GPDL

Motif scaffolding — building a new protein backbone that holds a functional motif in exactly its native geometry — is the workhorse task behind designed binders, enzymes and antibody mimetics. The leading approaches fine-tune structure predictors pretrained on multiple sequence alignments, most prominently RFdiffusion and RFjoint. They are accurate, but their scaffolds cluster tightly: an MSA-pretrained model is conditioned on co-evolutionary signal that de novo proteins, by construction, lack. Methods trained from scratch sample more broadly and pay for it in accuracy.

GPDL — Generative Protein Design by Language model — from Hai-Feng Chen's group at Shanghai Jiao Tong University, replaces the MSA-pretrained prior with a protein language model one. It orchestrates frozen components rather than training a new generator: a structure seeding module adapted from ESMFold proposes a prompt structure around the specified motif, ESM-IF1 reads that structure back into candidate sequences, and an optimization module refines the sequence by Markov chain Monte Carlo under a simulated-annealing schedule, scored against ESMFold. The two modules run independently or in sequence, trading speed against quality.

The work appeared as a preprint in October 2023, was revised in December 2024 under the GPDL name, and was published in the International Journal of Biological Macromolecules in 2025.

#Key Features

• Language-model prior instead of MSA pretraining: The seeding module keeps ESMFold's folding trunk and structure module with the esm2_t36_3B_UR50D language model, so the design prior comes from sequence space rather than co-evolutionary couplings that de novo backbones lack.
• Distance-matrix conditioning: Motif geometry enters as 25 pairwise distances per residue pair — the four backbone atoms plus a virtual Cβ — each expanded into a 16-dimensional Gaussian radial basis over 2 to 22 Å and added into the pair representation through a single-layer adapter, making the conditioning invariant to rigid-body transformation of the input motif.
• Two complementary modules: GPDL-Inpainting runs from a released checkpoint for fast seeding; GPDL-Hallucination optimizes the sequence sample against the frozen folding oracle with a user-set iteration count.
• Diversity as the headline metric: Designs are scored by unique designable clusters, combining designability and diversity in one number, and by multidimensional scaling over pairwise backbone TM-scores.
• Zero-install entry point: A Colab notebook runs the full pipeline without any local environment, which matters because the reference installation needs two conda environments.

#Technical Details

The seeding module is fine-tuned on an inpainting objective: given the motif sequence with alanine padding and the motif distance matrix, recover the full structure. Its checkpoint (inpaint_weight_11.pt) ships in the repository. Seed structures are encoded by ESM-IF1, and the optimization module mutates the sequence under an MCMC schedule with losses on motif RMSD, self-consistency TM-score, predicted aligned error and pLDDT; ESMFold's weights are never updated, only the sequence sample. ProteinMPNN, vendored as a fork in the repository, redesigns sequences for the resulting backbones.

On the 24-case RFdiffusion motif-scaffolding benchmark, GPDL solves 22, the highest count among RFdiffusion, RFdesign, Chroma and TDS, and generates 33.5% more unique designable clusters than RFdiffusion. Evaluation follows the RFdiffusion protocol with a TM-score cutoff of 0.5 and ESMFold as the filter; results hold with OmegaFold or single-sequence AlphaFold 2. On a 30-protein orphan benchmark split by MMseqs2 at 30% maximum sequence identity to training, global TM-score, PAE, motif RMSD and pLDDT are comparable across groups, with orphans achieving the higher success rate. For the PD-L1 binding interface on PD-1 (PDB 5IUS), where RFdiffusion produced a single successful design, GPDL produced four at 5,000 iterations. Metal-binding designs for a di-iron protein (1BCF) and a double EF-hand (1PRW) were redesigned with ProteinMPNN, folded with AlphaFold 3 including the ions, and held their coordination geometry through 200 ns of all-atom molecular dynamics.

#Applications

GPDL targets small functional proteins built around a known active site or binding epitope: enzyme active sites, binders against a defined interface, and metal-coordination sites. Its diversity advantage matters most when a single motif needs many structurally distinct candidates — for experimental screening, or when the first fold class fails to express. The tunable iteration count lets a user spend more compute only on the motifs that need it.

#Impact

GPDL is evidence that converting a structure predictor into a designer, without training a diffusion or flow-matching generator, remains competitive on the standard motif-scaffolding benchmark while covering more of structure space. The MIT-licensed repository, released checkpoint and Colab notebook make it straightforwardly reusable. The stated limitations are material: all validation is in silico, with no wet-lab characterization; very short motifs such as the three-residue catalytic triad of ketosteroid isomerase (1QJG) remain difficult for GPDL and the methods it is compared against; and controllable generation — steering toward specified secondary structure or a bound small molecule — is not supported, with a Van der Waals ligand penalty tested and found not to improve accuracy.