An Isoform-Centric, Structure-Aware Framework for Protein Function Prediction and Evaluation, Instantiated in 3DisoDeepPF

Understanding and accurately predicting function across protein isoforms has been a long-standing challenge with profound implications for both biological and translational research. However, due to the scarcity of high-quality isoform-resolved annotations, most computational methods for protein function prediction have been developed and benchmarked using a single reference sequence per gene. To address this gap, we present an isoform-centric framework for protein family (Pfam) domain and Gene Ontology (GO) term prediction. We implemented this framework in 3DisoDeepPF, a deep multi-label learning model that integrates a dense graph of sequence and structure similarity with multimodal representations, and applied it to a breast cancer isoform-specific atlas. 3DisoDeepPF improves GO-term and Pfam-domain prediction over all baselines and state-of-the art models in both conventional and isoform-resolved settings. It also remains robust under homology-controlled isoform-level evaluation and resolves directional Pfam remodeling among isoforms from the same gene. An evidence-tracing module links predicted labels to associated proteins, as illustrated by a CIB1 breast cancer isoform case study. Together, 3DisoDeepPF provides a framework for resolving disease-relevant isoform function, and can support hypothesis generation and future prioritization of cancer-associated isoform.