Protein Folding

Protein folding is the process by which a polypeptide chain — a one-dimensional sequence of amino acids — reaches the three-dimensional structure that lets it do its biological job. The thermodynamic version of the problem is the Levinthal problem: the number of accessible conformations is astronomical, yet folding completes on a millisecond-to-second timescale, which means the energy landscape is funnelled rather than flat. The kinetic version is harder still: folding in vivo happens on the ribosome, against a crowded cytosol, with the help of chaperones that shepherd nascent chains away from aggregation, and sometimes against active perturbations from membranes, osmolytes, or post-translational modifications. Methodological work in the modern era organises around four axes: physics-based and learned models of the energy landscape (the descendants of force fields, now coupled to generative models from machine learning); cotranslational folding (recognising that the chain folds while it is still being synthesised, with the ribosome and its associated chaperones part of the energy landscape); non-trivial folding behaviour (proteins that adopt two folds, fold in asymmetric environments, or refold in response to the cellular state); and the chemical environment (how small-molecule osmolytes, lipids, and stressors shift folding equilibria).

Generative models meet physics-based simulation

Coarse-grained molecular dynamics has long been the workhorse for sampling the protein-folding energy landscape, but its accuracy is bounded by the quality of the force field, and traditional force fields are difficult to train on the high-quality data that now exists. Arts et al. (2023) close that gap with “Two for One”: diffusion models that simultaneously act as generative samplers and as learnable force fields for coarse-grained simulations. The trick is that a denoising diffusion model trained on equilibrium configurations encodes the underlying free-energy surface implicitly, so the same network that generates plausible configurations from noise also defines a force field that can be used in molecular dynamics. The result is a clean methodological route to learn force fields directly from structural data, rather than fitting them to a small set of hand-curated reference systems, and it places generative modelling alongside replica exchange and metadynamics as a tool for sampling folding landscapes. Schafer et al. (2023) use a complementary methodology — large-scale structural and evolutionary analysis — to identify proteins that have been evolutionarily selected to adopt two folds from the same sequence, a phenomenon that classical funnel theory does not predict. The finding sharpens the connection between sequence and structure: in some regions of sequence space, evolution actively prefers ambiguity, and the resulting two-fold proteins are a strong test case for any predictive model that assumes a single ground-state structure.

Cotranslational folding and chaperone coordination

Folding does not start in the test tube; it starts on the ribosome, as the nascent chain extrudes through the exit tunnel. Streit et al. (2024) make this concrete by showing that the ribosome lowers the entropic penalty of protein folding: by restricting the conformational ensemble of the nascent chain through its tethering to the ribosomal surface, the ribosome acts as a part of the folding apparatus rather than as a passive synthesis machine. The thermodynamic consequence is that some folding events become favourable on the ribosome that would be unfavourable in solution. Roeselová et al. (2024) extend the picture to the chaperone layer, dissecting the mechanism of chaperone coordination during cotranslational folding in bacteria: trigger factor, DnaK, and downstream chaperones engage the nascent chain at distinct extents of synthesis, and the handovers between them are themselves a regulatory step that determines folding outcome. Together the two papers refocus the folding problem from “the equilibrium structure of a complete chain in dilute solution” to “the trajectory of a chain through synthesis, chaperone-binding, and release”, which is the regime in which most folding actually happens.

Folding in non-trivial environments

The classical folding problem assumes a homogeneous aqueous environment, but biology folds in many environments at once. Machin et al. (2023) show that protein–lipid charge interactions control the folding of outer membrane proteins into asymmetric membranes: the two leaflets of an outer membrane carry different charges, and that asymmetry biases the orientation and folding kinetics of nascent β-barrel proteins inserting into the membrane. The work converts the membrane from a passive solvent into an active component of the folding energy landscape. Pepelnjak et al. (2024) attack the small-molecule side of the same question with an in situ analysis of how osmolytes thermally stabilise proteomes: by combining limited proteolysis with thermal denaturation across an entire proteome, they identify which proteins each osmolyte stabilises and propose mechanisms for the selectivity. The result is a system-wide picture of how the chemical environment of the cell shifts folding equilibria, and it gives a template for studying any small molecule (drug, chaperone, stress response metabolite) that perturbs the proteome’s folded state. Open methodological questions span all four axes: can learned force fields be made transferable across protein families, do cotranslational folding trajectories converge to the same structure as refolding in vitro, and how should two-fold proteins be represented in structure-prediction models that currently assume one ground state per sequence?