Representing local protein environments with atomistic foundation models
１．This paper introduces a novel method to represent local protein environments using embeddings derived from atomistic foundation models (AFMs). These representations capture both structural and chemical details of proteins with high fidelity.
２．The authors show that AFM-derived embeddings naturally encode secondary structure motifs and amino acid identities, outperforming handcrafted descriptors in accuracy and generalizability.
３．A key highlight is the use of these representations to build a chemical shift predictor for NMR spectroscopy. The model achieves state-of-the-art accuracy and, for the first time, provides confidence estimates using data-driven likelihoods in the AFM embedding space.
４．These embeddings enable computation of meaningful similarity metrics and likelihoods over protein environments. Such metrics allow unsupervised clustering, anomaly detection, and fine-grained structural quality assessments.
５．The concept of canonical local environments is introduced: a residue-centered region including nearby atoms (within 5Å), encoded using backbone atoms and bonded hydrogens. This standardization ensures representations are comparable across residues.
６．AFM embeddings respond sensitively and smoothly to local structural perturbations such as sidechain rotations or secondary structure transitions. This suggests the latent space is physically meaningful and continuous.
７．The authors benchmark multiple AFM families (MACE, Egret, OrbNet, AIMNet) across prediction tasks. MACE embeddings perform best overall, while AIMNet excels in pKa prediction due to multitask training.
８．In a case study, the model correctly captures ring current effects on chemical shifts from aromatic residues, reproducing the expected 180° periodicity—something existing models fail to do.
９．They show that AFM-derived likelihoods can quantify prediction uncertainty: lower-likelihood environments have systematically higher chemical shift errors, making this a valuable tool for downstream confidence estimation.
１０．Finally, the embeddings even support partial reconstruction of protein structures. Using only the descriptor, an optimization process can guide AlphaFold3 to generate reasonable alternative conformations.
📜Paper:
arxiv.org/abs/2505.23354v1 #ProteinStructure #NMR #MachineLearning #AFM #ComputationalBiology #StructuralBiology
