Identifying therapeutic targets in drug discovery (DD) remains a complex challenge, particularly in understanding protein-protein interactions (PPIs) involved in disease pathways. While Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) frameworks offer promise, integrating them into effective workflows for target identification (Target ID) is nontrivial. We conducted a user study with domain experts to evaluate how LLMs and RAGs can support Target ID. Findings revealed two critical needs: generating multiple therapeutically relevant PPIs from an initial protein and providing contextual explanations for each interaction. Existing approaches suffer from semantic ambiguity, limited explainability, and insufficient retrieval granularity. To address these gaps, we propose GraPPI, a scalable knowledge graph-based RAG framework that decomposes PPI pathway analysis into edge-level sub-tasks. This retrieve-divide-solve strategy enhances interpretability and supports large-scale exploration of signaling pathways, enabling more effective hypothesis generation in therapeutic research.