MODEL PROTEIN FOLDING · AI · STRUCTURE PREDICTION · under KEEPER ERIS

AlphaFold2 & CASP14 · Solved and Unsolved

At CASP14 (2020), AlphaFold2 reached a median domain GDT_TS of 92.4 and 0.96 Å backbone accuracy, against 2.8 Å for the next best method. It largely solved single-chain structure prediction, but not how proteins fold, how they move, or disorder.

Dave (t249u): "can you build some pages on protein folding, the math and the difficulty, edutaiton please, use this wasm engine"

LIT CASP14: median domain GDT_TS 92.4. Median backbone r.m.s.d.₉₅ was 0.96 Å for AlphaFold2 versus 2.8 Å for the next best method, and all-atom 1.5 Å versus 3.5 Å. Summed z-score 244.0 versus 90.8. GDT_TS > 70 on 87 of 92 domains (Jumper et al., Nature 596:583, 2021).

LIT The AlphaFold Protein Structure Database grew to 214,684,311 predicted structures in July 2022, covering almost all of UniProt. AlphaFold 3 (Abramson et al., Nature 630:493, 2024) extended prediction to complexes with nucleic acids and ligands. The 2024 Nobel Prize in Chemistry went to Hassabis and Jumper, and to David Baker for protein design.

LIT Scores: GDT_TS = 25 × (fraction of Cα within 1, 2, 4 and 8 Å, summed). TM-score = (1/L) Σ 1/(1 + (dᵢ/d₀)²) with d₀ = 1.24 ∛(L − 15) − 1.8. pLDDT bands: ≥ 90 very high (#0053D6), 70–90 confident, 50–70 low, < 50 very low.

WALL Not solved by AlphaFold2: the folding pathway and kinetics (Levinthal’s “how”), conformational ensembles and dynamics, intrinsically disordered regions (low pLDDT), and reliable effects of single mutations. It predicts a structure, it doesn’t simulate folding.

MODEL The lab compares a made-up helix–turn–helix with noisy “predictions” in the same frame, with no superposition step. Real GDT/TM searches over superpositions. The “pLDDT” bars here are just 100 − 10 × error, only to show the colour scale.

MODEL Two agents, as Dave set for every world. The ARTIST folds the 36-residue HP chain S4 PPPHHPPHHPPPPPHHHHHHHPPHHPPPPHHPPHPP on the cubic (3D) lattice with pivot, corner and kink moves. The CRITIC judges each move (Metropolis at temperature T) and scores the best fold as −E/18 (E_ref = -18, best found by our annealing; literature optimum −18). It also has Dave’s I-13 VM recount every kept fold from raw coordinates. Rendered with Three.js r160 from a pinned CDN, devicePixelRatio capped at 2. Drag to orbit, scroll to zoom.

LIT on load: the seal is recomputed, the page’s facts are re-checked as arithmetic (throws on mismatch), 200 seeded ARTIST rounds must improve, the workspace loop logs 5 entries, and the OBJ and SVG exports must agree. Then the engine boots, its SHA-256 is checked, and the page’s VM results are compared with independent JS.

⏸ Pause Reseed Export training + workspace log (.json) Export fold (.svg) Export geometry (.obj)

Gallery of kept folds (WebGL snapshots):

Shared board MODEL ARTIST ↔ CRITIC

Background art: AMBER MODEL a generative field seeded from this page’s seal.

proving…

Seal sha256(name|slug|blurb): 0fd953da9db9b544f5efcf3e7a9f4667c6b98ce33a863dcf02a37d9624ef21c1

↑ Protein Folding index · ← ERIS, keeper of Mayhem