[Paper Review] Joint segmentation of many aCGH profiles using fast group LARS

Kevin Bleakley, Jean‐Philippe Vert|ArXiv.org|Oct 7, 2009

Medical Imaging Techniques and Applications25 references3 citations

TL;DR

This paper proposes a fast group LARS algorithm for joint segmentation of multiple aCGH profiles by enforcing shared breakpoints across all profiles, enabling simultaneous detection of recurrent copy number variations (CNVs) while smoothing data into piecewise-constant regions. The method achieves O(knp) time complexity and O(np) memory usage, allowing rapid identification of biologically relevant breakpoints even in noisy data, with validation on a bladder cancer dataset confirming known and novel CNVs.

ABSTRACT

Array-Based Comparative Genomic Hybridization (aCGH) is a method used to search for genomic regions with copy numbers variations. For a given aCGH profile, one challenge is to accurately segment it into regions of constant copy number. Subjects sharing the same disease status, for example a type of cancer, often have aCGH profiles with similar copy number variations, due to duplications and deletions relevant to that particular disease. We introduce a constrained optimization algorithm that jointly segments aCGH profiles of many subjects. It simultaneously penalizes the amount of freedom the set of profiles have to jump from one level of constant copy number to another, at genomic locations known as breakpoints. We show that breakpoints shared by many different profiles tend to be found first by the algorithm, even in the presence of significant amounts of noise. The algorithm can be formulated as a group LARS problem. We propose an extremely fast way to find the solution path, i.e., a sequence of shared breakpoints in order of importance. For no extra cost the algorithm smoothes all of the aCGH profiles into piecewise-constant regions of equal copy number, giving low-dimensional versions of the original data. These can be shown for all profiles on a single graph, allowing for intuitive visual interpretation. Simulations and an implementation of the algorithm on bladder cancer aCGH profiles are provided.

Motivation & Objective

• To address the challenge of accurately segmenting multiple aCGH profiles into regions of constant copy number while leveraging shared biological patterns across profiles.
• To overcome limitations of existing methods that treat profiles independently or require kernel window selection, discretization, or high computational cost.
• To develop a scalable, fast algorithm that explicitly encodes the biological prior that patients with the same disease share key CNV breakpoints.
• To produce smoothed, low-dimensional representations of aCGH data that preserve shared structural features and enable downstream analysis such as classification.

Proposed method

• Formulates joint segmentation as a group LARS problem with a fused lasso-type constraint that enforces all profiles to share the same set of breakpoints.
• Reformulates the optimization problem to avoid explicit construction of the large p²n² matrix, enabling O(knp) time and O(np) memory complexity.
• Uses a solution path algorithm to sequentially identify the most significant shared breakpoints in order of importance, based on their contribution to reducing residual error.
• Applies piecewise-constant smoothing by replacing probe values between consecutive breakpoints with their mean, producing low-dimensional, interpretable profiles.
• Leverages the shared breakpoint structure to ensure consistent vector representation across all profiles, suitable for machine learning applications.
• Employs a fast active set strategy to efficiently compute the solution path without storing the full design matrix.

Experimental results

Research questions

• RQ1 Can a joint segmentation method that enforces shared breakpoints across multiple aCGH profiles improve detection of biologically relevant copy number variations compared to independent segmentation?
• RQ2 How can the computational cost of joint segmentation be minimized while maintaining accuracy and scalability for large aCGH datasets?
• RQ3 To what extent does enforcing shared breakpoints enhance robustness to noise and improve detection of recurrent CNVs in cancer genomics?
• RQ4 Can the resulting smoothed, piecewise-constant profiles serve as effective low-dimensional features for downstream classification tasks?
• RQ5 What is the optimal stopping criterion for selecting the number of breakpoints in the joint segmentation process?

Key findings

• The algorithm identified 80 shared breakpoints in 57 bladder cancer aCGH profiles in just 0.72 seconds on a 2008 MacBook Pro, demonstrating O(knp) scalability.
• The method successfully recovered nearly all known recurrent CNVs in bladder cancer, including frequent duplications at 8q22-24, 17q21, and 20q, and deletions at 8p22-23, 13q, 17p, 11p, and chromosome 9.
• Even profiles with minimal CNVs, such as one showing only a deletion on chromosome 9, were smoothly and accurately represented after forced alignment to shared breakpoints.
• The algorithm’s ability to detect shared breakpoints first, even in noisy conditions, confirms its robustness and biological relevance.
• The smoothed profiles, when superimposed on a single graph, enabled intuitive visual interpretation of recurrent CNVs across the entire cohort.
• The method produces consistent, low-dimensional vector representations of aCGH data that are suitable for classification and other machine learning applications.