Each sphere is a residue in a toy 26-residue chain, colored by chemical class: pink = hydrophobic (H), cyan = polar (P) — the same two-letter alphabet used in classic lattice protein models. An AI-style optimizer nudges every residue's 3D position downhill on an energy surface each iteration, the same core idea behind iterative structure-refinement networks like AlphaFold and behind force-field relaxation in AI materials search tools like GNoME.
E = Σ_bonds ½k(|r_i+1 − r_i| − ℓ0)²
+ Σ_i,j [ k_rep / d_ij − εH·Hi·Hj·max(0, 1 − d_ij/rH) ]
update: r_i ← r_i − η·∇_i E + √(2T)·ξ , T = (1 − confidence)
- AI confidence sets the search temperature T. Low confidence = noisy, exploratory steps (like an uncertain prediction); high confidence = clean descent straight to a compact fold.
- Fold speed sets how many optimizer iterations run per second.
- Fold / Unfold start the descent or reset the chain to an extended, unfolded state.
- New Sequence reshuffles which residues are hydrophobic vs polar, changing the entire energy landscape — a new "protein" to fold.
- Autonomous Screening mimics a self-driving lab: it repeatedly folds a fresh random sequence, records its converged energy, and keeps the best (lowest-energy) candidate found so far — a simplified version of how AI-guided autonomous labs screen thousands of molecules or materials for a target property.
Hydrophobic residues pulling inward while polar ones stay exposed is a real driving force in actual protein folding; here it is deliberately simplified so the descent, and the AI-search analogy, are easy to watch happen in real time.