Provocative but serious. If prediction is near-experimental for many proteins, do we still need crystallography/cryo-EM at scale, or does the wet lab become a niche validation step? Genuinely curious where practitioners land.
No — and not close. Prediction gives you a static, single conformation. Experiments give you dynamics, ligand-bound states, post-translational modifications, and ground truth for novel folds where prediction is least reliable. It moves the bottleneck; it doesn't remove it.
@hcole That's the strongest case against my framing. So the claim should be narrower: it automates the 'rough static structure' step, freeing experiments for the hard parts (dynamics, complexes, mechanism)?
@dortiz Exactly that. It's a force multiplier — you spend wet-lab time where it actually adds information instead of grinding out structures a model could have given you.
From the drug side: a predicted apo structure is a starting point, but you still need experimental structures of the complex with your candidate to do real medicinal chemistry. Induced fit and cryptic pockets are exactly where predictions are shakiest.
Historically, automating a step increases total demand for the surrounding science, not decreases it (Jevons-style). AlphaFold made structure cheap, which exploded the number of downstream questions that now need experimental follow-up. More prediction → more (targeted) experiments.
@lenaf This is the take I trust most. 'Replace' is the wrong verb. The honest prediction is reallocation: fewer routine structures, far more functional/dynamic studies.
Statistical caution that's relevant here: the impressive accuracy is averaged over benchmarks dominated by MSA-rich proteins. The proteins you can't easily crystallize overlap heavily with the ones prediction is worst at (orphans, disordered regions). So predictions are weakest exactly where experiments are most needed — which argues against obsolescence.
Worth saying the two are increasingly combined, not competing: predicted models are used as search models for molecular replacement and to fit cryo-EM density faster. The future looks like AI + experiment in one loop, not one killing the other.
@amir_r Yes — predicted models accelerating phasing/fitting is already standard. That's the clearest evidence it's a tool inside the experimental pipeline, not a replacement for it.
한 가지 더: '예측이 충분히 정확하다'는 판단 자체를 검증하려면 결국 실험이 필요해요. 새로운 폴드일수록 신뢰도를 믿을 근거가 약하니까, 실험은 최소한 calibration 용도로라도 계속 남을 수밖에 없을 것 같습니다.