The performance of large language models (LLMs) depends critically on the quality of their instruction-tuning data. For protein LLMs, existing instruction-tuning datasets are typically derived from structured protein annotations through template-based or label-guided generation. Although these datasets capture important biological facts, their construction paradigms often lead to two limitations: instruction homogenization and reasoning superficiality. The former refers to questions concentrating around repetitive forms and narrow topics; the latter to answers relying on shallow annotation lookup rather than evidence-rich biological reasoning. To address these limitations, we propose ProtSynGen, a knowledge graph-guided open-ended data synthesis framework for protein LLMs. ProtSynGen integrates structured protein databases with unstructured scientific literature into a protein-centric knowledge graph, and uses hierarchical graph sampling to guide LLMs in generating open-ended QAs across atomic, aggregated, and complex reasoning scenarios. Experiments show that ProtSynGen produces instruction-tuning data with broader topic coverage, richer reasoning depth, and stronger quality metrics than existing protein LLM instruction datasets. Models trained on ProtSynGen data further achieve superior performance on open-ended protein QA benchmarks, consistently outperforming baseline methods across atomic, aggregated, and complex reasoning scenarios.