Transcriptomic profiling of bulk and single-cell data typically involves aligning to or quantifying a reference gene annotation, such as RefSeq or Gencode, to produce a count matrix for downstream analyses. However, these static references do not capture the full diversity and complexity of the human transcriptome, which harbours a plethora of cell-type specific long non-coding RNAs and alternatively spliced isoforms. Long read sequencing is facilitating the identification and full-length characterization of these transcriptional products, unveiling numerous novel transcripts in any given sample. We hypothesize that ad hoc transcriptomic references--generated from and only from the RNA sequences at hand--will increase the granularity of single-cell analyses by improving the precision of gene expression profiling. We first performed a comprehensive benchmarking of different transcriptome assembly algorithms on spike-in RNAs to assess their relative performances. We then tested this strategy on single-cell cDNA data from a pediatric acute myeloid leukemia bone marrow aspiration, sequenced in parallel with Illumina and Nanopore. A de novo long read assembly was compared to the ‘gold-standard’ static reference strategy to reveal over 7000 new transcripts in this sample. The ad hoc reference has a noticeable impact on the generation of clusters using dimensionality reduction tools, revealing a possible new group of leukemic cells. Furthermore, 112 novel intergenic lncRNAs were identified to be differentially expressed and specific to the leukemic cell clusters, revealing novel biomarkers for the monitoring of minimal residual disease as well as new targets for precision therapeutics and subsequent functional investigations. We propose that this strategy has great potential for molecular medicine, albeit further benchmarking is required.