We received the fastq files of whole-exome sequencing results of tumor and matched normal sample data of 10 AML patients from the Korea Genome Organization in December 2012. There were no patient-related medical or characteristic data. We aligned the sequencing reads to the human reference genome (hg 19, GRCh37) from USCC by BWA 0.6.2 [12] (Figs. 1 and 2). To filter the known single nucleotide polymorphisms (SNPs), we used dbSNP bulid 137. We removed PCR duplicates and filtering low-quality SNPs by Samtools 0.1.18 [13], Picard 1.68 [14], and GATK 2.3.4 [15]. After the filtering process, the SAM file was converted to VCF file by VCF Tools 0.1.10 [16]. For detecting somatic mutation candidates, we obtained the difference in VCF files between tumor and normal samples. For annotation of these somatic mutation candidates, we used the ANNOVAR tool [17].

We used formal concept analysis (FCA) for the construction of hierarchical relationships among samples sharing highly mutated genes [18]. FCA is a useful method in conceptual clustering of objects, attributes, and their binary relationship. In FCA, the sets of formal objects and formal attributes together with their relation to each other form a "formal context," which can be represented by a crosstable. In our case, the objects are 10 AML samples, and the attributes are 45 highly mutated genes. We defined the formal context as K = (G, M, I), where G is a set of objects (i.e., samples), M is a set of attributes (i.e., mutated genes), and I ⊆ G × M is the incidence relations where (g, m) ⊆ I if object g has attribute m. For A ⊆ G and B ⊆ M, we define the operators A' = {m ∈ M|gIm for all g ∈ A} (i.e., the set of attributes common to the objects in A) and B' = {g ∈ G|gIm for all m ∈ B} (i.e., the set of objects common to the attributes in B). A pair of (A, B) is a formal concept of k(G, M, I) if and only if A ⊆ G, B ⊆ M, A' = B, and A = B'. A is called the extent and B is the intent of the concept (A, B). The extent consists of all objects belonging to the concept while the intent contains all attributes shared by the objects. The concept of a given context is naturally ordered by the subconcept-superconcept relation, defined by (A1, B1) ≤ (A2, B2): <=> A1 ⊆ A2 (<=> B2 ⊆ B1).

The ordered set of all concepts of the context (G, M, I) is denoted by C(G, M, I) and is called the concept lattice of (G, M, I). We represent the structure of this concept lattice with a Hasse diagram, in which nodes are the concepts and the edges correspond to the neighborhood relationship among the concepts. All concepts above an object label (below the attribute label) include that object (attribute). The top element of a lattice is a unit concept, representing a concept that contains all objects. The bottom element is a zero concept having no object.

We have identified 12,908 somatic mutation candidates in 10 AML sequenced exomes, including 1,281 point mutations, 625 insertion/deletions (Indels) (Table 1, Fig. 3). The point mutations include 7,483 synonymous single nucleotide variations (SNVs), 4,297 nonsynonymous SNVs, 282 stopgain SNVs, 14 stoploss SNVs, and 5 frameshift substitutions, and the Indels include 247 frame shift insertions, 310 frameshift deletions, and 68 nonframe shifts (Fig. 4). For each patient, the average nonsynonymous mutation count was 429.7 (SD, 97.16).

About 342 to 665 genes have nonsynonymous somatic mutation candidates at least once in each AML sample (Table 2). Recurrent mutated genes were observed in all samples.

The most frequently mutated genes across all samples were USP9Y and MUC5B; these genes were mutated in 5 samples. These genes were also highly mutated in each sample; for USP9Y genes, it had 6 nonsynonymous mutations in sample 3. We have selected 45 highly mutated genes (1.5%) from 2981 mutated genes. We defined highly mutated genes as genes having 3 or more nonsynonymous mutations in each sample (Table 3). In a comparison of mutations with the COSMIC database [19], among 45 highly mutated genes, 21 genes matched to hematopoietic and lymphoid tissue malignancy terms, and 21 genes matched to other cancer types. In 3 genes, there was no matched term in COSMIC (Table 4).

We used the concept lattice to construct the hierarchical relationship between the samples that had 45 highly mutated genes. Concept Biolattice analysis is a mathematical framework based on concept lattice analysis for better biological interpretation of genomic data. The top element of a lattice is a unit concept, representing a concept that contains all objects. The bottom element is a zero concept having no object [20, 21]. For comparing with the Concept lattice (Fig. 5), we also performed hierarchical clustering analysis by Ward method. In hierarchical clustering, cluster 1 has 5 samples (nos. 1, 2, 5, 9, and 10), cluster 2 has 2 samples (nos. 4 and 7), and others have 1 sample each (Fig. 6). We divided the samples into 4 subgroups by interpretation of the concept lattice result (Fig. 7). Lattice subgroup 1 shared SYNE1 gene mutation, and samples 3, 4, and 7 were included in this subgroup. Subgroup 2 was comprised of 5 samples (nos. 1, 2, 5, 6, and 9) that had MUC5 gene mutations in common. Samples 10 and 8 could be isolated by the uniqueness of their mutated gene sharing pattern.