Learning linear combinations of multiple kernels is an appealing strategy when the right choice of features is unknown. Previous approaches to multiple kernel learning (MKL) promote sparse kernel combinations and hence support interpretability. Unfortunately, L1-norm MKL is hardly observed to outperform trivial baselines in practical applications. To allow for robust kernel mixtures, we generalize MKL to arbitrary Lp-norms. We devise new insights on the connection between several existing MKL formulations and develop two efficient interleaved optimization strategies for arbitrary p>1. Empirically, we demonstrate that the interleaved optimization strategies are much faster compared to the traditionally used wrapper approaches. Finally, we apply Lp-norm MKL to real-world problems from computational biology, showing that non-sparse MKL achieves accuracies that go beyond the state-of-the-art.

MOTIVATION: At the heart of many important bioinformatics problems, such as gene finding and function prediction, is the classification of biological sequences. Frequently the most accurate classifiers are obtained by training support vector machines (SVMs) with complex sequence kernels. However, a cumbersome shortcoming of SVMs is that their learned decision rules are very hard to understand for humans and cannot easily be related to biological facts. RESULTS: To make SVM-based sequence classifiers more accessible and profitable, we introduce the concept of positional oligomer importance matrices (POIMs) and propose an efficient algorithm for their computation. In contrast to the raw SVM feature weighting, POIMs take the underlying correlation structure of k-mer features induced by overlaps of related k-mers into account. POIMs can be seen as a powerful generalization of sequence logos: they allow to capture and visualize sequence patterns that are relevant for the investigated biological phenomena. AVAILABILITY: All source code, datasets, tables and figures are available at http://www.fml.tuebingen.mpg.de/raetsch/projects/POIM. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

We study the problem of learning kernel machines transductively for structured output variables. Transductive learning can be reduced to combinatorial optimization problems over all possible labelings of the unlabeled data. In order to scale transductive learning to structured variables, we transform the corresponding non-convex, combinatorial, constrained optimization problems into continuous, unconstrained optimization problems. The discrete optimization parameters are eliminated and the resulting differentiable problems can be optimized efficiently. We study the effectiveness of the generalized TSVM on multiclass classification and label-sequence learning problems empirically.

In many applications it is desirable to learn from several kernels. "Multiple kernel learning" (MKL) allows the practitioner to optimize over linear combinations of kernels. By enforcing sparse coefficients, it also generalizes feature selection to kernel selection. We propose MKL for joint feature maps. This provides a convenient and principled way for MKL with multiclass problems. In addition, we can exploit the joint feature map to learn kernels on output spaces. We show the equivalence of several different primal formulations including different regularizers. We present several optimization methods, and compare a convex quadratically constrained quadratic program (QCQP) and two semi-infinite linear programs (SILPs) on toy data, showing that the SILPs are faster than the QCQP. We then demonstrate the utility of our method by applying the SILP to three real world datasets.

Semi-Supervised Support Vector Machines (S3VMs) are an appealing method for using unlabeled data in classification: their objective function favors decision boundaries which do not cut clusters. However their main problem is that the optimization problem is non-convex and has many local minima, which often results in suboptimal performances. In this paper we propose to use a global optimization technique known as continuation to alleviate this problem. Compared to other algorithms minimizing the same objective function, our continuation method often leads to lower test errors.

It is common in classification methods to first place data in a vector space and then learn decision boundaries. We propose reversing that process: for fixed decision boundaries, we "learn" the location of the data. This way we (i) do not need a metric (or even stronger structure) - pairwise dissimilarities suffice; and additionally (ii) produce low-dimensional embeddings that can be analyzed visually. We achieve this by combining an entropy-based embedding method with an entropy-based version of semi-supervised logistic regression. We present results for clustering and semi-supervised classification.

We report and compare the performance of different learning algorithms based on data from cortical recordings. The task is to predict the orienta- tion of visual stimuli from the activity of a population of simultaneously recorded neurons. We compare several ways of improving the coding of the input (i.e., the spike data) as well as of the output (i.e., the orienta- tion), and report the results obtained using different kernel algorithms.

MOTIVATION: Large scale gene expression data are often analysed by clustering genes based on gene expression data alone, though a priori knowledge in the form of biological networks is available. The use of this additional information promises to improve exploratory analysis considerably. RESULTS: We propose constructing a distance function which combines information from expression data and biological networks. Based on this function, we compute a joint clustering of genes and vertices of the network. This general approach is elaborated for metabolic networks. We define a graph distance function on such networks and combine it with a correlation-based distance function for gene expression measurements. A hierarchical clustering and an associated statistical measure is computed to arrive at a reasonable number of clusters. Our method is validated using expression data of the yeast diauxic shift. The resulting clusters are easily interpretable in terms of the biochemical network and the gene expression data and suggest that our method is able to automatically identify processes that are relevant under the measured conditions.

We estimate the number of microarrays that is required in order to gain reliable results from a common type of study: the pairwise comparison of different classes of samples. Current knowlegde seems to suffice for the construction of models that are realistic with respect to searches for individual differentially expressed genes. Such models allow to investigate the dependence of the required number of samples on the relevant parameters: the biological variability of the samples within each class; the fold changes in expression; the detection sensitivity of the microarrays; and the acceptable error rates of the results. We supply experimentalists with general conclusions as well as a freely accessible Java applet at http://cartan.gmd.de/~zien/classsize/ for fine tuning simulations to their particular actualities. Since the situation can be assumed to be very similar for large scale proteomics and metabolomics studies, our methods and results might also apply there.

MOTIVATION: We present an extensive evaluation of different methods and criteria to detect remote homologs of a given protein sequence. We investigate two associated problems: first, to develop a sensitive searching method to identify possible candidates and, second, to assign a confidence to the putative candidates in order to select the best one. For searching methods where the score distributions are known, p-values are used as confidence measure with great success. For the cases where such theoretical backing is absent, we propose empirical approximations to p-values for searching procedures. RESULTS: As a baseline, we review the performances of different methods for detecting remote protein folds (sequence alignment and threading, with and without sequence profiles, global and local). The analysis is performed on a large representative set of protein structures. For fold recognition, we find that methods using sequence profiles generally perform better than methods using plain sequences, and that threading methods perform better than sequence alignment methods. In order to assess the quality of the predictions made, we establish and compare several confidence measures, including raw scores, z-scores, raw score gaps, z-score gaps, and different methods of p-value estimation. We work our way from the theoretically well backed local scores towards more explorative global and threading scores. The methods for assessing the statistical significance of predictions are compared using specificity--sensitivity plots. For local alignment techniques we find that p-value methods work best, albeit computationally cheaper methods such as those based on score gaps achieve similar performance. For global methods where no theory is available methods based on score gaps work best. By using the score gap functions as the measure of confidence we improve the more powerful fold recognition methods for which p-values are unavailable. AVAILABILITY: The benchmark set is available upon request.

Various bioinformatics problems require optimizing several different properties simultaneously. For example, in the protein threading problem, a scoring function combines the values for different parameters of possible sequence-to-structure alignments into a single score to allow for unambiguous optimization. In this context, an essential question is how each property should be weighted. As the native structures are known for some sequences, a partial ordering on optimal alignments to other structures, e.g., derived from structural comparisons, may be used to adjust the weights. To resolve the arising interdependence of weights and computed solutions, we propose a heuristic approach: iterating the computation of solutions (here, threading alignments) given the weights and the estimation of optimal weights of the scoring function given these solutions via systematic calibration methods. For our application (i.e., threading), this iterative approach results in structurally meaningful weights that significantly improve performance on both the training and the test data sets. In addition, the optimized parameters show significant improvements on the recognition rate for a grossly enlarged comprehensive benchmark, a modified recognition protocol as well as modified alignment types (local instead of global and profiles instead of single sequences). These results show the general validity of the optimized weights for the given threading program and the associated scoring contributions.

MOTIVATION: In order to extract protein sequences from nucleotide sequences, it is an important step to recognize points at which regions start that code for proteins. These points are called translation initiation sites (TIS). RESULTS: The task of finding TIS can be modeled as a classification problem. We demonstrate the applicability of support vector machines for this task, and show how to incorporate prior biological knowledge by engineering an appropriate kernel function. With the described techniques the recognition performance can be improved by 26% over leading existing approaches. We provide evidence that existing related methods (e.g. ESTScan) could profit from advanced TIS recognition.