Transcript reconstruction from mammal RNA-Seq data remains a challenging problem due to several biases, such as those from sequencing or mapping, the complexity of mammalian transcriptome generation from alternative splicing, fragmentary characteristics of reads, and from the unbalanced sequencing. Here, IsoRef, a reference-based transcriptome assembler for RNA-Seq data, is proposed. IsoRef investigates information from not only sequencing data, but from transcript annotation as well, in order to build accurate splice graphs. A ?ow balancing technique is proposed to reduce the impact of false positive transcripts and to narrow the search space of true positive transcripts. For each of two in silico datasets, IsoRef predicted 1,400 additional correct transcripts than StringTie; for each of the fve actual datasets, IsoRef identifed at least 1,500 additional correct transcripts than StringTie, which improves the transcript-level and gene-level accuracy compared to StringTie with a maximum improvement of 20%. IsoRef is available at deepomics.org/module-instances/2CA682222F734424/.