NanoAAI: An interpretable sequence–structure framework for nanobody–antigen binding prediction

Functional antibody fragments are increasingly used in diagnostics and therapeutic engineering, but experimental identification of antigen-specific binders remains costly and difficult to scale. Existing protein--protein and antibody--antigen interaction models are predominantly sequence-based and often provide limited antibody-specific, residue-level interpretation. Here we present NanoAAI, a sequence--structure framework for predicting binding between antibody fragments and antigens. NanoAAI combines protein language-model embeddings with residue-level molecular graphs, aligns sequence and structural representations through intra- and intermolecular contrastive learning, and incorporates complementarity-determining-region (CDR) priors through CDR-aware cross-attention. We evaluated NanoAAI on three binding tasks spanning different antibody-fragment formats and found that it achieved the strongest overall performance among the evaluated methods. Interpretability analyses showed that NanoAAI attention signals were enriched in CDR regions, particularly CDR3. By further incorporating AlphaFold2-predicted structures, NanoAAI enabled antigen-wise retrieval of nanobody sequences without experimentally resolved structures. Overall, NanoAAI provides an accurate, interpretable, and structure-aware computational framework for antibody-fragment--antigen binding prediction and candidate nanobody prioritization.