CytoGPS: A large-scale karyotype analysis of CML data

Introduction

Precision medicine leverages an individual’s genotype to optimize their likelihood of achieving and maintaining health through prevention and targeted therapy. Karyotyping, the practice of visually examining and recording chromosomal abnormalities, is one of the earliest and most common genotyping techniques [1]. The utility of karyotyping for diagnosis and prognosis has been demonstrated in numerous genetic and neoplastic diseases. The results often have important clinical implications for patient care, and are regularly and widely used for clinical decision making [1]. Conventional karyotypic analysis is a component of the standard-of-care for virtually all hematologic malignancies [1]. Because cytogenetic data have been used clinically for decades, many institutions have databases that contain tens of thousands of karyotypes. For example, the National Cancer Institute (NCI) hosts the Mitelman Database of Chromosome Aberrations and Gene Fusions, which contains more than 69,000 karyotypes collected since 1971 [2]. These karyotypes represent a wealth of historic data that is waiting be unlocked for biomedical research.

Karyotypes are stored as text using a standard human-interpretable notation, the International System for Human Cytogenetic Nomenclature (ISCN). ISCN codes are complex, variable, and have not been translated into a machine-readable format able to identify clinically relevant patterns. Further, several versions of the ISCN standard have been introduced over the decades. Following the release of the earliest version of ISCN in 1971, the standard has been revised nine times, most recently in 2018 [3]. These revisions often contain significant changes from previous standards, compounding the difficulty of translation across and between cytogenetic data sets. Thus, to date little of the information embedded in karyotypic databases has been mined for research purposes.

Current use of karyotype data for research is limited to those patterns that are visually apparent to cytogeneticists. Although the ISCN is meant to be readable by humans, it is often difficult to interpret due to the volume, complexity, and variability of the information contained therein. To the knowledge of the authors, no commercial or open source software converts ISCN-based karyotypes into a computational model. The closest is CyDAS [4], which has two significant drawbacks: 1) it employs regular expressions to parse karyotypes, and therefore cannot support automatic diagnosis and recovery from syntactic failures, and 2) CyDAS cannot parse some special karyotypes, such as derivative chromosomes. Due to the lack/limitation of existing software, potential clinically relevant patterns in long and complex karyotypes remain hidden and thus, unused. To leverage this wealth of existing karyotypic data in its entirety, we developed a computational tool, the CytoGenetic Pattern Sleuth (CytoGPS), which is designed to translate raw karyotypes into a computable form [5]. This form represents chromosomal abnormalities as loss, gain, or fusion (LGF) events at the resolution of cytogenetic bands. Karyotypes transformed by CytoGPS can be analyzed in aggregate with a diverse array of approaches.

Here, we describe the development of the CytoGPS model and, as proof of principle, apply it to all 4,969 karyotypes in the Mitelman database with a diagnosis of chronic myelogenous leukemia (CML). CML is a cytogenetically well-understood disease, defined by presence of the Philadelphia translocation, t(9;22)(q34;q11.2) [6]. We expected the database to include some cases where this translocation would not appear, for two reasons. First, some of the data extend back to 1971, before the t(9;22)(q34;q11.2) became a diagnostic standard of CML, and might be classified differently using current criteria. Second, in up to 10% of CML cases, the t(9;22)(q34;q11.2) cannot be detected on routine karyotyping either because 3 or more chromosomes are involved in the translocation, or it is a so-called “cryptic” translocation, beyond the limits of resolution of routine karyotyping [7]. Regardless, we demonstrate that CytoGPS can find both known subtypes defined by the most important secondary cytogenetic abnormalities in CML as well as novel abnormalities.

Materials and Methods

Results

Discussion

Most of the specific abnormalities that define each of these clusters have been reported in the literature; thus providing verification and validation of CytoGPS as a new, viable methodology. Isochromosome 17 has been noted as relevant in CML cases particularly due to the resulting loss of the TP53 gene [15]. Trisomy 8 is another common cytogenetic event in CML; it is the most common secondary event after the Philadelphia chromosome [17]. This finding is supported by our analysis, both by the number of cases with trisomy 8 and the abundance of trisomy 8 as a defining characteristic in multiple clusters. Rarer events such as fusions associated with 3q (the defining characteristic of Cluster 1) have also been noted in the literature [18].

Many of these clusters can be “paired” based on the presence or absence of the t(9;22) translocation. Clusters 5 and 7 are both defined as having trisomy 8 yet Cluster 5 has the t(9;22) translocation while Cluster 7 does not. Both clusters have a comparable number of cases despite the ascertainment bias that CML is defined by the t(9;22) translocation and thus we would expect a much larger number of cases with the translocation than without. This mirrored cluster structure is due to two major reasons: 1) the age of the samples and 2) cryptic translocations. The first reason, the age of the samples, relates to the timing of CML’s disease definition. The Mitelman database goes back to 1971, pre-dating the disease definition of CML by the t(9;22) translocation. Many of the CML samples that were analyzed in the study thus pre-date this disease definition. The second reason that many of these samples do not appear to have the t(9;22) translocation relates to cryptic translocations. A cryptic translocation is defined as the event where a translocation has occurred, but it is too small to be detectable by conventional cytogenetics [19]. Such events would appear only from a more molecular assay such as Fluorescence in situ Hybridization (FISH) [20]. That is, many of these cases that do not appear to have the t(9;22) translocation probably have a cryptic translocation, thus, it went undetected and was not displayed in the ISCN karyotype.

Another interesting finding is the relationship of common cytogenetic abnormalities to each other. As seen in Figure 2, the majority of gain events split from the majority of loss events at the top of the dendrogram. This indicates that when a patient has a gain event they are more likely to have additional gains rather than losses. This phenomenon has been observed in other hematologic malignancies such as chronic lymphocytic leukemia (CLL) [21]. This further strengthens the argument that CytoGPS is able to uncover the underlying biology contained in karyotype data. Further studies will determine the clinical relevance of these findings.

The methods described here can be applied to any collection of ISCN karyotypes. We expect leukemias and lymphomas, which predominate in many institutional cytogenetics databases, to be the first diseases to the studied in detail. The free version of CytoGPS can be accessed via the web browser CytoGPS.org [5]. CytoGPS exports the binary vectors of each clone-karyotype both as Excel spreadsheets and as JSON files, which can be read into a wide variety of statistical software. As shown in this paper, these can be used for unsupervised clustering analyses, but both the binary vectors and the cluster assignments can also be used in supervised analyses to find markers or predictors of clinically important variable such as response to therapy.