Acute Leukemia Classification Using Transcriptional Profiles From Low-Cost Nanopore mRNA Sequencing

Abstract

Purpose: Most cases of pediatric acute leukemia occur in low- and middle-income countries, where health centers lack the tools required for accurate diagnosis and disease classification. Recent research shows the robustness of using unbiased short-read RNA sequencing to classify genomic subtypes of acute leukemia. Compared with short-read sequencing, nanopore sequencing has low capital and consumable costs, making it suitable for use in locations with limited health infrastructure.

Materials and methods: We show the feasibility of nanopore mRNA sequencing on 134 cryopreserved acute leukemia specimens (26 acute myeloid leukemia [AML], 73 B-lineage acute lymphoblastic leukemia [B-ALL], 34 T-lineage acute lymphoblastic leukemia, and one acute undifferentiated leukemia). Using multiple library preparation approaches, we generated long-read transcripts for each sample. We developed a novel composite classification approach to predict acute leukemia lineage and major B-ALL and AML molecular subtypes directly from gene expression profiles.

Results: We demonstrate accurate classification of acute leukemia samples into AML, B-ALL, or T-lineage acute lymphoblastic leukemia (96.2% of cases are classifiable with a probability of > 0.8, with 100% accuracy) and further classification into clinically actionable genomic subtypes using shallow RNA nanopore sequencing, with 96.2% accuracy for major AML subtypes and 94.1% accuracy for major B-lineage acute lymphoblastic leukemia subtypes.

Conclusion: Transcriptional profiling of acute leukemia samples using nanopore technology for diagnostic classification is feasible and accurate, which has the potential to improve the accuracy of cancer diagnosis in low-resource settings.

FIG 1.

Process overview. (A) Sample preparation and sequencing: extract RNA from buffy coat, perform long-range RT-PCR to produce full-length cDNA, and sequence on MinION sequencer. (B) Informatics: raw sequence data are base-called on the local sequencing computer and then transferred to a remote cloud-computing platform where sequences are aligned to the reference transcriptome and a gene expression matrix is constructed and used as input to a machine learning lineage/genomic subtype classifier. (C) Classification: conceptual model of classification approach. We train a set of pairwise PLSDA classifiers, consisting of each type versus each other and each type versus everything else. The set of binary predictions produced by these models is then used to build a new feature matrix, which serves as the input to a SVM classifier for each of the lineage and AML genomic subtypes. The SVM classifiers produce prediction probabilities for each type. AML, acute myeloid leukemia; PCR, polymerase chain reaction; PLSDA, partial least-squares discriminant analysis; RT-PCR, reverse transcriptase polymerase chain reaction; SVM, support vector machine.

FIG 2.

Association of classification accuracy with the prediction probability. Each point represents a separate case, sorted by known lineage and subtype. The color of each point represents the predicted classification on the basis of ONT RNA sequencing. (A) Classification accuracy of leukemia lineage. (B) Classification accuracy of the AML core genomic subtype. (C) Classification accuracy of the B-ALL genomic subtype. For the leukemia lineage and AML genomic subtype classification, all cases with a prediction probability of > 80% were correctly classified. AML, acute myeloid leukemia; B-ALL, B-lineage acute lymphoblastic leukemia.

FIG 3.

Transcript versus read length as an indicator of RNA degradation and/or cDNA/PCR fidelity. There is significant variability in quality and proportion of full-length reads, to which classification is very robust. (A) 036, high quality with an average transcript completeness of 58.7%. (B) 043, typical with an average transcript completeness of 36.9%. (C) 048, degraded with an average transcript completeness of 25.5%. (D) 063, degraded with an average transcript completeness of 25.4%. PCR, polymerase chain reaction.