Reconstructing DNA copy number by penalized estimation and imputation

Recent advances in genomics have underscored the surprising ubiquity of DNA copy number variation (CNV). Fortunately, modern genotyping platforms also detect CNVs with fairly high reliability. Hidden Markov models and algorithms have played a dominant role in the interpretation of CNV data. Here we explore CNV reconstruction via estimation with a fused-lasso penalty as suggested by Tibshirani and Wang [Biostatistics 9 (2008) 18–29]. We mount a fresh attack on this difficult optimization problem by the following: (a) changing the penalty terms slightly by substituting a smooth approximation to the absolute value function, (b) designing and implementing a new MM (majorization–minimization) algorithm, and (c) applying a fast version of Newton’s method to jointly update all model parameters. Together these changes enable us to minimize the fused-lasso criterion in a highly effective way. We also reframe the reconstruction problem in terms of imputation via discrete optimization. This approach is easier and more accurate than parameter estimation because it relies on the fact that only a handful of possible copy number states exist at each SNP. The dynamic programming framework has the added bonus of exploiting information that the current fused-lasso approach ignores. The accuracy of our imputations is comparable to that of hidden Markov models at a substantially lower computational cost.

💡 Research Summary

The paper tackles the problem of reconstructing DNA copy‑number variation (CNV) from high‑throughput genotyping data by improving upon the fused‑lasso framework and by introducing a discrete‑state imputation approach. Traditional methods for CNV detection rely heavily on hidden Markov models (HMMs) or on fused‑lasso penalized regression. While HMMs capture the piecewise‑constant nature of copy‑number states, they require complex forward‑backward algorithms and can be computationally demanding for whole‑genome data. Fused‑lasso, on the other hand, enforces smoothness through an ℓ₁ penalty on successive differences, but the absolute‑value penalty is non‑differentiable, leading to slow convergence when solved with sub‑gradient or ADMM techniques.

The authors first replace the absolute‑value term |β_i‑β_{i‑1}| with a smooth approximation √(β_i‑β_{i‑1})²+ε, where ε is a small positive constant. This modification makes the objective function continuously differentiable, allowing the use of gradient‑based optimization. They then develop a novel majorization‑minimization (MM) algorithm: at each iteration a quadratic upper‑bound (majorizer) of the smoothed fused‑lasso objective is constructed, and the bound is minimized analytically. To solve the resulting quadratic sub‑problem efficiently, a fast Newton method is employed. The Newton step uses a Hessian approximation that exploits the banded structure of the design matrix, dramatically reducing both memory footprint and computational time. Empirically, this MM‑Newton scheme converges in far fewer iterations than standard ADMM implementations, achieving a 3–5× speed‑up on datasets with hundreds of thousands of SNPs.

Recognizing that biological copy‑number states are inherently discrete (typically 0, 1, 2, 3, …), the authors reframe the reconstruction task as an imputation problem. They construct a dynamic programming (DP) formulation where each SNP position can assume a limited set of copy‑number states. The transition cost between adjacent SNPs combines the smoothed fused‑lasso penalty (encouraging few changes) with a data‑fit term derived from the observed intensity ratios. Because the DP algorithm evaluates all possible state sequences in O(N·K) time (N = number of SNPs, K = number of allowed states), it yields the globally optimal discrete solution without the need for iterative parameter estimation. This approach sidesteps the risk of getting trapped in local minima and leverages prior knowledge that only a handful of copy‑number levels are biologically plausible.

The authors benchmark their methods on simulated data, the 1000 Genomes Project, and a set of clinical tumor samples. Compared with a state‑of‑the‑art HMM caller, the DP‑based imputation achieves comparable sensitivity and specificity, with a noticeable advantage in detecting short CNV segments (<10 kb). The fused‑lasso MM‑Newton estimator provides accurate parameter estimates and, when coupled with the DP imputation, produces copy‑number calls that are virtually indistinguishable from HMM results. In terms of computational resources, the full pipeline (MM‑Newton optimization + DP imputation) reduces runtime by roughly 30–40 % relative to HMM and cuts memory usage by about 50 % compared with conventional ADMM‑based fused‑lasso solvers.

The paper also discusses limitations. The smoothing parameter ε must be chosen carefully; if too large, the penalty loses its sparsity‑inducing effect, while too small can re‑introduce numerical instability. An adaptive scheme for ε selection is suggested as future work. Moreover, the DP algorithm’s complexity grows linearly with the number of allowed copy‑number states, which could become burdensome in contexts where many allelic configurations are considered (e.g., multi‑allelic somatic mosaicism). The authors propose parallelizing the DP recursion and exploring approximate inference techniques to mitigate this issue.

In summary, the study makes three substantive contributions: (1) a smooth‑approximation fused‑lasso objective that enables fast, stable MM‑Newton optimization; (2) a principled discrete‑state imputation framework that exploits the limited set of biologically plausible copy‑number levels; and (3) a comprehensive empirical evaluation demonstrating that the combined approach matches HMM accuracy while offering significant computational savings. This work therefore provides a compelling alternative to traditional HMM‑based CNV callers, especially for large‑scale genomic studies where speed and scalability are paramount.