RADiAnce: Retrieval-Augmented Diffusion for Protein Binders

• The paper's main contribution is introducing a retrieval-augmented latent diffusion framework that designs protein binders by integrating known interface exemplars.
• It employs a shared contrastive latent space and contrastive atomic VAE to align binding sites with binder interfaces, enabling cross-domain transfer.
• Quantitative experiments on peptides and antibodies demonstrate improved interaction overlap, binding affinity, and structural realism compared to baselines.

Searching arXiv for the RADiAnce paper and closely related retrieval-augmented diffusion work. Search query: arXiv (Zhang et al., 12 Oct 2025) RADiAnce protein binders retrieval augmented diffusion Retrieval-Augmented Diffusion for Aligned interface (RADiAnce) is a framework for protein binder design that leverages known interfaces to guide the generation of novel binders targeting specific sites (Zhang et al., 12 Oct 2025). It addresses the requirement that binder design recover not only plausible amino-acid sequences or stable backbones, but also realistic and functional interaction patterns at the interface, including hydrogen bonds, hydrophobic contacts, salt bridges, aromatic stacking, and correct residue orientation and packing (Zhang et al., 12 Oct 2025). The method unifies retrieval and generation in a shared contrastive latent space, retrieves relevant interface exemplars for a given binding site with simple dot-product similarity, and conditions a latent diffusion generator on those exemplars to enable cross-domain interface transfer across peptides, antibodies, and protein fragments (Zhang et al., 12 Oct 2025).

1. Problem formulation and design rationale

RADiAnce is motivated by a limitation of structure-based generative models that generate binders only conditioned on the target binding site. In the formulation reported for RADiAnce, this weakens rationality, interpretability, and generalization because the model does not explicitly leverage known interface examples already present in databases (Zhang et al., 12 Oct 2025). The central premise is therefore knowledge-guided generation: retrieve similar known interfaces and use them as prompts for the generator.

A molecular complex is represented as a pair of graphs, with binder molecule $x$ and binding site $y$ (Zhang et al., 12 Oct 2025). The binding site is defined as residues within a 10 Å cutoff of the binder, based on $C_\beta$ atoms (Zhang et al., 12 Oct 2025). In this setting, the target conditional distribution is expressed as

$p_\theta(x \mid y, v),$

where $v$ denotes retrieved interface exemplars (Zhang et al., 12 Oct 2025).

The framework is explicitly designed for cross-domain interface transfer. The reported domains are peptides, antibodies, and protein fragments, and the paper argues that transferable interaction logic can persist across these binder modalities (Zhang et al., 12 Oct 2025). This suggests that RADiAnce treats interfaces as reusable structural and physicochemical motifs rather than as domain-isolated templates.

2. Shared contrastive latent space and interface retrieval

The defining architectural idea of RADiAnce is a shared contrastive latent space in which binding sites are encoded as keys and binder interfaces are encoded as values (Zhang et al., 12 Oct 2025). Retrieval and generation therefore operate in the same representation space rather than being loosely coupled post hoc stages.

The paper constructs a key-value database

$\mathcal D = \{(k_k, v_k)\mid k=1,2,3,\dots\},$

where $k_k$ is a binding site graph embedding and $v_k$ is the corresponding binder interface embedding (Zhang et al., 12 Oct 2025). At inference, for a query binding site $y$, the model retrieves similar interface exemplars

$$v = \{v_k \mid (k_k, v_k)\in \mathcal D,\; k_k \text{ similar to } y\},$$

and generation proceeds under $y$0 (Zhang et al., 12 Oct 2025).

After training, the encoded database is written as

$y$1

and a query site $y$2 is encoded into a query key $y$3 (Zhang et al., 12 Oct 2025). Similarity is computed by inner product,

$y$4

followed by top-$y$5 retrieval,

$y$6

The paper also states that the ground-truth ligand of the query complex is excluded from retrieval to prevent trivial memorization (Zhang et al., 12 Oct 2025). This is important for distinguishing retrieval-guided generation from exemplar copying.

The retrieval mechanism is therefore not described as a generic nearest-neighbor heuristic. It depends on a latent space in which matching binding-site/interface pairs are explicitly aligned during training, so that retrieved values are intended to be semantically and structurally compatible with the query site (Zhang et al., 12 Oct 2025).

3. Contrastive atomic VAE and retrieval-space learning

RADiAnce uses a contrastive atomic VAE to learn compressed latent representations for both binding sites and binder interfaces (Zhang et al., 12 Oct 2025). The encoder maps binder and binding site independently into latent point clouds,

$y$7

with each latent variable containing a scalar embedding and a 3D coordinate embedding (Zhang et al., 12 Oct 2025). The decoder reconstructs the binder conditioned on both the binder latent and the site latent.

The reconstruction loss combines cross-entropy for block or residue type with MSE for time-dependent vector fields used in coordinate reconstruction: $y$8 The latent space is regularized by KL divergence on both scalar and coordinate latents: $y$9 where $C_\beta$0 is the center of mass of block $C_\beta$1 (Zhang et al., 12 Oct 2025).

To align site and interface latents for retrieval, RADiAnce applies a double-sided contrastive loss: $C_\beta$2 Here, $C_\beta$3 is the temperature, $C_\beta$4 is the batch size, and $C_\beta$5 denotes Euclidean inner product (Zhang et al., 12 Oct 2025). The full training objective is

$C_\beta$6

The reported interpretation is that joint training of reconstruction, latent regularization, and retrieval alignment is crucial, and retrieval performance drops if the VAE loss is removed (Zhang et al., 12 Oct 2025). A plausible implication is that the retrieval space is not merely discriminative; it is also constrained to remain generatively meaningful.

4. Retrieval-conditioned latent diffusion and conditioning strategy

Generation in RADiAnce is performed by a retrieval-conditioned latent diffusion generator operating in the same latent space as the retrieval module (Zhang et al., 12 Oct 2025). For each latent block state $C_\beta$7, the forward process adds Gaussian noise: $C_\beta$8 or equivalently,

$C_\beta$9

so that

$p_\theta(x \mid y, v),$0

The paper states that $p_\theta(x \mid y, v),$1 follows a cosine noise schedule (Zhang et al., 12 Oct 2025).

The reverse process learns

$p_\theta(x \mid y, v),$2

with noise-prediction parameterization

$p_\theta(x \mid y, v),$3

Training uses the denoising loss

$p_\theta(x \mid y, v),$4

Conditioning is implemented by an E(3)-equivariant transformer with cross-attention inserted between self-attention and feed-forward layers (Zhang et al., 12 Oct 2025). If $p_\theta(x \mid y, v),$5 are latent hidden states and $p_\theta(x \mid y, v),$6 are retrieved template embeddings, then

$p_\theta(x \mid y, v),$7

with learnable projections $p_\theta(x \mid y, v),$8, and the residual cross-attention update is

$p_\theta(x \mid y, v),$9

The paper also reports comparison with AdaLN-Zero and in-context fusion, with cross-attention performing best (Zhang et al., 12 Oct 2025).

The inference workflow is modular: encode the binding site $v$0 into query key $v$1, compute dot-product similarity against all database keys, retrieve top-$v$2 binder latents, sample latent noise $v$3, iteratively denoise conditioned on $v$4 and $v$5, and decode the final latent $v$6 into binder structure $v$7 (Zhang et al., 12 Oct 2025). Because the latent database can be replaced at inference, the paper presents this modularity as useful for different task domains (Zhang et al., 12 Oct 2025).

5. Evaluation setup, benchmarks, and reported performance

RADiAnce is evaluated on peptide design, antibody design, and cross-domain reference data (Zhang et al., 12 Oct 2025).

Domain / reference set	Dataset	Size / source
Peptide design	PepBench	4,157 training, 114 validation, 93 test cases from LNR
Antibody design	SAbDab	9,473 training, 400 validation, 60 test cases from RAbD
Cross-domain references	ProtFrag	70,498 monomer-derived protein fragments

The retrieval database is restricted to training data to prevent leakage (Zhang et al., 12 Oct 2025). Retrieval is evaluated with RC-$v$8, specifically RC-0.1%, RC-0.5%, and RC-5%, and with Interaction Type Overlap (ITO), which compares the interaction-pattern distribution of retrieved exemplars with the ground truth based on interaction types such as hydrogen bonds and hydrophobic contacts (Zhang et al., 12 Oct 2025). The retrieval baseline is random retrieval. The summary reported in the paper is that the contrastive VAE substantially improves retrieval over random, antibody RC-5% reaches 100%, peptide RC-5% reaches 67.74%, and ITO improves strongly with cross-domain data, especially with ProtFrag included (Zhang et al., 12 Oct 2025).

Generation is assessed with AAR, RMSD, $v$9, IMP, and ISM; the paper also reports DockQ, FoldX $\mathcal D = \{(k_k, v_k)\mid k=1,2,3,\dots\},$0, Binding Site Recovery, and diversity (Zhang et al., 12 Oct 2025). For each test complex, 100 samples are generated per model (Zhang et al., 12 Oct 2025). Reported baselines for peptides are RFDiffusion + ProteinMPNN + Rosetta, PepFlow, PepGLAD, and UniMoMo, while baselines for antibodies are MEAN, DyMEAN, DiffAb, GeoAB-R, GeoAB-D, and UniMoMo (Zhang et al., 12 Oct 2025).

The main reported quantitative findings are domain-specific. In peptide codesign, RADiAnce is described as best or near-best across metrics, with AAR 39.42, RMSD 2.29 Å, $\mathcal D = \{(k_k, v_k)\mid k=1,2,3,\dots\},$1 1.963 kJ/mol, IMP 41.94%, and ISM 52.15% (Zhang et al., 12 Oct 2025). In antibody CDR design, the method is reported to achieve higher AAR, lower RMSD, more favorable $\mathcal D = \{(k_k, v_k)\mid k=1,2,3,\dots\},$2, better IMP, and excellent ISM across CDRs, with example values HCDR3: AAR 54.66, RMSD 0.9443 Å, $\mathcal D = \{(k_k, v_k)\mid k=1,2,3,\dots\},$3 and LCDR3: AAR 76.75, RMSD 0.4478 Å, $\mathcal D = \{(k_k, v_k)\mid k=1,2,3,\dots\},$4 (Zhang et al., 12 Oct 2025).

A recurrent conclusion in the reported experiments is that retrieval alone is insufficient. The ablation study states that a baseline without retrieval has lower performance, a baseline with retrieval but without alignment yields limited gains and can be worse on raw Vina score, and the full model combining alignment and retrieval produces the largest improvements across metrics (Zhang et al., 12 Oct 2025). Although the paper title and primary task concern protein binders rather than small molecules, the same methodological principle is echoed in retrieval-enhanced aligned diffusion for receptor-based ligand generation in READ (Xu et al., 17 Jun 2025).

6. Cross-domain transfer, interpretability, and limitations

A major contribution attributed to RADiAnce is the demonstration that interfaces from one domain can improve generation in another (Zhang et al., 12 Oct 2025). The experiments explicitly study retrieval from antibodies, peptides, and protein fragments and report that cross-domain retrieval improves interaction overlap, binding affinity, geometrical plausibility, and reconstruction of contact patterns relative to single-domain-only retrieval (Zhang et al., 12 Oct 2025). The paper’s explanation is that many binding motifs are shared across binder modalities, including hydrogen-bond networks, local packing, and shape complementarity (Zhang et al., 12 Oct 2025). This suggests that the latent space captures transferable interface logic rather than only scaffold-specific signatures.

The analysis of retrieval quantity and quality is also specific. The paper reports that more retrieved samples generally help up to a point, too many retrieved templates can add noise, random or reverse retrieval hurts performance, no retrieval is clearly worse than retrieval-based settings, and adaptive retrieval with a similarity threshold works better than fixed top-$\mathcal D = \{(k_k, v_k)\mid k=1,2,3,\dots\},$5 in some cases (Zhang et al., 12 Oct 2025). These observations frame RADiAnce as a quality-sensitive retrieval-augmented generator rather than a monotonic “more exemplars is better” system.

Interpretability is illustrated through a case study on GPIIb/IIIa binder design, where retrieval-augmented generation reproduces key hydrogen-bond motifs involving arginine at the binding site and tyrosine in the binder (Zhang et al., 12 Oct 2025). Without retrieval, the model fails to reconstruct these interactions well; with retrieval, it inherits the critical motif and generates a more realistic interface (Zhang et al., 12 Oct 2025). The paper presents this as an interpretability advantage because generation can be related back to retrieved exemplars.

The stated limitations are centered on retrieval quality and benchmark construction. The paper notes that poor retrieval can inject noise, better structural descriptors may improve future systems, and more robust, structure-aware conditioning strategies are needed (Zhang et al., 12 Oct 2025). It also suggests stricter filtering in future benchmarks, including structural similarity constraints such as TM-score, to reduce leakage (Zhang et al., 12 Oct 2025). In practical terms, the method is presented as relevant to therapeutic antibodies, peptide inhibitors, protein-fragment binders, and de novo binder optimization, and the paper reports a de novo antibody design pipeline without a predefined bound scaffold that iterates design with AlphaFold3 and Rosetta (Zhang et al., 12 Oct 2025).

Taken together, RADiAnce is characterized by a specific methodological synthesis: a contrastive latent space aligning binding sites and interfaces, retrieval of site-relevant interface exemplars, and latent diffusion conditioned on those exemplars (Zhang et al., 12 Oct 2025). The paper’s broader implication is that protein binder design benefits from retrieving transferable interface knowledge and generating with it, rather than relying exclusively on site-conditioned generation from scratch (Zhang et al., 12 Oct 2025).