A common task in biological research is to predict function for proteins by comparing sequences between proteins of known and unknown function. This is often done using pair-wise sequence alignment algorithms (e.g. BLAST). A problem with this approach is the assumption of a simple equivalence between a minimum sequence similarity threshold and the function similarity between proteins. This assumption is based on the binary concept of homology in that proteins are or not homologous. The relationship between sequence and function however is more complex as well as pertinent for predicting protein function, e.g. evaluating BLAST alignments or developing training sets for profile models based on functional rather than homologous groupings. Our motivation for this study was to model sequence and function similarity between proteins to gain insights into the "sequence-function similarity relationship between proteins for predicting function. Using our model we found that function similarity generally increases with sequence similarity but with a high degree of variability. This result has implications for pair-wise approaches in that it appears sequence similarity must be very high to ensure high function similarity. Profile models which enable higher sensitivity are a potential solution. However, multiple sequences alignments (a necessary prerequisite) are a problem in that current algorithms have difficulty aligning sequences with very low sequence similarity, which is common in our data set, or are intractable for high numbers of sequences. Given the importance of predicting protein function and the need for multiple sequence alignments, algorithms for accomplishing this task should be further refined and developed.

MOTIVATION: Enrichment tests are used in high-throughput experimentation to measure the association between gene or protein expression and membership in groups or pathways. The Fisher's exact test is commonly used. We specifically examined the associations produced by the Fisher test between protein identification by mass spectrometry discovery proteomics, and their Gene Ontology (GO) term assignments in a large yeast dataset. We found that direct application of the Fisher test is misleading in proteomics due to the bias in mass spectrometry to preferentially identify proteins based on their biochemical properties. False inference about associations can be made if this bias is not corrected. Our method adjusts Fisher tests for these biases and produces associations more directly attributable to protein expression rather than experimental bias. RESULTS: Using logistic regression, we modeled the association between protein identification and GO term assignments while adjusting for identification bias in mass spectrometry. The model accounts for five biochemical properties of peptides: (i) hydrophobicity, (ii) molecular weight, (iii) transfer energy, (iv) beta turn frequency and (v) isoelectric point. The model was fit on 181 060 peptides from 2678 proteins identified in 24 yeast proteomics datasets with a 1% false discovery rate. In analyzing the association between protein identification and their GO term assignments, we found that 25% (134 out of 544) of Fisher tests that showed significant association (q-value ≤0.05) were non-significant after adjustment using our model. Simulations generating yeast protein sets enriched for identification propensity show that unadjusted enrichment tests were biased while our approach worked well.

Mediator-based data integration systems resolve exploratory queries by joining data elements across sources. In the presence of uncertainties, such multiple expansions can quickly lead to spurious connections and incorrect results. The BioRank project investigates formalisms for modeling uncertainty during scientific data integration and for ranking uncertain query results. Our motivating application is protein function prediction. In this paper we show that: (i) explicit modeling of uncertainties as probabilities increases our ability to predict less-known or previously unknown functions (though it does not improve predicting the well-known). This suggests that probabilistic uncertainty models offer utility for scientific knowledge discovery; (ii) small perturbations in the input probabilities tend to produce only minor changes in the quality of our result rankings. This suggests that our methods are robust against slight variations in the way uncertainties are transformed into probabilities; and (iii) several techniques allow us to evaluate our probabilistic rankings efficiently. This suggests that probabilistic query evaluation is not as hard for real-world problems as theory indicates.

There is a significant need for data integration capabilities in the scientific domain, which has manifested itself as products in the commercial world as well as academia. However, in our experiences in dealing with biological data it has become apparent to us that existing data integration products do not handle uncertainties in the data very well. This leads to systems that often produce an explosion of less relevant answers which subsequently leads to a loss of more relevant answers by overloading the user. How to incorporate functionality into data integration systems to properly handle uncertainties and make results more useful has become an important research question. In this paper we describe an enhanced general-purpose data integration system which incorporates uncertainty metrics within a formal probabilistic framework. Additionally, for evaluation purposes, we have implemented a use case scenario which utilizes biological data sources and performed a study which provides validation of system query results.

Genomic medicine aims to revolutionize health care by applying our growing understanding of the molecular basis of disease. Research in this arena is data intensive, which means data sets are large and highly heterogeneous. To create knowledge from data, researchers must integrate these large and diverse data sets. This presents daunting informatic challenges such as representation of data that is suitable for computational inference (knowledge representation), and linking heterogeneous data sets (data integration). Fortunately, many of these challenges can be classified as data integration problems, and technologies exist in the area of data integration that may be applied to these challenges. In this paper, we discuss the opportunities of genomic medicine as well as identify the informatics challenges in this domain. We also review concepts and methodologies in the field of data integration. These data integration concepts and methodologies are then aligned with informatics challenges in genomic medicine and presented as potential solutions. We conclude this paper with challenges still not addressed in genomic medicine and gaps that remain in data integration research to facilitate genomic medicine.