The　use　of　multiple　drugs,　termed　Polypharmacy,　is　a　common　prescription　for　treatment.　However,　Polypharmacy　is　likely　to　cause　unknown　side　effects,　which　can　seriously　affect　patients’　health.　Given　the　complexity　of　the　Drug-Drug　Interactions　(DDIs),　recent　approaches　mainly　leverage　knowledge　graphs　to　predict　DDIs,　formulated　as　a　multi-relational　link　prediction　problem.　The　accuracy　of　such　approaches　relies　on　the　comprehensibility　of　the　established　knowledge　graphs.　In　this　paper,　we　explore　the　factors　of　DDIs　in-depth　and　propose　an　effective　DDIs　prediction　method　based　on　reasonable　domain　knowledge.　Specifically,　we　established　a　comprehensive　knowledge　graph,　which　takes　the　interrelationships　among　drugs,　genes,　and　enzymes　into　account,　on　top　of　a　baseline　knowledge　graph.　We　then　train　and　apply　a　Convolutional　Neural　Network　(CNN)　to　predict　the　type　and　probability　of　DDIs　among　drug　pairs.　We　verify　the　impact　of　different　factors　on　performance.　Under　the　same　data　conditions　as　the　baseline,　the　results　of　which　show　that　the　accuracy　of　our　proposal　is　12%　higher　than　the　baseline　approach.　©　2020,　Springer　Nature　Switzerland　AG.