MOTIVATION: The heterogeneity of cancer cannot always be recognized by tumor morphology, but may be reflected by the underlying genetic aberrations. Array comparative genome hybridization (array-CGH) methods provide high-throughput data on genetic copy numbers, but determining the clinically relevant copy number changes remains a challenge. Conventional classification methods for linking recurrent alterations to clinical outcome ignore sequential correlations in selecting relevant features. Conversely, existing sequence classification methods can only model overall copy number instability, without regard to any particular position in the genome. RESULTS: Here, we present the heterogeneous hidden conditional random field, a new integrated array-CGH analysis method for jointly classifying tumors, inferring copy numbers and identifying clinically relevant positions in recurrent alteration regions. By capturing the sequentiality as well as the locality of changes, our integrated model provides better noise reduction, and achieves more relevant gene retrieval and more accurate classification than existing methods. We provide an efficient L1-regularized discriminative training algorithm, which notably selects a small set of candidate genes most likely to be clinically relevant and driving the recurrent amplicons of importance. Our method thus provides unbiased starting points in deciding which genomic regions and which genes in particular to pursue for further examination. Our experiments on synthetic data and real genomic cancer prediction data show that our method is superior, both in prediction accuracy and relevant feature discovery, to existing methods. We also demonstrate that it can be used to generate novel biological hypotheses for breast cancer.

This paper studies boosting algorithms that make a single pass over a set of base classi(cid:2)ers. We (cid:2)rst analyze a one-pass algorithm in the setting of boosting with diverse base classi(cid:2)ers. Our guarantee is the same as the best proved for any boosting algo- rithm, but our one-pass algorithm is much faster than previous approaches. We next exhibit a random source of examples for which a (cid:147)picky(cid:148) variant of Ad- aBoost that skips poor base classi(cid:2)ers can outperform the standard AdaBoost al- gorithm, which uses every base classi(cid:2)er, by an exponential factor. Experiments with Reuters and synthetic data show that one-pass boosting can sub- stantially improve on the accuracy of Naive Bayes, and that picky boosting can sometimes lead to a further improvement in accuracy.

In 3D shape classification scenarios with classes arranged in a hierarchy from most general to most specific, the use of an independent classifier for each class can produce predictions that are inconsistent with the parent-child relationships of the hierarchy. To be consistent, an example shape must not be assigned to a class unless it is also assigned to its parent class. This paper presents a Bayesian framework for combining multiple classifiers based on a class hierarchy. Given a set of independent classifiers for an arbitrary type of shape descriptor, we combine their possibly inconsistent predictions in our Bayesian framework to obtain the most probable consistent set of predictions. Such error correction is expected to improve accuracy on the overall classification by utilizing the structure of the hierarchy. Our experiments show that over the 170-class hierarchical Princeton Shape Benchmark using the Spherical Harmonic Descriptor (SHD) our algorithm improves the classification accuracy of the majority of classes, in comparison to independent classifiers. Our method is also more effective than straightforward heuristics for correcting hierarchical inconsistencies.

MOTIVATION: Assigning functions for unknown genes based on diverse large-scale data is a key task in functional genomics. Previous work on gene function prediction has addressed this problem using independent classifiers for each function. However, such an approach ignores the structure of functional class taxonomies, such as the Gene Ontology (GO). Over a hierarchy of functional classes, a group of independent classifiers where each one predicts gene membership to a particular class can produce a hierarchically inconsistent set of predictions, where for a given gene a specific class may be predicted positive while its inclusive parent class is predicted negative. Taking the hierarchical structure into account resolves such inconsistencies and provides an opportunity for leveraging all classifiers in the hierarchy to achieve higher specificity of predictions. RESULTS: We developed a Bayesian framework for combining multiple classifiers based on the functional taxonomy constraints. Using a hierarchy of support vector machine (SVM) classifiers trained on multiple data types, we combined predictions in our Bayesian framework to obtain the most probable consistent set of predictions. Experiments show that over a 105-node subhierarchy of the GO, our Bayesian framework improves predictions for 93 nodes. As an additional benefit, our method also provides implicit calibration of SVM margin outputs to probabilities. Using this method, we make function predictions for multiple proteins, and experimentally confirm predictions for proteins involved in mitosis. SUPPLEMENTARY INFORMATION: Results for the 105 selected GO classes and predictions for 1059 unknown genes are available at: http://function.princeton.edu/genesite/ CONTACT: [email protected].