An algorithmic approach for breakage-fusion-bridge detection in tumor genomes

Abstract

Breakage-fusion-bridge (BFB) is a mechanism of genomic instability characterized by the joining and subsequent tearing apart of sister chromatids. When this process is repeated during multiple rounds of cell division, it leads to patterns of copy number increases of chromosomal segments as well as fold-back inversions where duplicated segments are arranged head-to-head. These structural variations can then drive tumorigenesis. BFB can be observed in progress using cytogenetic techniques, but generally BFB must be inferred from data such as microarrays or sequencing collected after BFB has ceased. Making correct inferences from this data is not straightforward, particularly given the complexity of some cancer genomes and BFB’s ability to generate a wide range of rearrangement patterns. Here we present algorithms to aid the interpretation of evidence for BFB. We first pose the BFB count-vector problem: given a chromosome segmentation and segment copy numbers, decide whether BFB can yield a chromosome with the given segment counts. We present a linear time algorithm for the problem, in contrast to a previous exponential time algorithm. We then combine this algorithm with fold-back inversions to develop tests for BFB. We show that, contingent on assumptions about cancer genome evolution, count vectors and fold-back inversions are sufficient evidence for detecting BFB. We apply the presented techniques to paired-end sequencing data from pancreatic tumors and confirm a previous finding of BFB as well as identify a chromosomal region likely rearranged by BFB cycles, demonstrating the practicality of our approach.

Genomic instability allows cells to acquire the functional capabilities needed to become cancerous (1), so understanding the origin and operation of genomic instability is crucial to finding effective treatments for cancer. Numerous mechanisms of genomic instability have been proposed (2), including the faulty repair of double-stranded DNA breaks by recombination or end-joining and polymerase hopping caused by replication fork collapse (3). These mechanisms are generally not directly observable, so their elucidation requires the deciphering of often subtle clues after genomic instability has ceased.

In contrast, the breakage-fusion-bridge (BFB) mechanism creates gross chromosomal abnormalities that can be seen in progress using methods that have been available for decades (4). BFB begins when a chromosome loses a telomere (Fig. 1 A and B). Then, during replication, the two sister chromatids of the telomere-lacking chromosome fuse together (Fig. 1 C and D). During anaphase, as the centromeres of the chromosome migrate to opposite ends of the cell (Fig. 1E), the fused chromatids are torn apart (Fig. 1F). Each daughter cell receives a chromosome missing a telomere, and the cycle can begin again. As this process repeats, it can lead to the rapid accumulation of amplifications and rearrangements that facilitate the transition to malignancy (5).

Fig. 1.

Schematic BFB process.

This process produces several identifiable cytogenetic signatures such as anaphase bridges and dicentric chromosomes. However, as cancer genomics has shifted to high-throughput techniques, the signatures of BFB have become less clear. Methods such as microarrays and sequencing do not allow for direct observation of BFB, and similar to other mechanisms, BFB must be inferred from its footprint in complex data.

Multiple groups have begun to address the problem of finding evidence for BFB in high-throughput data. For example, Bignell et al. found a pattern of inversions and exponentially increasing copy numbers “[bearing] all the architectural hallmarks at the sequence level” of BFB (6). Kitada and Yamasaki found a pattern of copy counts and segment organization consistent with a particular set of BFB cycles (7). Hillmer et al. used paired-end sequencing to find patterns of inversions and amplification explainable by BFB (8).

The procedures of these investigators, among others (9–11), share an element in common: They determine whether a particular observation is consistent with or could be explained by BFB. Although this is helpful, it does not on its own allow one to infer whether or not BFB occurred. Indeed, in a previous work (12) we examined short patterns of copy number increases consisting of five or six chromosome segments. We found that many such patterns, whether produced by BFB or not, were consistent or nearly consistent with BFB. Thus, finding that such a pattern was consistent with BFB would only be weak evidence that it had been produced by BFB. This finding highlights the need for a rigorous and systematic approach to the interpretation of modern data for BFB to avoid being misled by the complexity of cancer genomes and the BFB mechanism itself.

Here we present a framework for interpreting high-throughput data for signatures of BFB. We incorporate observations of breakpoints as well as copy numbers to create a scoring scheme for chromosomes. Through simulations, we find appropriate threshold scores for labeling a chromosome as having undergone BFB based on varying models of cancer genome evolution and tolerances for error. This framework complements the work of previous groups by not only finding breakpoint and copy number patterns consistent with BFB but also showing under what assumptions they are more likely to be observed if BFB occurred than if it did not.

The key technical contribution that underlies our scoring scheme is a fast algorithm for determining if a given pattern of copy counts is consistent with BFB. This algorithm is related to a previously described algorithm (12) in that it takes advantage of a distinctive feature of BFB: When fused chromatids are torn apart, they may not tear at the site of fusion. This yields chromosomes with either a terminal deletion or a terminal inverted duplication. When a chromosome undergoes this process repeatedly, it results in particular patterns of copy number increases. The running time of the earlier algorithm grew exponentially with the amount of amplification and the number of segments in a copy number pattern. This greatly narrowed the scope of copy number patterns that could be investigated. This was particularly limiting as it appeared that copy number patterns with more segments would be more useful for identifying BFB, but these patterns could not be evaluated in a reasonable amount of time with the previous method. The algorithm presented here is linear time and therefore allows complex copy number patterns to be checked in a trivial amount of time.

We begin by describing the kinds of high-throughput data that can provide evidence for BFB. We then lay out some formalization needed to precisely describe scoring methods of samples based on BFB evidence implied from such data. Next, we define related computational problems, followed by an outline of algorithms for these problems. In Results, we detail the simulations used to measure the performance of our scoring system. Based on simulation parameters, we find false and true positive rates for different BFB signatures. We apply our methods to two datasets. The first is copy number data from 746 cancer cell lines (13). We find three chromosomes that have long copy number patterns consistent with BFB, but the false positive rates from our simulations suggest that these may be false discoveries. We also examine paired-end sequencing data from pancreatic cancers. We find two chromosomes that likely have undergone BFB, one that was also identified in ref. 14 and one additional finding.

High-Throughput Evidence for BFB

We consider two experimental sources for evidence for BFB: microarrays and sequencing. Microarrays allow for the estimation of the copy number of segments of a chromosome by measuring probe intensities (15). Sequencing also yields copy number estimates by measuring depth of sequence coverage (16). In addition, sequencing can reveal genomic breakpoints where different portions of the genome are unexpectedly adjacent. This is generally the extent of evidence available from either technique. Sequencing does not allow for a full reconstruction of a rearranged chromosome, as the repetitive nature of the genome leads to multiple alternative assemblies. Neither method can resolve segment copy numbers by orientation, so copy numbers from both forward and reversed chromosome segments are summed. Nevertheless, BFB should leave its signature both in breakpoints and copy counts, and we examine each in turn.

Breakpoints.

During BFB, the telomere-lacking sister chromatids are fused together. This causes the ends of the sister chromatids to become adjacent but in opposite orientations (Fig. 1D). This adjacency is unlikely to be disrupted by subsequent BFB cycles and will remain in the final sequence as two duplicated segments arranged head-to-head. If the chromosome is paired-end sequenced, the rearrangement will appear as two ends that map very near each other but in opposite orientations. This type of rearrangement has been termed a “fold-back inversion” (14), and regions of a chromosome rearranged by BFB should have an enrichment of these fold-back inversions. Reliable indications for fold-back inversions may or may not be available, depending on the type of experiment and its intensity.

Copy Counts.

Each BFB cycle duplicates some telomeric portion of the chromosome undergoing BFB. These repeated duplications lead to certain characteristic copy number patterns. We would like to evaluate copy numbers observed from microarrays or sequencing and determine if the copy numbers contain the footprint of BFB. Previous groups have searched for such a footprint by manually inspecting copy number data and searching for a set of BFB cycles that could produce the observed copy numbers (6, 7). This approach is challenging and labor intensive, but developing a more general approach turns out to be rather difficult. A key technical contribution of this paper is the development of efficient algorithms to evaluate copy counts for consistency with BFB.

Formalizing BFB.

Creating an efficient method for evaluating copy numbers requires some formalization, so we begin with some definitions and basic results.

We represent a chromosome as a string ABC…, where each letter corresponds to a contiguous segment of the chromosome. For example, the string ABCD would symbolize a chromosome arm composed of four segments, where A is the segment nearest the centromere. More generally, we use σ_l for the lth segment in a chromosome. So, ABCD could be written σ₁σ₂σ₃σ₄. A bar notation is used to signify that a segment is reversed. Greek letters α, β, γ, ρ denote concatenations of chromosomal segments, and a bar will again mean that the concatenation is reversed. For example, if , . An empty string is denoted by ɛ.

Consider the following BFB cycle on a chromosome X • ABCD, where “•” represents the centromere, X is one chromosomal arm, and ABCD is the four-segmented other chromosomal arm which has lost a telomere. The cycle starts with the duplication of the chromosome into two sister chromatids and their fusion at the ends of the “D” segments, generating the dicentric chromosome . During anaphase, the two centromeres migrate to opposite poles of the cell and a breakage of the dicentric chromosome occurs between the centromeres, say between and , providing one daughter cell with the chromosome , and another daughter cell with the chromosome X•ABC (chromosomes and X•ABC are equivalent). The now-amplified segment D in the first daughter cell may confer some proliferative advantage, causing its descendants to increase in frequency. The daughter cells also lack a telomere on one chromosome arm and may undergo additional BFB cycles. One possible subsequent cycle could, for example, cause an inverted duplication of the suffix , yielding the chromosome . As these BFB cycles continue, the count of segments on the modified chromosome can increase significantly.

The notation will be used to indicate that the string β can be obtained by applying 0 or more BFB cycles over the string α, as formally described in Definition 1.

Definition 1:

For two strings α, β, say that if β = α, or α = ργ for some strings ρ,γ such that γ ≠ ɛ, and .

We say that β is an l-BFB string if for some consecutive chromosomal region α = σ_lσ_l+1… starting at the lth segment σ_l, . Say that β is a BFB string if it is an l-BFB string for some l. As examples, CDE = σ₃σ₄σ₅ is a 3-BFB string, and so are CDE and CDEE. The empty string ɛ is considered an l-BFB string for every integer l > 0.

Denote by the count vector of α, where α represents a chromosomal arm σ₁σ₂…σ_k with k segments, and n_l is the copy number of σ_l and in α. For example, for , . Say that a vector is a BFB count vector if there exists some 1-BFB string α such that .

Handling Experimental Imprecision.

Experimental methods do not provide the accurate copy number of a given chromosome segment. Instead, some measurement error is expected. Moreover, in a cancer genome, it is plausible that a region undergoing BFB may also be rearranged by other mechanisms. So, when we evaluate a count vector for consistency with BFB, we must also consider whether the count vector is “nearly” consistent with BFB. For this, we define a distance measure δ between count vectors, where reflects a penalty for assuming that the real copy counts are whereas the measured counts are . We implement such a distance measure based on the Poisson likelihood of the observation, as follows: Let be the Poisson probability of measuring a copy number n, given that the segment’s true copy number is n′. Assuming measurement errors are independent, the probability for measuring a count vector , where the true counts are , is given by Define the distance of from by

For every pair of count vectors and of the same length , the closer to 0 the greater is the similarity between and .

BFB Count-Vector Problem.

With these definitions, we can now precisely pose a set of problems that needs to be solved to evaluate copy number patterns for consistency with BFB.

BFB Count-Vector Problem Variants.

• • Input: A count vector .

• • The decision variant: decide if is a BFB count vector.

• • The search variant: if is a BFB count vector, find a BFB string α such that .

• • The distance variant: Identify a BFB count vector such that is minimized. Output δ.

Outline of the BFB Count-Vector Algorithms

We defer the full details of the algorithms we have developed to the accompanying SI Text, presenting here only essential properties of BFB strings and some intuition of how to incorporate these properties in algorithms for BFB count-vector problems. We focus on the search variant of the problem, where the goal of the algorithm is to output a BFB string α consistent with the input counts, if such a string exists.

Properties of BFB Palindromes.

Call an l-BFB string β of the form an l-BFB palindrome.^† For an l-BFB string α, the string is an l-BFB palindrome by definition (choosing ρ = ɛ and γ = α in Definition 1). In ref. 12 it was shown that every prefix of a BFB string is itself a BFB string; thus, for an l-BFB palindrome , the prefix α of β is also an l-BFB string. Hence, it follows that α is an l-BFB string if and only if is an l-BFB palindrome. For a BFB string α with and a corresponding BFB palindrome , we have that . Thus, a count vector is a BFB count vector if and only if there is a 1-BFB palindrome β such that . Considering BFB palindromes instead of BFB strings will facilitate the algorithm description.

Define an l-block as a palindrome of the form , where β′ is an (l + 1)-BFB palindrome. For example, from the 4-BFB palindromes and , we can produce the 3-blocks and . It may be asserted that an l-block is a special case of an l-BFB palindrome. Next, we show how l-BFB palindromes may be decomposed into l-block substrings.

For a string α ≠ ɛ, denote by top(α) the maximum integer t such that σ_t or occur in α, and define top(ɛ) = 0. For two strings α and β, say that α ≤^t β if top(α) ≤ top(β), and that α <^t β if top(α) ≤ top(β). For example, for α = AB and , top(α) = 2 and top(β) = 4, therefore α <^t β.

Definition 2:

A string α is a convexed l-palindrome if α = ɛ, or α = γβγ such that γ is a convexed l-palindrome, β is an l-BFB palindrome, and γ <^t β.

Although every l-BFB palindrome α is also a convexed l-palindrome (because α = ɛαɛ), not every convexed l-palindrome is a valid BFB string. For example, is a convexed 1-palindrome (choosing , ), yet it is not a 1-BFB string. Instead, we have the following claim, proven in SI Text:

Claim 1.

A string α is an l-BFB palindrome if and only if α = ɛ, α is an l-block, or α = βγβ, such that β is an l-BFB palindrome, γ is a convexed l-palindrome, and γ ≤^t β.

From Definition 2 and Claim 1, it follows that an l-BFB palindrome α is a palindromic concatenation of l-blocks. In addition, for the total count 2n_l of σ_l and in α, α contains exactly n_l l-blocks, where each block contains one occurrence of σ_l and one occurrence of . When n_l is even, α is of the form , each β_i is an l-block. When n_l is odd, α is of the form . In the latter case, say that is the center of α, where in the former case say that the center of α is ɛ. Note that every l-block β appearing in α and different from its center occurs an even number of times in α. If the center of α is an l-block, this particular block is the only block which appears an odd number of times in α, whereas if it is an empty string then no block appears an odd number of times in α.

Now, let β be a 1-BFB palindrome with a count vector . It is helpful to depict β so that each character σ_l is at its own layer l, increasing with increasing l, as shown in Fig. 2A. As β is a concatenation of 1-blocks, we can consider the collection B¹ = {m₁β₁, m₂β₂,…,m_qβ_q} of these blocks, where each count m_i is the number of distinct repeats of β_i in β. For example, for the string in Fig. 2A, B¹ = {2β₁, β₂, 2β₃, 4β₄}, where |B¹| = n₁ = 9, and β₂ is the center of β. Masking from strings in B¹ all occurrences of A and , each 1-block in B¹ becomes a 2-BFB palindrome . Such 2-BFB palindromes may be further decomposed into 2-blocks, yielding a 2-block collection B² (in Fig. 2B, B² = {2β₅, β₆, 2β₇}, where |B²| = n₂ = 5). In general, for each 1 ≤ l ≤ k, masking in β all letters σ_r and such that r < l defines a corresponding collection of l-block substrings of β. Each collection B^l contains exactly n_l elements, as each l-block in the collection contains exactly two out of the 2n_l occurrences of σ_l in the string (where one occurrence is reversed). The collection B^l+1 is obtained from B^l by masking occurrences of σ_l and from the elements in B^l, and decomposing the obtained (l + 1)-BFB palindromes into (l + 1)-blocks. We may define B^k+1 = (where denotes an empty collection), because after masking in β all segments σ₁,…,σ_k we are left with an empty collection of (k + 1)-blocks.

Fig. 2.

Layer visualization of a BFB palindrome , where . A possible BFB sequence that produces α is . , and . (A–D) Layers 1–4 of β, respectively. In each layer l, the l-blocks composing the collection B^l are annotated as substrings of the form β_i.

The algorithm we describe for the search variant of the BFB count-vector problem exploits the above-described property of BFB palindromes. Given a count vector , the algorithm processes iteratively the counts in the vector one by one, from n_k down to n₁, producing a series of collections B^k, B^k−1,…, B¹. Starting with B^k+1 = , each collection B^l in the series is obtained from the preceding collection B^l+1 in a two-step procedure: First, (l + 1)-blocks from B^l+1 are concatenated in a manner that produces an (l + 1)-BFB palindrome collection B′ of size n_l (B′ may contain empty strings, which can be thought of as concatenations of zero elements from B^l+1). Then, B^l is obtained by “wrapping” each element β′ ∈ B′ with a pair of σ_l characters to become an l-block . We will refer to the first step in this procedure as collection folding, and to the second step as collection wrapping. For example, in Fig. 2D, the elements in B⁴ = {4β₁₀} are folded to form a 4-palindrome collection B′ = {2β₁₀β₁₀,ɛ} of size n₃ = 3. After wrapping each element of B′ by C to the left and to the right, we get the 3-block collection . Algorithm SEARCH-BFB in Fig. 3 gives the pseudocode for the described procedure, excluding the implementation of the folding phase, which is kept abstract here. We next discuss some restrictions over the folding procedure, and point out that greedy folding is nontrivial. Nevertheless, in SI Text we show an explicit implementation of a folding procedure, which guarantees that the search algorithm finds a BFB string provided that the input is a valid BFB count vector.

Fig. 3.

Algorithm for the BFB count-vector problem.

Required Conditions for Folding.

Recall that the input of the folding procedure is an l-block collection B and an integer n, and the procedure should concatenate all strings in B in some manner to produce an l-BFB palindrome collection B′ of size n. Because both l-blocks and empty strings are special cases of l-BFB palindromes, when n ≥ |B| it is always possible to obtain B′ by simply adding n − |B| empty strings to B. Nevertheless, when n < |B|, there are instances for which no valid folding exists, as shown next.

For a pair of collections B and B′, B + B′ is the collection containing all elements in B and B′. When B″ = B + B′, we say that B = B″ − B′ (note that B″ − B′ is well-defined only when B″ contains B′). For some (possibly rational) number x ≥ 0, denote by xB the collection {⌊xm₁⌋β₁, ⌊xm₂⌋β₂,…, ⌊xm_q⌋β_q}. The operation mod2(B) yields the subcollection of B containing a single copy of each distinct element β with an odd count in B. For example, for B = {2β₁, β₂, 5β₃, 6β₄}, mod2(B) = {β₂, β₃}. Observe that .

Claim 2.

Let B be an l-BFB palindrome collection such that mod2(B) = . Then, it is possible to concatenate all elements in B to obtain a single l-BFB palindrome.

Proof:

By induction on the size of B. By definition, mod2(B) = implies that the counts of all distinct elements in B are even. When B = , the concatenation of all elements in B yields an empty string ɛ, which is an l-BFB palindrome as required. Otherwise, assume the claim holds for all collections B′ smaller than B. Let β ∈ B be an element such that for every β′ ∈ B, top(β′) ≤ top(β), and let B′ = B − {2β}. Note that mod2(B′) = (the count parity is identical for every element in both B and B′), and from the inductive assumption it is possible to concatenate all elements in B′ into a single l-BFB palindrome α′. From Claim 1, the string α = βα′β is an l-BFB palindrome, obtained by concatenating all elements in B.□

Claim 3.

Let B be an l-block collection. There is a folding B′ of B such that |B′| = |mod2(B)| + 1.

Proof:

Recall that . All element counts in the collection are even, hence , and from Claim 2 it is possible to concatenate all elements in into a single l-BFB palindrome α. Thus, the collection B′ = mod2(B) + α is a folding of B of size |mod2(B)| + 1.□

Claim 4.

For every folding B′ of an l-block collection B, |mod2(B′)| ≥ |mod2(B)|.

Proof:

Let β ∈ mod2(B) be an l-block repeating an odd number of times m in B. Therefore, β appears as a center of at least one element β′ that occurs an odd number of times in B′ (otherwise, β has an even number of distinct repeats as a substring of elements in B′, in contradiction to the fact that m is odd). Hence, for each β ∈ mod2(B) there is a corresponding unique element β′ ∈ mod2(B′), and so |mod2(B′)| ≥ |mod2(B)|.□

The SEARCH-BFB algorithm described in Fig. 3 tries in each iteration l to fold the block collection B^l+1 obtained in the previous iteration into an (l + 1)-BFB palindrome collection of size n_l. When n_l ≥ |mod2(B^l+1)| + 1, there always exists a folding as required: B^l+1 may be folded into a collection of size |mod2(B^l+1)| + 1 due to Claim 3, and additional n_l − |mod2(B^l+1)| − 1 empty strings may be added to get a folding of size n_l. On the other hand, when n_l < |mod2(B^l+1)|, no folding as required exists, due to Claim 4. In the remaining case of n_l = |mod2(B^l+1)|, the existence of an n_l-size folding of B^l+1 depends on the element composition of B^l+1, as exemplified next.

Consider the run of Algorithm SEARCH-BFB over the input count vector . Here, k = 3, and the algorithm starts by initializing the collection B⁴ = . In the first loop iteration l = 3, and the algorithm first tries to fold the empty collection B⁴ into a 4-BFB palindrome collection containing n₃ = 2 elements. Because there are no elements in B⁴ to concatenate, the only way to perform this folding is by adding to B⁴ two empty strings, yielding the collection B′ = {2ɛ}, which after wrapping becomes . In the next iteration l = 2, and B³ should be folded into a collection B′ of size n₂ = 3. Among the possibilities to perform this folding are the following: B′^a = {2β₁, ɛ}, and B′^b = {β₁β₁, 2ɛ}, which after wrapping become , and , respectively. Note that |mod2(B^2a)| = |mod2(B^2b)| = 1. Nevertheless, it is possible to fold B^2a in the next iteration into the collection {β₂β₃β₂} of size n₁ = 1, whereas B^2b cannot be folded into such a collection. The reason is that the only concatenation of all elements in B^2b into a single palindrome is the concatenation β₃β₄β₃, yet , and Claim 1 implies that this concatenation is not a valid BFB palindrome.

In SI Text, we define a property called the signature of a collection, and show how the exact minimum folding size depends on this signature (Table S1). We also show how to fold a collection in a manner that optimizes this signature, and guarantees for valid BFB count-vector inputs that the search algorithm finds an admitting BFB string.

Running Time

For a count vector , let be the number of segments in a string corresponding to . Let denote a number proportional to the number of bits in the representation of , assuming each count n_i is represented by O(log n_i) bits. In SI Text, we complete the implementation details of algorithms for the decision, search, and distance variants of the BFB count-vector problem, and show that these algorithms have the asymptotic running times of (bit operations), O(N), and O(N^{log N}) (under some realistic assumptions), respectively. For the decision and search variants, these running times are optimal, being linear in the input (for the decision variant) or output (for the search variant) lengths.

In practical terms, this has a significant effect on our ability to evaluate copy number signatures of BFB compared with the previous exponential time algorithm (12). To determine if a count vector consistent with BFB is in fact strong evidence for BFB, we have to check many count vectors. Analyzing the simulations we explain below required testing tens of millions of different count vectors, so even a small improvement in running time can have a large impact of the scope of analysis we can perform. However, the running time improvement with this algorithm is not small. For example, a count vector that took 9 s with the previous algorithm can be processed by the present algorithm in 1.2 × 10⁻⁵ s. A count vector that was abandoned after 30 h with the old algorithm now takes only 8.1 × 10⁻⁶ s. Thus, the improvement in running time is not of merely theoretical interest, and was required for making the current study possible.

Detecting Signatures of BFB

We can now describe the two features we will use to determine if a chromosome has undergone BFB. The first feature is based on the fold-back inversions that BFB produces. For a given region, we can find all of the breakpoints identified by sequencing and determine what proportion are fold-back inversions. We call this the fold-back fraction. The second feature relies on our algorithm that solves the BFB count-vector problems we have posed. For a given count vector, we can find the distance to the nearest count vector that could be produced by BFB using the distance metric we defined above. We call this the count-vector distance. For a particular count vector, we define a score s that combines these two features:

Here, f refers to the fold-back fraction, δ refers to the count-vector distance, and λ refers to the weight we give to the count-vector distance versus the fold-back fraction when calculating the score. When λ = 1, we are only looking at count-vector distance, whereas when λ = 0, we are only using fold-back fraction and ignoring the count vectors.

Results

To determine whether our two proposed features could identify BFB against the complex backdrop of a cancer genome, we simulated rearranged chromosomes. Our overall goal was to simulate cancer chromosomes that were highly rearranged yet had not undergone BFB to see if evidence for BFB appeared in them, suggesting that using such evidence would lead to false positives. Conversely, we also wanted to simulate chromosomes whose rearrangements included BFB to determine if a proposed BFB signature was sensitive enough to identify BFB when it occurred. Because it is not clear how to faithfully simulate cancer genome rearrangements, we used a wide range of simulation parameters so we could understand how different assumptions affect the features’ ability to identify BFB.

We began with a pair of unrearranged chromosomes and then introduced 50 rearrangements to each. Each rearrangement was an inversion, a deletion, or a duplication. Duplications were either direct or inverted and could be tandem or interspersed. The type of each rearrangement was chosen from a distribution. In some chromosome pairs, we imitated BFB by successively duplicating and inverting segments of one end of one chromosome for each round of BFB. The number of BFB rounds varied from 2 to 10. Then, we calculated the copy counts and breakpoints for the chromosome pair and introduced error to the copy counts according to a random model and also randomly deleted or inserted breakpoint observations. For each combination of rearrangement type distribution and number of BFB rounds, we simulated 5,000 chromosome pairs with BFB and 15,000 without BFB. Complete details are in SI Text.

We first examined the usefulness of count-vector distance alone in identifying BFB by setting λ = 1 in our score function (Eq. 1). For each chromosome pair, we found all contiguous count vectors of a given length and calculated their scores, as described above and in SI Text. We used the minimum score s over all of these subvectors in the chromosome as a score for the whole chromosome. Then, for varying thresholds, we classified all chromosomes with a score lower than the threshold as having been rearranged by BFB. The performance of this classification varied with the parameters used to simulate the chromosomes, but typical results can be seen in Fig. 4A. The solid lines show receiver operating characteristic (ROC) curves for different count-vector lengths for the simulation with eight rounds of BFB and a distribution that yields roughly equal probabilities of the other rearrangement types. Consistent with previous observations, short count vectors that are perfectly consistent with BFB can be found in many chromosomes, even if BFB did not occur. So, even with a score threshold of zero, they would still be classified as consistent with BFB. For example, 63% of chromosomes without any true BFB rearrangements in Fig. 4A had a count vector of length 6, perfectly consistent with BFB.

Fig. 4.

Simulation and pancreatic cancer results. (A) ROC curves for different count-vector lengths with and without fold-back fractions for the simulation of eight BFB rounds and equally likely other rearrangements. (B) Observed copy counts and copy counts compatible with BFB on the short arm of chromosome 12 in pancreatic cancer sample PD3641. The presence of fold-back inversions and the count vector’s consistency with BFB suggests that this portion of chromosome 12 underwent BFB cycles.

In contrast, examining longer count vectors produced a better classification. For instance, setting the score threshold to 0.10, count vectors of length 12 could achieve a true positive rate (TPR) of 67% and a false positive rate (FPR) of only 10%. However, this performance must be considered in the context of an experiment seeking evidence for BFB. Chromosomes that have undergone BFB are probably rare. If only 1 in 100 chromosomes tested underwent BFB, then a test with an FPR of even 1% will produce mostly false discoveries. Achieving this FPR with count vectors of length 12 with the chromosomes in Fig. 4A would result in a TPR of only 16%. A more appropriate target FPR for screening many samples, say 0.1%, could not be achieved with count vectors alone.

Next, we incorporated fold-back inversions into the scoring function by setting λ = 0.5. ROC curves using this approach are shown by dashed lines in Fig. 4A. Incorporating fold-back fractions into the scoring leads to better discrimination of chromosomes with and without BFB rearrangements; the test in Fig. 4A that combines count vectors of length 12 and fold-back inversions can achieve a TPR of 48% with an FPR of 0.1% by setting the score threshold to 0.27. This suggests that it could detect BFB in a large dataset without being overwhelmed by false discoveries.

Of course, these conclusions depend on our simulation resembling actual cancer rearrangements and BFB cycles. A true specification of cancer genome evolution is unknown and in any case varies from cancer to cancer. Recognizing this complication, we repeated the analysis in Fig. 4A for the different rearrangement distributions, number of BFB rounds, and count-vector lengths. For each combination, we recorded the score threshold needed to achieve FPRs of 0.1, 1, and 5%, and the respective expected TPRs. The full results are shown in Dataset S1 and ROC curves are shown in Figs. S1–S5. Generally, different simulations showed the same trends. Fold-back inversions alone were better at identifying BFB than count vectors alone, but the combination of both features provided the best classification. By examining a wide range of simulation parameters, we illustrate how changes in assumptions about cancer genome evolution and BFB influence the appropriateness and expected outcomes of tests for BFB.

We applied our method to a publicly available dataset of copy number profiles from 746 cancer cell lines (13). We found three chromosomes with count vectors of length 12 nearly consistent with BFB: chromosome 8 from cell line AU565, chromosome 10 from cell line PC-3, and chromosome 8 from cell line MG-63 (SI Text). Although the patterns of copy counts on these chromosomes do bear the hallmarks of BFB, our simulations suggest that labeling chromosomes as having undergone BFB based on these count vectors would lead to an FPR between 1 and 10%. Given that thousands of chromosomes were examined, many of which were highly rearranged, the consistency of these copy counts with BFB may be spurious.

We also applied our method to paired-end sequencing data from seven previously published pancreatic cancer samples (14). We estimated copy numbers from the reads and used breakpoints as reported by the original investigators (Fig. S6). We examined count vectors of length 8 and chose a threshold score of 0.18, which would give an FPR of 0.1% based on simulations where the non-BFB rearrangement types are roughly equally likely. We identified two chromosomes that showed evidence for BFB, both from the same sample, PD3641. The first was the long arm of chromosome 8. This chromosome was identified by the original investigators as likely being rearranged by BFB. Our analysis suggests that, barring rearrangements that differ significantly from any of our simulations, this chromosome did indeed undergo BFB cycles. We also found evidence for BFB rearrangements from a count vector spanning 10 megabases on the short arm of chromosome 12 (Fig. 4B and Fig. S7). Thus, we were able to recover evidence for BFB previously identified by hand curation. And, by combining count vector and fold-back analysis, we found an additional strong BFB candidate that would not be apparent without modeling and simulation.

Discussion

Some 80 years after Barbara McClintock’s discovery of the BFB mechanism, it is seeing renewed interest in the context of tumor genome evolution. Recent publications have claimed, based on empirical observations of segmentation counts and other features, that their data counts are consistent with BFB. The main technical contribution of the paper is an efficient algorithm for detecting if given segmentation counts can indeed be created by BFB cycles. That algorithm turns out to be nontrivial, requiring a deep foray into the combinatorics of BFB count vectors, even though its final implementation is straightforward and fast. Experimenting with the implementation reveals that in fact (i) there is a big diversity of count vectors created by true BFB cycles, not all of which are easily recognizable as BFB; and (ii) at least for short count vectors, it is often possible to create BFB-like vectors by non-BFB operations. Thus, being “consistent with BFB” and “caused by BFB” are not equivalent. Fortunately, our results also suggest that using longer count vectors, and additional information of fold-backs, gives stronger prediction of BFB, even in the presence of noise and diploidy. Although assembly of these highly rearranged genomes continues to be difficult, recent advances in long single-molecule sequencing will provide additional spatial information that will improve the resolving power of our algorithm. As more cancer genomes are sequenced, including single-cell sequencing, the method presented here will be helpful in determining the extent and scope of BFB cycles in the evolution of the tumor genome.

Materials and Methods

Details on the algorithm, and on the simulation methods are available in SI Text.

Code Availability.

Java and Python code used to analyze chromosomes is available at www.bitbucket.org/mckinsel/bfb.

Estimating Pancreas Tumor Copy Number.

The pancreas tumor data were downloaded from the European Genome-Phenome Archive, accession no. EGAS00000000064. The data were paired-end reads; each end was 37 bases long. We aligned the reads with Bowtie (17). Then we used readDepth (18) for segmentation and integer copy number estimation.