Spatial transcriptomics (ST) maps gene expression within intact tissues, but its cost and technical complexity limit broad adoption. A growing body of deep-learning methods now seeks to infer ST directly from routine histology images, especially H&E-stained slides, with the aim of turning archived pathology into virtual molecular data. This review examines more than 40 ST prediction models, comparing their data requirements, modelling strategies, evaluation practices and translational potential. Recent advances in model architecture, histology foundation models and higher resolution in situ ST platforms have improved prediction performance and expanded the range of potential applications. However, most current methods still show limited robustness and generalisability, with performance constrained primarily by the quantity, quality and diversity of available ST training data. Comparisons between models also remain difficult due to inconsistent pre-processing pipelines, training datasets and evaluation strategies. The field is increasingly recognising that progress depends not only on more sophisticated models, but also on standardised benchmarks, improved data harmonisation and clearer evaluation frameworks. Despite these limitations, ST prediction is already showing utility in research settings, including biomarker discovery, tissue domain inference, molecular super-resolution and large-scale analysis of archived histology cohorts. Emerging applications also include patient stratification, virtual molecular profiling and multi-modal pathology systems that integrate histology, transcriptomics and language models. As datasets continue to expand and models become more reliable, ST prediction has the potential to become an important component of next-generation digital pathology workflows.