AlphaFold 3 predicts the structure and interactions of life's molecules

Original Article ↗ Hacker News Discussion ↗

Commercial strategy, access, and licensing

  • AF3 is released as a closed-source, non‑commercial web server with a 10‑predictions‑per‑day limit and restricted ligands; full capabilities (notably docking) are not exposed.
  • Many see “safety” justifications as pretext for monetization via Alphabet’s Isomorphic Labs, which is positioned as the commercial arm for drug discovery.
  • This shift from AF2’s open release to AF3’s closed model is criticized as a “tax” on academia and a missed chance for public-good science.
  • Others argue Alphabet is structurally bound to monetize and that capitalism’s incentives helped bring AF about in the first place.
  • Open alternatives (RoseTTAFold, OpenFold, RosettaFold All‑Atom, etc.) are cited; several posters expect AF3‑like models to be replicated and open‑sourced.

Capabilities, accuracy, and validation

  • AF3 extends beyond proteins to DNA, RNA, ligands, ions, and chemical modifications, and improves protein–protein complex modeling.
  • Critics note press materials emphasize relative gains but underplay absolute accuracy (e.g., ~70% on some tasks, ~50% better than traditional docking in cited reports).
  • AF2 and successors are widely regarded as “very good but not perfect”; remaining issues include disordered regions, domain orientations, side‑chain details, and out‑of‑distribution sequences.
  • Validation relies heavily on blind assessments like CASP (using unpublished experimental structures) and comparison to crystallography and cryo‑EM. AF predictions are often framed as “exceptionally useful hypotheses,” not ground truth.

Reproducibility, transparency, and Nature standards

  • Several posters highlight that AF3’s Nature paper provides only pseudocode and a hosted service, not runnable code or full training details.
  • This is seen as in tension with Nature’s stated policies on code availability and as harmful to reproducibility, with results now depending on a single proprietary server.

Safety, misuse, and biosecurity

  • Some argue tight control is justified because improved structure and docking prediction could ease design of engineered pathogens or gain‑of‑function work.
  • Others counter that serious actors already have the tools needed for bioweapons, and that limiting AF3 mostly protects commercial advantage rather than safety.

Scientific and philosophical implications

  • Extensive debate over black‑box ML vs interpretable, physics‑based models:
    • Pro‑ML voices emphasize empirical success: better predictions trump elegant but weaker theories.
    • Skeptics worry about overreliance on opaque models, lack of mechanistic understanding, and “epicycle‑like” science.
  • There is broad expectation that AF‑style models will drastically accelerate hypothesis generation in biology and drug discovery, even if they do not replace experimental work.