N6-methyladenosine (m⁶A) is the most prevalent post-transcriptional modification in eukaryotic cells, playing a crucial role in regulating various biological processes. Dysregulation of m⁶A status is implicated in multiple human diseases, including cancer. Several prediction frameworks have been proposed for high-accuracy identification of putative m⁶A sites; however, none have targeted direct prediction of cancer-associated (or pro-cancer) m⁶A residues at the base-resolution level. Here, we report m6A-CAPred, a computational tool for predicting pro-cancer m⁶A sites learned from a comprehensive dataset of experimentally validated m⁶A sites. Our findings indicate that sequence information alone achieves limited performance. However, by leveraging domain-related knowledge (genome-derived features), m6A-CAPred successfully captures distinct domain characteristics between potentially pro-cancer m⁶A modifications and normal ones, with an average AUROC of 0.885 tested on an independent dataset. Leveraging the power of machine learning, we then performed transcriptome-wide prediction for large-scale screening of potentially pro-cancer m⁶A sites. Somatic variants derived from 33 types of TCGA cancer projects were extracted for additional validation, and the results showed that SNP density clearly differentiated the predicted pro-cancer and normal m⁶A sites.