PepGLAD

PepGLAD's contribution is the specific problem it solves in the sequence-structure co-design task. Most prior peptide design work is target-agnostic and produces 1D sequences, or generates a bare backbone and leaves side chains for a later stage. PepGLAD encodes full-atom residues of variable size into fixed-dimensional latents through a variational autoencoder, runs diffusion in that latent space, and decodes back to residue space. Because amino acid identities change during generation and therefore change the atom count, fixed-size diffusion is a genuine obstacle — this architecture is the paper's answer to it. The second design decision, a receptor-specific affine transformation that converts 3D coordinates into a shared standard space, addresses variable binding geometry. Binding sites differ in center offset and shape, which means their coordinate distributions have disparate expectations and covariance rather than resembling a standard normal. Mapping them into a common space is what lets the diffusion process transfer across pockets, and the paper reports gains in diversity and binding affinity as well as recovery of reference structures for binding conformation generation. On strengths, the code release is practical for a research group: pre-trained weights, training scripts for fine-tuning on custom targets, and inference scripts for new pockets, all with the diffusion hyperparameters exposed in code rather than hidden behind an interface. The paper also contributes a benchmark curated from the Protein Data Bank and literature with 1D sequences and 3D structures, which is useful if you want to compare generators on the same footing. On weaknesses, the honest framing is that this is an artifact you deploy and maintain. It assumes fluency in PyTorch, PyTorch Geometric, and diffusion models before you generate your first peptide. Validation is in silico; nothing in the release provides experimental confirmation of the peptides you produce, and there is no wet-lab service, SLA, or support contract. You bring the GPU compute and any integration into an existing screening or structure-prediction pipeline. Where it fits: academic and industrial computational biology groups with GPU clusters that are already extending or benchmarking geometric latent diffusion models, and teams who want to fine-tune a peptide diffusion model on proprietary targets. Where it does not: researchers without deep learning experience, anyone who needs a hosted API or point-and-click tool, and teams expecting synthesis-ready peptides directly out of the model. If your comparison set is backbone-first pipelines like RFdiffusion, the joint full-atom output is the differentiator worth evaluating.