CASPTracker

Also: the Pencil Room, what structure prediction cannot yet do, each with a source that says it is still open.

Also: BlindTest, a small blind test every week: a new experiment against a prediction that was public first.

Thirty years of the Critical Assessment of Structure Prediction (CASP) — the biennial blind experiment that has measured protein-structure prediction since 1994. Each card below is one CASP experiment: its organiser, the widely-reported headline result, the group(s) publicly credited with topping the primary assessment, and a link to a public source. Nothing is installed; nothing about you is collected.

This is the tool — an honest headline-results record, not a mirror of the per-target GDT_TS / lDDT / TM-score / RMSD tables (those live on predictioncenter.org) — see the coverage gauge below. We report what the organiser and assessors reported; we show the metric, not a synthesised “winner”; and where a specific GDT_TS figure is not a clear matter of public record, we leave it blank rather than invent it. Editorial is CC BY 4.0; each experiment links a public source.

1. CASP1 1994 Founding (1994–1998) GDT_TS not itemised

   First CASP experiment, organised by John Moult and colleagues; results presented at Asilomar, California (December 1994).

   The first blind, community-wide assessment of protein structure prediction. Predictions were made for target sequences whose experimental structures were not yet public, establishing the double-blind format that still defines CASP. Comparative (homology) modelling showed early promise; ab initio prediction remained largely unsolved.

   Topped the primary assessmentNo single overall winner was crowned; CASP1 established the assessment framework rather than a leaderboard.

   Sources

   • Moult et al., "A large-scale experiment to assess protein structure prediction methods", Proteins (1995) · Proteins: Structure, Function, and Bioinformatics
   • Protein Structure Prediction Center — CASP history · Prediction Center
2. CASP2 1996 Founding (1994–1998) GDT_TS not itemised

   Second CASP experiment; results meeting at Asilomar, California (December 1996). Organised by the Protein Structure Prediction Center.

   CASP2 consolidated the assessment categories (comparative modelling, fold recognition / threading, and ab initio). Threading methods drew particular attention for recognising folds with low sequence identity.

   Topped the primary assessmentAssessed by category by independent assessors; no single overall champion. Reported in the CASP2 special issue of Proteins.

   Sources

   • Moult et al., "Critical assessment of methods of protein structure prediction (CASP): Round II", Proteins Suppl. 1 (1997) · Proteins
   • Protein Structure Prediction Center — CASP2 · Prediction Center
3. CASP3 1998 Founding (1994–1998) GDT_TS not itemised

   Third CASP experiment; results meeting at Asilomar, California (December 1998).

   CASP3 saw the introduction of fragment-assembly approaches to ab initio prediction; David Baker's Rosetta fragment-insertion method emerged as a leading approach for free-modelling targets.

   Topped the primary assessmentRosetta (Baker lab) was widely noted for its performance on ab initio / new-fold targets; category assessment reported in the Proteins CASP3 supplement.

   Sources

   • Moult et al., "Critical assessment of methods of protein structure prediction (CASP): Round III", Proteins Suppl. 3 (1999) · Proteins
   • Protein Structure Prediction Center — CASP3 · Prediction Center
4. CASP4 2000 Founding (1994–1998) GDT_TS not itemised

   Fourth CASP experiment; results meeting at Asilomar, California (December 2000).

   CASP4 is widely regarded as the point where fragment-assembly ab initio methods, led by Rosetta, produced genuinely useful models for some new-fold targets. The GDT_TS score became a central assessment metric around this period.

   Topped the primary assessmentRosetta (Baker lab) was a standout on the ab initio / new-fold category; overall assessment reported in the Proteins CASP4 supplement.

   Sources

   • Moult et al., "Critical assessment of methods of protein structure prediction (CASP): Round IV", Proteins Suppl. 5 (2001) · Proteins
   • Protein Structure Prediction Center — CASP4 · Prediction Center
5. CASP5 2002 Template era (2000–2010) GDT_TS not itemised

   Fifth CASP experiment; results meeting at Asilomar, California (December 2002).

   CASP5 continued the rise of fragment-assembly methods and saw increasing use of meta-servers that combine multiple prediction servers. Comparative modelling remained the most reliable category.

   Topped the primary assessmentAssessed by category; Baker-lab Rosetta and leading server groups featured prominently. Reported in the Proteins CASP5 supplement.

   Sources

   • Moult et al., "Critical assessment of methods of protein structure prediction (CASP): Round V", Proteins Suppl. 6 (2003) · Proteins
   • Protein Structure Prediction Center — CASP5 · Prediction Center
6. CASP6 2004 Template era (2000–2010) GDT_TS not itemised

   Sixth CASP experiment; results meeting at Gaeta, Italy (December 2004).

   CASP6 emphasised template-based modelling accuracy and model refinement. Automated servers narrowed the gap with human expert groups on many template-based targets.

   Topped the primary assessmentAssessed by category; no single overall champion. Reported in the Proteins CASP6 supplement.

   Sources

   • Moult et al., "Critical assessment of methods of protein structure prediction — Round 6", Proteins Suppl. 7 (2005) · Proteins
   • Protein Structure Prediction Center — CASP6 · Prediction Center
7. CASP7 2006 Template era (2000–2010) GDT_TS not itemised

   Seventh CASP experiment; results meeting at Asilomar, California (November 2006).

   CASP7 introduced a dedicated model-quality-assessment (QA) category and continued the trend of high-accuracy template-based modelling. Server + human hybrid pipelines dominated the reliable categories.

   Topped the primary assessmentAssessed by category; the Baker (Rosetta) and Zhang groups were among the strongest performers. Reported in the Proteins CASP7 supplement.

   Sources

   • Moult et al., "Critical assessment of methods of protein structure prediction — Round VII", Proteins Suppl. 8 (2007) · Proteins
   • Protein Structure Prediction Center — CASP7 · Prediction Center
8. CASP8 2008 Template era (2000–2010) GDT_TS not itemised

   Eighth CASP experiment; results meeting in Cagliari, Sardinia, Italy (December 2008).

   CASP8 confirmed the strength of consensus / meta-server approaches. The Zhang group's I-TASSER pipeline was widely recognised as a leading automated method for tertiary-structure prediction.

   Topped the primary assessmentZhang lab (I-TASSER) was among the top-performing groups for automated tertiary-structure prediction. Reported in the Proteins CASP8 supplement.

   Sources

   • Moult et al., "Critical assessment of methods of protein structure prediction — Round VIII", Proteins Suppl. 9 (2009) · Proteins
   • Protein Structure Prediction Center — CASP8 · Prediction Center
9. CASP9 2010 Template era (2000–2010) GDT_TS not itemised

   Ninth CASP experiment; results meeting in Asilomar, California (December 2010).

   CASP9 continued incremental gains in template-based modelling and refinement. Contact-assisted and consensus methods matured, but ab initio prediction of large novel folds remained the central unsolved problem.

   Topped the primary assessmentAssessed by category; leading automated groups (including Zhang / I-TASSER and Baker / Rosetta lineages) featured. Reported in the Proteins CASP9 supplement.

   Sources

   • Moult et al., "Critical assessment of methods of protein structure prediction (CASP) — round IX", Proteins Suppl. (2011) · Proteins
   • Protein Structure Prediction Center — CASP9 · Prediction Center
10. CASP10 2012 Co-evolution era (2010–2016) GDT_TS not itemised

    Tenth CASP experiment; results meeting in Gaeta, Italy (December 2012).

    CASP10 marked the emergence of co-evolution / residue–residue contact prediction from multiple sequence alignments as a route to fold prediction for targets lacking templates — the methodological seed of the later deep-learning surge.

    Topped the primary assessmentAssessed by category; the Baker and Zhang groups again featured among the strongest. Reported in the Proteins CASP10 supplement.

    Sources

    • Moult et al., "Critical assessment of methods of protein structure prediction (CASP) — round X", Proteins Suppl. (2014) · Proteins
    • Protein Structure Prediction Center — CASP10 · Prediction Center
11. CASP11 2014 Co-evolution era (2010–2016) GDT_TS not itemised

    Eleventh CASP experiment; results meeting in Riviera Maya, Mexico (December 2014).

    CASP11 saw contact-prediction-assisted modelling improve free-modelling results, with co-evolution methods increasingly integrated into fold prediction pipelines.

    Topped the primary assessmentAssessed by category; leading server and human groups performed strongly on template-based targets. Reported in the Proteins CASP11 supplement.

    Sources

    • Moult et al., "Critical assessment of methods of protein structure prediction (CASP) — round XI", Proteins Suppl. (2016) · Proteins
    • Protein Structure Prediction Center — CASP11 · Prediction Center
12. CASP12 2016 Co-evolution era (2010–2016) GDT_TS not itemised

    Twelfth CASP experiment; results meeting in Gaeta, Italy (December 2016).

    CASP12 is widely cited as the point at which deep-learning-based contact prediction (e.g. RaptorX-Contact using residual neural networks) delivered a clear, measurable jump in free-modelling accuracy, foreshadowing the CASP13 deep-learning breakthrough.

    Topped the primary assessmentAssessed by category; deep-learning contact-prediction groups drew particular notice on hard targets. Reported in the Proteins CASP12 supplement.

    Sources

    • Moult et al., "Critical assessment of methods of protein structure prediction (CASP) — Round XII", Proteins Suppl. (2018) · Proteins
    • Protein Structure Prediction Center — CASP12 · Prediction Center
13. CASP13 2018 Deep-learning breakthrough (2018–2020) GDT_TS not itemised

    Thirteenth CASP experiment; results meeting in Riviera Maya, Cancún, Mexico (December 2018).

    The first AlphaFold (DeepMind), entered as group "A7D", topped the tertiary-structure (free-modelling) rankings — the first time a deep-learning system led CASP. AlphaFold 1 predicted inter-residue distance distributions with deep neural networks and used them to guide folding, producing a clear step-change on the hardest free-modelling targets.

    Topped the primary assessmentAlphaFold (DeepMind, group A7D) placed first overall on tertiary-structure prediction; the Zhang group and other established labs led among non-DeepMind entrants.

    Sources

    • AlphaFold: Using AI for scientific discovery (CASP13 result) · Google DeepMind
    • Senior et al., "Improved protein structure prediction using potentials from deep learning" (AlphaFold, CASP13), Nature (2020) · Nature
    • Protein Structure Prediction Center — CASP13 · Prediction Center
14. CASP14 2020 Deep-learning breakthrough (2018–2020) GDT_TS ≈ 92.4

    Fourteenth CASP experiment; held virtually (December 2020) owing to the COVID-19 pandemic; organised by the Protein Structure Prediction Center.

    AlphaFold 2 (DeepMind, group "AF2") achieved a median backbone accuracy on the order of a GDT_TS of about 92.4 across targets — a level of accuracy competitive with experiment for many targets — and CASP organisers described protein structure prediction for single domains as, in large part, a solved problem. This is the result widely reported as the field's breakthrough moment.

    Topped the primary assessmentAlphaFold 2 (DeepMind) topped the tertiary-structure rankings by a wide margin; the Baker-lab and other groups led among non-DeepMind entrants.

    Sources

    • "AlphaFold: a solution to a 50-year-old grand challenge in biology" (CASP14, median GDT ~92.4) · Google DeepMind
    • Jumper et al., "Highly accurate protein structure prediction with AlphaFold", Nature (2021) · Nature
    • Protein Structure Prediction Center — CASP14 · Prediction Center
15. CASP15 2022 Post-AlphaFold (2022– ) GDT_TS not itemised

    Fifteenth CASP experiment; results meeting in Antalya, Türkiye (December 2022).

    With AlphaFold 2 open-source and widely available, CASP15 became a contest of AlphaFold-derived and enhanced pipelines rather than of a single system: top groups improved on plain AF2 by MSA engineering, sampling, and ensembling. CASP15 also ran a substantially expanded RNA / nucleic-acid category and emphasised protein assemblies (multimers).

    Topped the primary assessmentFor single-chain tertiary structure, AF2-based and AF2-enhanced groups (e.g. the Yang-Zhang / Zheng lineage and MSA-engineering groups) led; RoseTTAFold-family and specialised RNA-modelling groups featured in the newer categories. No single system swept every category.

    Sources

    • Protein Structure Prediction Center — CASP15 · Prediction Center
    • Kryshtafovych et al., "Critical assessment of methods of protein structure prediction (CASP) — Round XV", Proteins (2023) · Proteins
    • Das et al., CASP15 RNA-puzzles / nucleic-acid assessment coverage · Nature Methods
16. CASP16 2024 Post-AlphaFold (2022– ) GDT_TS not itemised

    Sixteenth CASP experiment; targets collected summer 2024; results conference in Punta Cana, Dominican Republic (December 1–4, 2024).

    CASP16 reaffirmed that single-domain protein folding is largely solved and shifted the frontier to complexes, ligands, and nucleic acids. AlphaFold 3, RoseTTAFold-All-Atom, and the Boltz / Chai families were newly available to all. Notably, in the protein-complex and antibody–antigen categories, a group layering traditional docking and extensive sampling on top of AF (the Kozakov–Vajda team) topped the field, showing hand-engineered pipelines still lead on the hardest complex cases.

    Topped the primary assessmentProtein complexes / multimers and antibody–antigen: Kozakov (Stony Brook) + Vajda (Boston University) team ("kozakovvajda"), the latter without using AF3/AFM directly on the antibody–antigen set. Oligomers: Cheng lab and Kihara lab. Stoichiometry: PreStoi (used by MULTICOM). Single-domain folding remained near-ceiling for AF-based methods.

    Sources

    • Success at CASP16 in predicting protein multimers (Kozakov–Vajda team) · Vajda Lab, Boston University
    • Assessment of protein complex predictions in CASP16 · PMC (NCBI)
    • Protein Structure Prediction Center — CASP16 · Prediction Center

No experiments match these filters. Clear filters