Before laboratory testing, bioinformatics problems often require a machine-learned predictor to identify the most likely choices among a wealth of possibilities. Researchers may advocate different predictors for the same problem, none of which is best in all situations. This paper introduces a case-based meta-predictor that combines a set of elaborate, pre-existing predictors to improve their accuracy on a difficult and important problem: protein-ligand docking. The method focuses on the reliability of its component predictors, and has broad potential applications in biology and chemistry. Despite noisy and biased input, the method outperforms its individual components on benchmark data. It provides a promising solution for the performance improvement of compound virtual screening, which would thereby reduce the time and cost of drug discovery.

As biological data about genes and their interactions proliferates, scientists have the opportunity to identify sets of proteins whose interactions make them worthy of further investigation. This paper reports on a knowledge discovery technique to support that work. Foretell is an algorithm originally designed to support search for solutions to constraint satisfaction problems. Recent adaptations enable Foretell to detect sets of genes that interact heavily with one another. We provide empirical results, and describe ongoing work on biological meaning and knowledge infusion from the user.

This paper investigates how readily an unsolvable constraint satisfaction problem can be reformulated so that it becomes solvable. We investigate small changes in the definitions of the problemís constraints, changes that alter neither the structure of its constraint graph nor the tightness of its constraints. Our results show that structured and unstructured problems respond differently to such changes, as do easy and difficult problems taken from the same problem class. Several plausible explanations for this behavior are discussed.