Multi-task learning for HIV therapy screening

Abstract

We address the problem of learning classifiers for a large number of tasks. We derive a solution that produces resampling weights which match the pool of all examples to the target distribution of any given task. Our work is motivated by the problem of predicting the outcome of a therapy attempt for a patient who carries an HIV virus with a set of observed genetic properties. Such predictions need to be made for hundreds of possible combinations of drugs, some of which use similar biochemical mechanisms. Multi-task learning enables us to make predictions even for drug combinations with few or no training examples and substantially improves the overall prediction accuracy.

Learning from Past Treatments and Their Outcome Improves Prediction of In Vivo Response to Anti-HIV Therapy

“…The sequence boosting method presented here, considered up to N-way interaction features, where N is the number of treatment events in the longest treatment record in the training data. Other approaches, such as Bayesian networks (Deforche et al, 2006) or transfer learning (Bickel et al, 2008) were applied to the same treatment outcome prediction problem, but none of them employed such a large number of features.…”

“…This concept was recently extended to predict in vivo response to combination treatments (Altmann et al, 2007(Altmann et al, , 2009Larder, 2007). However, so far computer-based methods make use of the patient's treatment history only by using binary indicators of previous exposure to a drug as additional features (Bickel et al, 2008;Larder, 2007;Rosen-Zvi et al, 2008). While this representation perfectly summarizes previous drug applications, it may miss important and informative cause-effect relationships, such as: the drug efavirenz (EFV) selects mutation RT103N, which leads to the administration of a new drug combination including lopinavir (LPV), but not any drug from the same class as EFV (see Figure 2 for an example).…”

Relaxed Transfer of Different Classes via Spectral Partition

“…There are many interpretations of transfer learning strategies, e.g., [3,4] reviewed in Section 5. One straightforward strategy is formulated as follows.…”

“…One of the main issues in transfer learning is how to transfer knowledge across different data distributions. A general approach is based on re-sampling (e.g., [3]), where the motivation of it is to "emphasize" the knowledge among "similar" and discriminating instances. Another line of work is to transfer knowledge based on the common features found in a subspace (e.g., [5]) or a projected feature space where the different tasks are similar to each other (e.g., [2]).…”

“…As an alternative solution, transfer learning was proposed to help extract some supervisory knowledge from Part of the work was done when the author was a visiting student at HKUST. related source data to help learn the target task (e.g., [3,4,16,22]). Existing transfer learning techniques implicitly assume that there is sufficient overlap between the source data and the target data, and categories of class labels are the same, in order to allow the transfer of knowledge.…”