Ion channels are critical therapeutic targets, yet identifying selective peptide modulators is challenging due to the structural complexity of venom toxins. We introduce a multimodal machine learning framework integrating primary protein sequences with 3D structural features, focusing on disulfide-rich scaffolds. We leverage Evolutionary Scale Modeling (ESM-2) to extract representations from 3,889 curated venom sequences. To capture structural realism, we incorporate AlphaFold 3D coordinates, representing peptides as molecular graphs. A key innovation is a graph attention mechanism weighted toward disulfide connectivity, reflecting the importance of cysteine knots in stabilizing toxin-channel interfaces. Preliminary results using ESM embeddings with a Random Forest classifier achieved 85% accuracy and a 0.82 F1-score in classifying Calcium, Sodium, and Potassium interactions. To improve out-of-distribution (OOD) generalization, we fuse transformer sequence embeddings with a specialized graph neural network (GNN). By explicitly modeling geometric constraints and disulfide dependencies, this framework advances AI-driven drug discovery for channelopathies and cancer.