--- id: protein-structure-prediction title: "Protein Structure Prediction: AlphaFold, RoseTTAFold, and AI-Driven Structural Biology" schema_type: article category: ai language: en confidence: high last_verified: "2026-05-24" created_date: "2026-05-24" generation_method: ai_assisted ai_models: - claude-4.5-sonnet derived_from_human_seed: true conflict_of_interest: none_declared is_live_document: false data_period: static completeness: 0.85 atomic_facts: - id: af-protein-structure-prediction-1 statement: >- Nature (May 2026) published a commentary identifying the next two frontiers of AI in structural biology: predicting full protein conformational landscapes (all functionally-relevant shapes a protein adopts) and routine de novo design of proteins with novel functions — going beyond static single-structure prediction that AlphaFold2 achieved in 2020. source_title: Nature Communications Biology (2026) — The latest AI breakthroughs in structural biology — doi:10.1038/s42003-026-10112-3 source_url: https://www.nature.com/articles/s42003-026-10112-3 confidence: high - id: af-protein-structure-prediction-2 statement: >- ScienceDirect (August 2025) published a comprehensive review of AI methods for protein folding and design — documenting that models like AlphaFold2, RoseTTAFold, and ESMFold achieve sub-angstrom accuracy for single-chain predictions but still struggle with conformational ensembles, intrinsically disordered regions, and protein dynamics that are critical for understanding enzyme catalysis and allosteric regulation. source_title: ScienceDirect Current Opinion in Structural Biology (2025) — AI methods for protein folding and design — doi:10.1016/j.sbi.2025.102985 source_url: https://www.sciencedirect.com/science/article/abs/pii/S0959440X25000843 confidence: high primary_sources: - id: ps-protein-structure-prediction-1 title: "The latest AI breakthroughs in structural biology: protein conformational landscapes and de novo design" type: academic_paper year: 2026 institution: Nature Communications Biology doi: 10.1038/s42003-026-10112-3 url: https://www.nature.com/articles/s42003-026-10112-3 - id: ps-protein-structure-prediction-2 title: Artificial intelligence methods for protein folding and design type: academic_paper year: 2025 institution: Current Opinion in Structural Biology / Elsevier url: https://www.sciencedirect.com/science/article/abs/pii/S0959440X25000843 known_gaps: - Predicting intrinsically disordered protein regions and dynamics - De novo design of enzymes with novel catalytic functions not found in nature disputed_statements: [] secondary_sources: - title: "AlphaFold 3: Accurate Structure Prediction of Biomolecular Interactions (Google DeepMind)" type: journal_article year: 2024 authors: - Abramson, Josh - Adler, Jonas - Dunger, Jack - et al. institution: Google DeepMind / Nature url: https://www.nature.com/articles/s41586-024-07487-w - title: "AlphaFold 2: Improved Protein Structure Prediction Using Deep Learning (CASP14)" type: journal_article year: 2021 authors: - Jumper, John - Evans, Richard - Pritzel, Alexander - et al. institution: Google DeepMind / Nature url: https://www.nature.com/articles/s41586-021-03819-2 - title: "RoseTTAFold: Accurate Prediction of Protein Structures and Interactions (Baker Lab)" type: journal_article year: 2021 authors: - Baek, Minkyung - DiMaio, Frank - Anishchenko, Ivan - et al. institution: University of Washington / Science url: https://doi.org/10.1126/science.abj8754 - title: Highly Accurate Protein Structure Prediction with AlphaFold (Seminal — Jumper et al. / Nature 2021) type: journal_article year: 2021 authors: - Jumper, John - Evans, Richard - Pritzel, Alexander - et al. institution: Google DeepMind / Nature url: https://www.nature.com/articles/s41586-021-03819-2 updated: "2026-05-24" --- ## TL;DR Protein structure prediction has been transformed by AI — from AlphaFold2's historic 2020 breakthrough predicting static 3D structures to current frontiers: simulating protein dynamics, predicting multi-protein complexes, and designing entirely new proteins with specified functions. The 2024 Nobel Prize in Chemistry recognized this revolution, but fundamental challenges remain. ## Core Explanation The protein folding problem: given an amino acid sequence (1D string), predict the 3D folded structure. Proteins fold in milliseconds in cells; experimentally determining structures via X-ray crystallography, cryo-EM, or NMR takes months to years per protein. AI approach: (1) Multiple Sequence Alignment (MSA) — find evolutionarily related sequences, extract co-evolution signals (residue pairs that mutate together are likely near each other in 3D); (2) Pair representation — encode pairwise relationships between all residue pairs; (3) Structure module — iteratively refine 3D coordinates using invariant point attention (IPA), respecting roto-translational equivariance. AlphaFold2 (Jumper et al., Nature 2021) achieved median GDT_TS of 92.4 on CASP14 — considered a solution to the single-chain protein folding problem. ## Detailed Analysis Key models: (1) AlphaFold2 (DeepMind, 2020-2021) — MSA + Evoformer + Structure Module. Trained on PDB (180K structures). Open-sourced, with 200M+ predictions in the AlphaFold Database; (2) RoseTTAFold (Baker Lab, 2021) — three-track architecture processing 1D sequence, 2D distance, and 3D coordinates simultaneously; (3) ESMFold (Meta, 2022) — uses language model embeddings instead of MSAs, enabling 60x faster inference at slightly lower accuracy; (4) AlphaFold-Multimer (2022) — extends to protein-protein complexes; (5) AlphaFold3 (2024) — diffusion-based architecture predicting complexes including proteins, DNA, RNA, ligands, and ions. Nature 2026 commentary identifies the conformational landscape problem: proteins are dynamic, not static — they sample multiple conformations essential for function (enzyme open/closed states, transporter inward/outward-facing). Current AI predicts one static structure, missing the ensemble. Frontier 1: predicting full energy landscapes and transition pathways. Frontier 2: de novo protein design — RFdiffusion (Baker Lab, 2023) generates novel protein backbones via diffusion models, then ProteinMPNN designs sequences folding into those backbones. This enables designing proteins that don't exist in nature — enzymes for plastic degradation, carbon capture, or targeted therapeutics. ## Further Reading - AlphaFold Database: 200M+ predicted structures (EMBL-EBI) - RFdiffusion: Diffusion models for protein backbone generation - CASP: Critical Assessment of Structure Prediction (biennial)