Multi-omics drug response prediction can shorten screening, reduce cost and provide more effective personalised cancer treatment. The activity of pathways, jointly shaped by mutations, copy-number changes, and methylation, directly influences drug sensitivity and resistance. Two limits hinder multi-omics fusion. First, at the structural level, features are non-grid and pathways show long-range dependencies, so standard CNNs struggle to model global context. Second, at the functional level, naive concatenation loses key biological meaning and fails to align functions. We propose PiM-X (Pathway-image Mamba with cross attention). To address functional alignment and non-grid inputs, we project multi-omics into pathway images. For each pathway and omics we perform sample-wise PCA, then fix pathway order by correlation, forming a compact grid that preserves local pathway neighborhoods. To address long-range dependencies, we apply a Mamba state-space encoder for efficient sequence modeling in linear time. A cross-attention module fuses global context from Mamba with local patterns from pathway images, capturing both distant relations and local structure. Against strong CNN/Transformer/SSM baselines, PiM-X raises AUC by 1.2% on GDSC and 2.3% on CCLE, with pathway-level results aligning with known biology.