Introduction
For two decades, the central goal of structural biology was to determine the structure of a protein — a single set of coordinates. Deep learning has largely solved that problem: AlphaFold2 and its successors now predict static folds at near-experimental accuracy for most proteins. But proteins are not statues. They breathe, flex, and cycle between functionally distinct conformational states — open/closed enzymes, active/inactive receptors, inward/outward-facing transporters, folded/disordered regions. Function is motion. A single static coordinate set captures, at best, one snapshot of a distribution, and can actively mislead when the biologically relevant state is a minor or transient one. This survey organizes the field by approach family — from static baselines through physics-based simulation, generative machine-learning ensemble models, cryo-EM heterogeneity analysis, and integrative multi-state modeling of disordered systems — giving for each what it is, how it works, key tools, evidence, strengths, and limitations, closing with an honest assessment of what is actually winning.
Family 1 — Static structure prediction as the baseline
What it is. End-to-end deep-learning models that map a protein sequence (plus, usually, an MSA) to a single 3D structure — the modern baseline every dynamics method builds on or reacts against.
Mechanism. AlphaFold2 uses an attention-based "Evoformer" over MSA and pairwise residue representations feeding a structure module that produces atomic coordinates end-to-end. RoseTTAFold uses a three-track architecture coupling 1D/2D/3D reasoning. ESMFold replaces the MSA with a protein language model (ESM-2), predicting structure from a single sequence at metagenomic scale. AlphaFold3 generalizes to a diffusion-based generative head over all-atom coordinates, extending to protein–ligand, protein–nucleic-acid, and multi-chain complexes.
Key tools. AlphaFold2; AlphaFold3; RoseTTAFold / RoseTTAFold All-Atom; ESMFold (ESM-2); open re-implementations (OpenFold; Boltz-1/2).
Evidence. AlphaFold2 achieved median backbone accuracy near experimental error at CASP14; ESMFold folded hundreds of millions of metagenomic sequences; AlphaFold3 extended high accuracy to biomolecular interactions.
Limitations — why a single structure is insufficient for dynamics. These models are trained to reproduce the one deposited PDB structure and by default return a single high-confidence conformation. They do not report populations, do not sample alternative functional states, and give no free energies or kinetics. For allosteric enzymes, GPCRs, transporters, and disordered regions — where the distribution of states is the biology — a single prediction is at best incomplete. This gap motivates every family below.
Citations. Jumper et al. 2021, Nature — 10.1038/s41586-021-03819-2 (PMID 34265844); Abramson et al. 2024, Nature (AlphaFold 3) — 10.1038/s41586-024-07487-w (PMID 38718835); Baek et al. 2021, Science (RoseTTAFold) — 10.1126/science.abj8754; Lin et al. 2023, Science (ESMFold) — 10.1126/science.ade2574.
Family 2 — Physics-based molecular dynamics & enhanced sampling
What it is. Simulation of protein motion by numerically integrating Newton's equations under a molecular force field. MD is the only family that produces genuine time-ordered trajectories — real "movies" — with physically grounded free energies and kinetics.
Mechanism. An empirical force field (AMBER, CHARMM, OPLS) assigns energies to bonds/angles/torsions/non-bonded terms; atoms are propagated in femtosecond steps. Because functional transitions occur on µs–s timescales, the field relies on enhanced sampling: metadynamics adds history-dependent bias along collective variables; replica-exchange MD (REMD) swaps temperature replicas to cross barriers; accelerated MD raises low-energy regions to speed rare events. Markov State Models (MSMs) stitch many short trajectories into a kinetic network of metastable states. AlphaFold-seeded MD uses predicted structures as diverse starting points. Special-purpose hardware (Anton, D. E. Shaw) reaches millisecond all-atom trajectories directly.
Key tools. Engines: GROMACS, AMBER, CHARMM, NAMD, OpenMM. Force fields: ff14SB, CHARMM36. Enhanced sampling: metadynamics (PLUMED), REMD, accelerated MD, weighted-ensemble/string methods. MSMs: PyEMMA, MSMBuilder. Hardware: Anton.
Evidence. Anton produced the first atomic-level millisecond folding and conformational-change trajectories; MSMs matured into a rigorous statistical framework; metadynamics/REMD/aMD are standard, heavily validated free-energy and rare-event methods.
Strengths. Physically principled; yields free energies, kinetics, and mechanism — the gold standard for how a protein moves, transition pathways, and coupling (e.g., ion/substrate coupling in transporters, ligand-induced allostery).
Limitations. Expensive (large systems / long timescales); force-field inaccuracies (especially IDPs and some ion interactions); enhanced sampling needs well-chosen collective variables; large collective transitions can remain under-sampled. This cost/timescale wall is what generative ML tries to leapfrog.
Citations. Laio & Parrinello 2002, PNAS (metadynamics) — 10.1073/pnas.202427399; Sugita & Okamoto 1999, Chem Phys Lett (REMD) — 10.1016/S0009-2614(99)01123-9; Hamelberg et al. 2004, JCP (aMD) — 10.1063/1.1755656; Husic & Pande 2018, JACS (MSM) — 10.1021/jacs.7b12191; Shaw et al. 2010, Science (Anton) — 10.1126/science.1187409; Abraham et al. 2015, SoftwareX (GROMACS) — 10.1016/j.softx.2015.06.001; Maier et al. 2015, JCTC (ff14SB) — 10.1021/acs.jctc.5b00255.
Family 3 — Generative / ML conformational ensemble models
What it is. Deep generative models that sample many conformations for a sequence, aiming to approximate the equilibrium (Boltzmann) distribution — populations included — orders of magnitude faster than MD. Crucial caveat: most are equilibrium-ensemble generators (states and populations, no time axis), not true trajectory/"movie" models. Only a few (e.g., MDGen) generate time-ordered dynamics, and those remain small-scale.
Mechanism. Sub-strategies: (a) MSA manipulation of AlphaFold — subsampling/clustering the input MSA induces AF2 to output alternative functional states (AF-Cluster; subsampling recovers transporter/receptor endpoints and approximate populations). (b) Repurposing folding networks as samplers — AlphaFlow/ESMFlow retrain under a flow-matching objective; EigenFold and Str2Str use diffusion/score-based sampling. (c) Learned Boltzmann samplers — Boltzmann Generators (normalizing flows with exact reweighting); DiG (Distributional Graphormer) predicts equilibrium distributions; BioEmu emulates equilibrium ensembles at scale. (d) Force-guided diffusion — ConfDiff injects force guidance into SE(3) diffusion. (e) Trajectory generation — MDGen models MD trajectories directly (true dynamics, tetrapeptide-scale).
Key tools. AF-Cluster; MSA-subsampling (del Alamo; AFsample2; subsampled AF2); AlphaFlow/ESMFlow; EigenFold; Str2Str; ConfDiff; Boltzmann Generators; DiG; idpGAN; BioEmu; MDGen.
Evidence. MSA-subsampled AF2 recovers both outward- and inward-facing transporter states across many transporters and GPCRs. BioEmu reports thousands of conformers per GPU-hour, capturing cryptic pockets and domain motions at accuracy approaching long MD but ~10⁵× faster. Boltzmann Generators demonstrated correctly reweighted equilibrium sampling. AF-Cluster resolved a known fold-switching protein's two states.
Strengths. Dramatically cheaper than MD; produce diverse conformers and (for the best methods) calibrated populations; excellent for cryptic-pocket discovery and structural-prior/hypothesis generation.
Limitations. Most give no kinetics or time ordering; population calibration is uneven; large collective transitions (full transporter alternating-access) are systematically under-sampled by equilibrium generators; many are single-chain (BioEmu is single-chain, Cα-level); accuracy is bounded by PDB training bias toward stable states.
Citations. Wayment-Steele et al. 2024, Nature (AF-Cluster) — 10.1038/s41586-023-06832-9; del Alamo et al. 2022, eLife (MSA subsampling) — 10.7554/eLife.75751; Kalakoti & Wallner 2025, Commun Biol (AFsample2) — 10.1038/s42003-025-07791-9; Monteiro da Silva et al. 2024, Nat Commun — 10.1038/s41467-024-46715-9; Zheng et al. 2024, Nat Mach Intell (DiG) — 10.1038/s42256-024-00837-3; Noé et al. 2019, Science (Boltzmann Generators) — 10.1126/science.aaw1147; Janson et al. 2023, Nat Commun (idpGAN) — 10.1038/s41467-023-36443-x; Jing et al. 2023 (EigenFold) — arXiv:2304.02198; Lu et al. 2024 (Str2Str) — arXiv:2306.03117; Jing et al. 2024 (AlphaFlow/ESMFlow) — arXiv:2402.04845; Wang et al. 2024 (ConfDiff) — arXiv:2403.14088; Lewis et al. 2025, Science (BioEmu) — 10.1126/science.adv9817; Jing et al. 2024 (MDGen) — arXiv:2409.17808.
Family 4 — Cryo-EM conformational heterogeneity
What it is. Methods that extract multiple conformational states — or continuous motion — from a single cryo-EM dataset, treating structural heterogeneity as signal. Each of the hundreds of thousands to millions of imaged particles is a frozen snapshot of a distinct conformation; the task is to recover the distribution.
Mechanism. cryoDRGN trains a VAE that learns a continuous low-dimensional latent space of conformations; cryoDRGN-AI extends this to neural ab initio reconstruction. 3D Variability Analysis (3DVA) in cryoSPARC solves for linear variability components (continuous flexibility + discrete states). 3DFlex models the map as a deformable canonical density plus a learned flow field. ManifoldEM places particle snapshots on a manifold to retrieve functional pathways and energy landscapes. RECOVAR uses regularized covariance estimation with kernel regression. DynaMight estimates per-particle deformations.
Key tools. cryoSPARC (+ 3DVA + 3DFlex); cryoDRGN / cryoDRGN-AI; ManifoldEM; RECOVAR; DynaMight; RELION (classification-based discrete heterogeneity).
Evidence. cryoDRGN resolved compositional and conformational heterogeneity in ribosome and spliceosome datasets; 3DFlex resolved continuous membrane-protein motion (e.g., TRPV1); ManifoldEM recovered functional energy landscapes from experimental snapshots.
Strengths. Grounded in real experimental data (not a model prior); resolves both discrete states and continuous motion; the definitive way to observe the conformational distribution a specific sample actually populates; extends to native context via cryo-ET.
Limitations. Requires a good particle stack; states below the reconstruction floor (rare conformers, or those lost to vitrification / air-water-interface bias) never enter the data; independent comparisons show learned deformation densities can diverge from ground truth, so orthogonal validation (smFRET, HDX-MS) is advisable.
Citations. Zhong et al. 2021, Nat Methods (cryoDRGN) — 10.1038/s41592-020-01049-4; Levy et al. 2025, Nat Methods (cryoDRGN-AI) — 10.1038/s41592-025-02720-4; Punjani & Fleet 2021, J Struct Biol (3DVA) — 10.1016/j.jsb.2021.107702; Punjani & Fleet 2023, Nat Methods (3DFlex) — 10.1038/s41592-023-01853-8; Dashti et al. 2020, Nat Commun (ManifoldEM) — 10.1038/s41467-020-18403-x; Gilles & Singer 2025, PNAS (RECOVAR) — 10.1073/pnas.2419140122; Schwab et al. 2024, Nat Methods (DynaMight) — 10.1038/s41592-024-02377-5; Punjani et al. 2017, Nat Methods (cryoSPARC) — 10.1038/nmeth.4169.
Family 5 — Intrinsically disordered proteins & integrative multi-state modeling
What it is. Methods for proteins (or regions) with no single folded structure — intrinsically disordered proteins (IDPs) — and integrative approaches combining sparse experimental measurements into an ensemble or multi-state model. Here the "structure" is fundamentally a distribution.
Mechanism. IDP ensembles are built by generating large candidate-conformer pools then selecting/reweighting a sub-ensemble whose averaged observables match experiment. Flexible-meccano generates disordered pools from residue-specific backbone statistics; the Ensemble Optimization Method (EOM) selects a sub-ensemble to fit SAXS; ENSEMBLE optimizes against combined NMR (chemical shifts, RDCs, PREs) and SAXS restraints. More broadly, the Integrative Modeling Platform (IMP) formalizes Bayesian combination of heterogeneous data (EM, cross-linking MS, SAXS, FRET, prior structures) to determine multi-state structures of large assemblies. ML generators (idpGAN) now emit disordered ensembles directly.
Key tools. Flexible-meccano; EOM (SAXS); ENSEMBLE (NMR+SAXS); IMP (integrative/Bayesian); idpGAN; NMR/SAXS-restrained MD.
Evidence. SAXS-based EOM and NMR-restrained ensembles are the established route to quantitative IDP descriptions; IMP underlies landmark integrative structures of large assemblies (e.g., the nuclear pore complex).
Strengths. Directly experiment-restrained; principled treatment of proteins with no single structure; IMP scales to very large heterogeneous assemblies; naturally multi-state.
Limitations. Under-determined — many ensembles fit the same averaged data, so results need maximum-entropy/regularization and are not unique; heavily dependent on data quality/quantity; labor-intensive; force-field and pool-generation biases propagate.
Citations. Russel et al. 2012, PLoS Biol (IMP) — 10.1371/journal.pbio.1001244; Bernadó et al. 2007, JACS (EOM) — 10.1021/ja069124n; Krzeminski et al. 2013, Bioinformatics (ENSEMBLE) — 10.1093/bioinformatics/bts701; Ozenne et al. 2012, Bioinformatics (Flexible-meccano) — 10.1093/bioinformatics/bts172; Jensen et al. 2014, Chem Rev — 10.1021/cr400688u.
What's Winning / The Honest State
- No single method wins; the field is converging on hybrids. The most reliable multi-state models in 2026 combine families: a fast generative or MSA-subsampled prior proposes states, physics-based MD/enhanced sampling gives mechanism and free energies, and experimental data (cryo-EM heterogeneity, NMR/SAXS, smFRET, HDX-MS) validates and anchors. Any purely computational ensemble is a hypothesis until experimentally checked.
- Static prediction is effectively solved; dynamics is not. AlphaFold-class models give near-perfect single structures but no distributions, populations, kinetics, or alternative functional states by default.
- Generative ensemble models are the fastest-moving frontier — with real caveats. BioEmu, DiG, AlphaFlow, MSA-subsampling deliver conformers at a tiny fraction of MD cost and are excellent for cryptic-pocket discovery. But most produce equilibrium ensembles, not trajectories; population calibration is uneven; large collective transitions are under-sampled. "Protein movie" claims are frequently over-stated — true trajectory generators (MDGen) remain peptide-scale.
- MD remains the gold standard for mechanism — the only family giving time-ordered pathways, kinetics, and free energies — at a cost generative models race to undercut.
- Cryo-EM heterogeneity is the strongest experimental grounding — cryoDRGN(-AI), 3DVA/3DFlex, RECOVAR, DynaMight recover multi-state distributions directly from data, but only for states that enter the particle stack, and learned densities need orthogonal validation.
BioMate Positioning
BioMate operates here as an execution platform, not a new algorithm. On its AWS-Batch compute layer it provides indexed, production cryo-EM workflows — including single-particle processing and conformational-heterogeneity analysis (the cryoSPARC / cryoDRGN-family methods of Family 4) — letting users resolve multiple states and continuous motion directly from their own particle data. It additionally offers a generative conformational-ensemble capability (Family-3-style equilibrium ensemble generation) producing many candidate conformers for a sequence, useful for cryptic-pocket discovery and as structural priors. BioMate's contribution is making these published methods runnable end-to-end at scale with live progress and downloadable results — orchestration and accessibility of the state of the art, not a proprietary replacement for it. The honest boundaries described throughout (equilibrium ensembles are not trajectories; computational ensembles need experimental validation; large collective transitions are hard) apply equally to results produced on the platform.
Verification note: all references confirmed against primary sources (publisher DOI, PubMed, or arXiv); no DOIs fabricated. Companion showcase: blog-protein-conformation.html. Related BioMate surveys: target identification, patient stratification, multi-omics integration, virtual cell modeling.