RADiAnce

RADiAnce is a generative framework for designing protein binders, molecules that bind a target interface, that combines retrieval with diffusion. Designing a binder means proposing an interface geometry and sequence that complements a given binding site, a task studied separately for peptides, antibodies, and protein fragments. RADiAnce unifies these binder types and, crucially, supplements pure generation with a retrieval step that reuses knowledge from previously seen interfaces.

The method learns a shared contrastive latent space in which binding sites and binder interfaces are aligned, so that for a new target it can retrieve relevant interface embeddings and use them to guide generation. A retrieval-conditioned latent diffusion model then integrates the retrieved interfaces through cross-attention and residual conditioning to produce a binder. Because the latent space is shared across binder types, RADiAnce transfers across domains: interfaces learned from one class of binders inform the design of another.

RADiAnce was introduced in October 2025 by Zishen Zhang, Xiangzhe Kong, Wenbing Huang, and Yang Liu, a collaboration between the Department of Computer Science and Technology and the Institute for AI Industry Research (AIR) at Tsinghua University and the Gaoling School of Artificial Intelligence at Renmin University of China. It was presented at NeurIPS 2025 and is described in an accompanying arXiv preprint.

#Key Features

• Retrieval-augmented generation: RADiAnce retrieves relevant interfaces for a target and conditions generation on them, injecting prior structural knowledge rather than generating from scratch.
• Shared contrastive latent space: A contrastive variational autoencoder encodes binding sites and binder interfaces into a common latent space aligned by contrastive learning, providing a unified similarity metric across binder types.
• Retrieval-conditioned latent diffusion: A latent diffusion model integrates retrieved interface embeddings through cross-attention and residual conditioning to guide binder generation.
• Cross-domain transfer: The unified representation lets the model transfer across peptides, antibodies, and protein fragments, with multi-domain training improving performance over single-domain baselines.
• Strong recovery metrics: The framework reports competitive amino-acid recovery, geometry, and binding-energy metrics across peptide and antibody design benchmarks.

#Technical Details

RADiAnce has two components. A contrastive variational autoencoder independently encodes binding sites and binder interfaces at all-atom resolution, supervised by reconstruction loss, KL regularization, and a contrastive objective that pulls matching site-interface pairs together while pushing non-matching pairs apart. A retrieval-conditioned latent diffusion model then generates in that same latent space, incorporating retrieved interface embeddings via cross-attention and residual conditioning. Training draws on three cross-domain datasets: PepBench (4,157 training and 114 validation peptide complexes, with 93 LNR test cases), SAbDab (9,473 training and 400 validation antibody entries, with 60 RAbD test cases), and ProtFrag (70,498 protein-fragment entries). On peptide design, RADiAnce reaches 39.42% amino-acid recovery with 2.29 Angstrom RMSD; on antibody CDR-H1 design it reaches 90.83% amino-acid recovery with 0.30 Angstrom RMSD and favorable predicted binding-energy changes, and cross-domain training measurably improves retrieval and design quality over single-domain training.

#Applications

RADiAnce is aimed at protein engineers and antibody designers who need candidate binders for a specified target interface and want to exploit the growing archive of solved complexes. By retrieving and reusing interface patterns rather than designing purely de novo, it can propose peptide, antibody-CDR, or protein fragment binders within one framework, which is useful when a target resembles interfaces seen in other binder classes. This makes it applicable to early-stage binder discovery across therapeutic and research settings.

#Impact

RADiAnce brings retrieval-augmented generation, a technique widely used in language modeling, to protein binder design, showing that explicitly retrieving and conditioning on known interfaces improves affinity and geometry over generation alone. Its shared latent space across peptides, antibodies, and protein fragments demonstrates cross-domain transfer, suggesting a route to binder generators that learn from all available interface data rather than siloed per-domain models. As a NeurIPS 2025 preprint, its results are computational and await experimental validation; the paper reports code availability in an appendix rather than through a widely mirrored public repository.