Submission 1 Overview: The FASTA file contains 4 sequences, all generated using BindCraft with the following configuration:

BindCraft Settings:

Default BindCraft filters and advanced settings were applied No specific interaction sites were specified Sequence lengths ranged from 10 to 250 residues predict_initial_guess = True (generates initial structure predictions to guide design) fix_mpnn_interface = False (allows ProteinMPNN to redesign the interface residues)

Submission 2 Overview: The FASTA file contains 2 sequences that were iteratively optimized based on initial high-scoring submissions.

Design Pipeline:

Warm-Start BindCraft Generation: Starting sequences were generated using a customized BindCraft implementation that employs a warm-start initialization strategy rather than standard random initialization:

Warm-Start Modification: Initializes sequence logits biased toward a known binding sequence (z₀[i, a] = β·δ(a = a_known[i]) + ε). Large-Scale Screening: 666 successful BindCraft trajectories were generated and evaluated using Boltz2 to predict ipSAE scores.

Machine Learning-Based Selection: A sequence classifier was developed to predict sequences likely to achieve ipSAE > 0.8 using ESM-C embeddings.

Rational Mutagenesis: Mutations were introduced in regions predicted to be less critical for binding (residues distant from the binding interface based on 3D structure analysis).

Final Selection: Mutated sequences were scored by the ESM-C classifier, and the highest-confidence predictions were submitted.