Geo2vec (geographic representation of transcript as vectors) explored different strategies for encoding sub-molecular geographic information of ribonucleotides. Three novel encoding methods, i.e., landmarkTX, gridTX, and chunkTX, as well as the widely used one-hot method are currently supported. LandmarkTX is a lightweight encoding scheme directly capturing the position of the target ribonucleotide (or site) relative to transcript landmarks, i.e., the distances to the two edges of the exon, coding sequence (CDS), and transcript, respectively. Meanwhile, gridTX and chunkTX are designed to describe the landscape of the entire transcript through grids (of equal widths) or regions (with unequal width), respectivel. By incorporating additional geographic information to classical sequence-based approach, we here provide new high-accuracy N6-methyladenosine (m6A) predictors GepSe and I-GepSe. GepSe (Geography plus Sequences) considers the longest transcript isoform for each site, while I-GepSe makes full use of all mapped isoforms and enables Isoform-aware deep learning for m6A.