AISAR finds hidden protein states

Proteins are not statues. They flex, breathe, and sometimes slip into rare shapes that matter more than their most common one. The problem is that structural biology usually gives you one tidy answer, while the molecule in solution is doing several things at once. That is the gap AISAR is trying to close. A paper published April 27 in Nature Communications says it can recover hidden protein states by generating lots of realistic structures with AI, then asking which combination best matches NMR data. (nature.com) ### What is AISAR actually doing? AISAR stands for AI Sampling with NMR Recall selection. The basic move is simple — instead of forcing experimental NMR signals into one final structure, it starts with a big pool of plausible conformers generated by AI-based sampling and then scores which sets of conformers best explain the observed NOESY peaks and other NMR observables. In other words, the models come first, and the experiment filters them afterward. (nature.com) ### Why is that different from standard NMR structure work? Traditional protein NMR workflows usually convert NOE signals into distance restraints and then build structures that satisfy those restraints. That works well when the protein behaves like one dominant shape. But if the sample is flickering between states, the restraints can get averaged together into something that is internally neat and(nature.com) by selecting multiple conformers that collectively explain the data. (nature.com) ### Why use AlphaFold-style models here? Because AI models are good at proposing physically realistic structures, even when experiments alone do not cleanly separate them. The earlier preprint version called the approach “AlphaFold-NMR,” and the final paper broadens that into AISAR. The key idea stayed the same — use enhanced AI sampling to generate a diverse structural ensemble, then let experiment decide which members are real contenders. (biorxiv.org) ### What did they show in a real protein? The clearest demo is Gaussia luciferase, or GLuc, a small enzyme that was already suspected to be dynamic. AISAR pulled out two interconverting states with large rearrangements in two lid regions, along with changes in binding pockets and cryptic surface cavities. One of those states is more open, exposing features that a single-state mod(biorxiv.org)lation start to get interesting. (nature.com) ### What is the strongest piece of evidence? The paper leans on NOESY Double Recall, which is basically a way to ask whether one structural state explains one subset of NMR contacts while another state explains a different subset. In GLuc, the analysis identified roughly 158 NOEs unique to state 1 and roughly 175 unique to state 2. That is the kind of split you want if you are arguing that the “extra” state is not just noise or overfitting. (nature.com) ### Was it only shown in one protein? No. The team also reports two distinct conformational states for the human tumor suppressor Cyclin-Dependent Kinase 2-Associated Protein 1, or CDK2AP1. That does not prove universal generality, but it does show the method is not locked to one quirky luciferase example. Different scaffold, same basic payoff — hidden states that single-state analysis can miss. (nature.com) ### So why does this matter beyond NMR specialists? Because a lot of protein function lives in low-population states — the brief shapes that open a pocket, expose an interface, or prepare an enzyme to react. If those states stay invisible, you can end up designing drugs, mutations, or mechanistic stories around the wrong structure. AISAR is basically a bridge between AI structure generation and experiment-driven reality checks. (nature.com) ### What is the bottom line? The news here is not that AI predicted another protein structure. It is that AI was used as a search engine for structural possibilities, and NMR was used to pick the ones the molecule actually visits. For dynamic proteins, that is a much more useful question. (nature.com)