Domain Adaption for Genomic Sequence Analysis

Supervised learning methods are machine learning techniques that use a set of labeled training examples (x_i; y_i) to infer a model that generalizes well to unseen examples x, and therefore allows to make predictions y for these instances. One of the underlying assumptions is that training and test data are drawn from the same distribution. However, in real-world applications, and specifically in computational biology, one often faces the situation that little or no training data is available for the problem of interest (the target domain), whereas a large corpus of labeled data exists for a related but different problem (the source domain). While it is highly desirable to exploit the labeled source data points to improve performance on the target domain, it clearly violates the basic assumptions of machine learning. Methods to tackle this so-called domain adaptation problem have recently caught increasing attention in the machine learning community and are subject of consideration in this work. In particularly we examine their effectiveness for the problem of splice site prediction across different organisms with varying evolutionary distance.

This work is a collaborative project with Bernhard Schoelkopf.

See splicing and domain adaptation for more details.

People specializing in this area

Graduate Students

Christian Widmer

More about Christian Widmer...

Alumnae and Alumni

Jose Miguel Leiva, Ph.D.

More about Jose Miguel Leiva, Ph.D....

Gabriele Schweikert, Ph.D.

My main interested in domain adaptation methods is in their application to sequence analysis and in particular to gene finding. To date gene finding methods require a large amount of labeled data (e.i. previously annotated genes) to train a model for the subsequent prediction of unknown genes in the rest of the genome. To achieve highly accurate predictions, training is performed for each organism separately, e.g. a model learnt on fly genes is not working well to predict genes in a worm genome as the characteristics of the genes have diverged during evolution. However, it is also known that the cellular machineries that are responsible for transcription, processing and translation of genes are highly conserved between organisms. Therefore the tasks of predicting genes in flies and worms are similar albeit different. By applying domain adaptation techniques we try to close the remaining gap.

More about Gabriele Schweikert, Ph.D....

Publications

• Widmer, C, Leiva, J, Altun, Y, and Rätsch, G (2010).
  Leveraging Sequence Classification by Taxonomy-based Multitask Learning
  In: Proc. RECOMB'10.
  Printable file

• Schweikert, G, Widmer, C, Schölkopf, B, and Rätsch, G (2008).
  An empirical Analysis of Domain Adaptation Algorithms
  In: Proc. NIPS 2008. Advances in Neural Information Processing Systems.
  Supplementary website /raetsch/suppl/genomedomainadaptation
  Printable file

• Widmer, C (2008).
  Domain Adaptation in Sequence Analysis
  Diplomarbeit, University of Tübingen.
  Printable file