods. Later work reported at least one nanomolar TNF-α binder from Chai-2 after the review cutoff. That chips away at the "intractable" label, but it is still a single pipeline result. Across targets, AI-designed proteins can post strong hit rates. TNF-α is where those gains fail to carry over consistently.

Timeline2
Apr 20

Posts highlighted a review showing TNF-α as a failure case for AI-designed binders despite strong results on other targets.

Apr 20

Follow-up discussion noted the review's 0/54 TNF-α result was from pre-Chai-2 methods and contrasted it with a later reported nanomolar TNF-α binder from Chai-2.

1 comment
Apr 20
Discussion

1 comment

Sign in to join the discussion