Text-guided protein generation requires effective conditioning mechanisms and evaluationprotocols that account for overlap in both protein sequences and textual descriptions. This thesisintroduces Cross-Modal Leakage-Aware Diffusion for Proteins (CLAD-Pro), combining an MCM-conditioned latent diffusion model with a cluster- and text-aware data-splitting pipeline. TheMulti-modal Cross-attention Mechanism (MCM) integrates token-level SciBERT representationsinto the denoising of ESM-2 protein representations through learned cross-modality tokens. Thesplitting pipeline groups similar descriptions and targets balanced evaluation categories withdifferent levels of sequence similarity to the training set. We compare MCM with facilitator-basedconditioning on SwissProtCLAP using computational measures of plausibility, diversity, referencesimilarity, and similarity to training proteins. Across the three assigned sequence-similaritycategories, MCM achieves higher mean and median global BLOSUM62-positive fractions andTM-scores, together with higher overall coverage and BLOSUM62-positive fractions for selectedreference annotations. However, the facilitator achieves more favourable predicted structuralconfidence and sequence pseudo-perplexity. Neither model consistently matches reference-proteinplausibility, and neither shows a consistent diversity advantage. Although the splitting pipelineproduces approximately balanced assigned categories, a sequence audit identifies training matchesabove the intended identity bounds, limiting conclusions about generalization. Overall, MCMimproves selected reference-reconstruction measures under the evaluated configuration, but thesegains do not establish improved textual adherence or biological function. Controlled conditioningexperiments and experimental validation are needed to determine the functional relevance of theobserved differences ... Read more