Modern Gene Finders: ab initio gene discovery benchmark with DNA language models

Detecting genes in DNA sequence is a fundamental step in enabling virtually any downstream analysis of the genome. A complete annotation pipeline must address two complementary tasks. One task is transcript position discovery, which determines where transcripts begin and end. The other task is transcript segmentation, which reconstructs exon intron structure within those intervals. In this work, we focus on transcript bondary discovery and treat it as an independent benchmarking problem. We introduce a mammalian benchmark that evaluates strand-aware localization of transcript boundaries for complete mRNA and lncRNA genes, using biologically grounded metrics based on transcription start sites and transcript termination sites. In addition, we introduce our own approach, which uses a DNA language model \texttt{ModernGENA} and a multi-stage pipeline to infer stranded transcript intervals and to recover multiple transcript boundary isoforms for the same gene.