Protein Folding: The HP Lattice Model Explained

Anfinsen's dogma and Levinthal's paradox

In the early 1960s, Christian Anfinsen denatured ribonuclease A completely and watched it spontaneously refold into its active form once the denaturant was removed — establishing Anfinsen's dogma: the native structure is the thermodynamically stable state determined solely by the amino acid sequence, with no additional template required. But a puzzle remains: a 100-residue protein sampling even 3 conformations per residue has ~3¹⁰⁰ ≈ 5×10⁴⁷ possible structures — exploring them all at 10¹³ per second would outlast the universe. This is Levinthal's paradox (1969): folding cannot be a random search, so the sequence must bias the pathway toward the native state.

The HP model: two letters, one energy rule

Proposed by Ken Dill in 1985, the HP model reduces all 20 amino acids to two types — H (hydrophobic, repelled by water) and P (polar, comfortable in water) — placed on a self-avoiding path across a 2D or 3D square lattice. The energy of a conformation comes from one rule: every pair of H residues that are lattice-adjacent but not sequence-adjacent contributes −1; everything else contributes zero. The goal is the conformation with the most H–H contacts. Despite the extreme simplification, this reproduces real behaviour — sequences with clustered H residues fold compactly, mirroring real hydrophobic-core proteins, because burying non-polar groups releases ordered water molecules and increases solvent entropy, accounting for an estimated 60–70% of real folding stability.

E = −1 × (number of non-bonded H–H lattice contacts)

Self-avoiding walks on 2D square lattice (growth ~2.638^n):
  length 20: ~1.7×10^7        length 50: ~2.6×10^19

Energy funnels and NP-hardness

Modern folding theory (Wolynes, Bryngelson, Onuchic) describes an energy landscape: for a random heteropolymer it's rugged, full of kinetic traps, but real proteins — shaped by evolution — have a funnel-shaped landscape where compactness generally tracks with lower free energy, resolving Levinthal's paradox as a guided descent rather than a random search. Finding the exact ground-state HP conformation, though, was proven NP-hard in 1998 (Crescenzi, Goldman, Papadimitriou, Piccolboni, Yannakakis): self-avoiding walks on a 2D lattice grow roughly as 2.638ⁿ (≈10¹⁹ for length 50), so exact search is infeasible beyond short sequences, and heuristics — Monte Carlo with the Metropolis criterion, genetic algorithms, and reinforcement learning — are used instead.

From lattice models to AlphaFold

Extensions like HPNX (four residue types) and the Miyazawa-Jernigan 20×20 empirical contact matrix progressively close the gap between the HP abstraction and real protein physics. The trajectory from lattice models through fragment-assembly tools (Rosetta, I-TASSER) to AlphaFold2's equivariant transformer architecture, which achieved near-experimental accuracy in 2021, illustrates layered progress in computational biology. AlphaFold2 doesn't use a lattice, but the same physical intuition — hydrophobic burial as the dominant driving force — is implicitly encoded in its training data of solved structures, which is exactly what the HP model makes explicit and computable.

Frequently asked questions

Try it live

Everything above runs in your browser — open Protein Folding Visualiser and watch Monte Carlo pivot moves with simulated annealing drive hydrophobic residues to cluster and minimise energy on a 22×22 lattice. Nothing is installed, nothing is uploaded.