Scientists in the laboratory of Rendong Yang, PhD, associate professor of Urology, have developed a new large language model that can interpret transcriptomic data in cancer cell lines more accurately than conventional approaches, as detailed in a recent study published in Nature Communications. 

Long-read RNA sequencing technologies have transformed transcriptomics research by detecting complex RNA splicing and gene fusion events that have often been missed by conventional short-read RNA-sequencing methods.

Among these technologies includes nanopore direct RNA sequencing (dRNA-seq), which can sequence full-length RNA molecules directly and produce more accurate analyses of RNA biology. However, previous work suggests this approach may generate chimera artifacts — in which multiple RNA sequences incorrectly join to form a single RNA sequence— and limit the reliability and utility of the data.

“If we cannot distinguish this artifact versus the true gene fusion event, that will severely affect your downstream quantification of gene expression values or the gene isoforms. Sometimes you may miss a particular gene isoform due to this chimeric artifact, which can affect how you determine the gene fusion drivers for a particular disease,” Yang said.

To address this challenge, Yang’s team developed a new genomic large language model, called DeepChopper, to facilitate the detection and removal of chimera artifacts in dRNAseq data.

“Leveraging recent advances in large language model that can interpret complex genetic patterns, DeepChopper processes long genomic contexts with single-nucleotide resolution. This capability enables precise identification of adapter sequences within base-called long reads, facilitating the detection and removal of chimera artifacts in dRNAseq data,” the authors wrote.

Detection of chimeric read artifacts in dRNA-seq data using DeepChopper
and validation with orthogonal sequencing platforms

(a) Overview of the architecture of the DeepChopper model. (b) Length distribution of predicted adapters by DeepChopper in VCaP dRNA-seq data. (c) Distribution of relative adapter position along read length in VCaP dRNA-seq data. Grey rectangle represents a long read from 5′ to 3′. Relative position is calculated as the ratio of the length before DeepChopper predicted adapter start position to the total read length. (d) Distribution of segments per read after trimming: 1 segment indicates 3′ end adapter trimmed, while 2 or more indicate internal adapters trimmed. (e) Chimeric alignments (in thousands) for VCaP dRNA-seq reads processed by Dorado with and without adapter trimming, and DeepChopper. DeepChopper greatly reduces chimeric alignments not supported by direct cDNA sequencing. (f) Distribution of base qualities from identified internal adapters by DeepChopper. Background colors indicate quality levels: green (high), yellow (medium), and red (low). (g) BLAST-like alignment tool (BLAT) identity distribution of the internal adapter sequences mapping against human reference genome. (h) The number of chimeric alignments (in thousands) for A549, HCT116, HepG2, K562, and MCF7 cell lines processed by Dorado with adapter trimming and DeepChopper. DeepChopper consistently reduces chimeric alignments not supported by cDNA sequencing across all cell lines. (i) Chimeric alignments from dRNA-seq of the WTC11 cell line, evaluated for support using additional ONT and Pacific Biosciences (PacBio) sequencing data with different protocols. DeepChopper reduces unsupported chimeric alignments across all methods compared to Dorado with adapter trimming.

To validate their tool, the investigators used DeepChopper to sequence the transcriptome of prostate cancer cells line, which further revealed the prevalence of chimera artifacts.

“We demonstrated that these artifacts significantly impact transcriptomic analysis by complicating gene fusion detection, transcript annotation, and alternative splicing studies,” the authors wrote.

The approach, Yang said, could be used to identify false-positive RNA biological events in cancer samples.

“That could be important to understand how underlying genomic instabilities occurred biologically rather than due to a technical artifact,” Yang said. “We can think about ways to utilize AI and large language models to address that rather than using a traditional rule-based approach.”

Availability – DeepChopper (v1.2.6), implemented in Rust and Python, is open source and available on GitHub (https://github.com/ylab-hi/DeepChopper)

Source – Northwestern University