SUMMARY: In response to NOT-GM-24-021, we apply for a supplement to our ongoing NIGMS award 5R01GM148970- 02, “DMS/NIGMS 1: Topological Study on Histological Images and Spatial Transcriptomics”. We propose to extend the parent award by integrating the developed methods with state-of-the-art deep learning methodology, in order to produce more powerful topology- and gene-informed feature representation for spatial transcriptomics data. The method will use topological features and gene features as guidance for contrastive learning and teacher-student consistency learning so that histology image patches are mapped into a powerful topology- and gene-informed features. To achieve the goal, we will develop contrastive learning or teacher-student consistency loss, in which we ensure that both gene-based and topology- based similarity measure will be preserved. We will use public ST data (10 Visium) on breast cancer and brain cancer, as well as proprietary ST data (4 triple negative breast cancer samples, from Indiana University). We will use gene correlation, as well as topological feature similarity, to determine similarities between different spots, and learn the feature mapping. To evaluate the quality of the obtained imaging feature, we will test on two downstream tasks: (1) direct image-to-gene mapping; (2) downstream tasks on whole slide prediction. For the first task, we will use the learnt feature to train a classifier to directly predict spot-level gene expression. We expect that the learnt new features have better prediction power of gene expression. Furthermore, for downstream tasks, we will combine the learnt gene features with whole slide prediction models such as Multiple Instance Learning (MIL). We expect that the new topology- and gene-informed features will achieve better prediction power in these downstream tasks. To perform the proposed tasks, however, we require a powerful GPU server. We expect such features will demonstrate strong prediction power in downstream tasks.