Example 05: Protein Folding Complex

4-chain protein complex binding affinity prediction. Structured canvas with matched-field cross-chain connectivity vs a flat baseline — the type hierarchy (chain → interface → complex) with array_element="matched_fields" learns binding affinity and stability prediction.

Source: examples/05_protein_folding_complex.py

Results

4×4 panel figure: Chain conformations, binding site contact maps, affinity predictions, stability analysis, secondary structure, topology diagrams, and training curves.

Animation: 4-chain protein complex docking simulation — free chains approaching, initial docking, conformational tightening, and stable complex formation with charge interactions, hydrogen bonds, and binding energy readout.

Type hierarchy

@dataclass
class Chain:
    sequence: Field = Field(4, 4)
    structure: Field = Field(2, 2)
    binding_site: Field = Field(2, 4, loss_weight=3.0,
                                semantic_type="protein binding interface")
    contacts: Field = Field(2, 4, loss_weight=2.0)

@dataclass
class Complex:
    interaction: Field = Field(2, 4, loss_weight=4.0)
    affinity: Field = Field(1, 1, loss_weight=5.0)
    stability: Field = Field(1, 2, loss_weight=3.0)
    chains: list  # 4 chains: heavy, light, antigen, cofactor

Flat baseline for comparison:

@dataclass
class FlatComplex:
    all_residues: Field = Field(8, 8)
    structure: Field = Field(4, 2)
    affinity: Field = Field(1, 1, loss_weight=5.0)

Connectivity

# Structured — matched-field cross-chain attention
bound_structured = compile_schema(
    Complex(chains=[Chain(), Chain(), Chain(), Chain()]),
    T=1, H=16, W=16, d_model=48,
    connectivity=ConnectivityPolicy(
        intra="dense",
        parent_child="hub_spoke",
        array_element="matched_fields",
    ),
    layout_strategy=LayoutStrategy.INTERLEAVED,
)
# 19 fields, 256 positions, 217 connections

# Flat baseline
bound_flat = compile_schema(
    FlatComplex(), T=1, H=16, W=16, d_model=48,
    connectivity=ConnectivityPolicy(intra="dense"),
)
# 3 fields, 256 positions, 9 connections

Key insight

array_element="matched_fields" connects binding site fields across chains — receptor chain binding sites attend to ligand chain binding sites of the same semantic type. The interaction region acts as a learned aggregation of cross-chain information.

Run it

python examples/05_protein_folding_complex.py
# Generates: assets/examples/05_protein.{png,gif,mp4}