Beyond tissueInfo: functional prediction using tissue expression profile similarity searches

Abstract

We present and validate tissue expression profile similarity searches (TEPSS), a computational approach to identify transcripts that share similar tissue expression profiles to one or more transcripts in a group of interest. We evaluated TEPSS for its ability to discriminate between pairs of transcripts coding for interacting proteins and non-interacting pairs. We found that ordering protein-protein pairs by TEPSS score produces sets significantly enriched in reported pairs of interacting proteins [interacting versus non-interacting pairs, Odds-ratio (OR) = 157.57, 95% confidence interval (CI) (36.81-375.51) at 1% coverage, employing a large dataset of about 50 000 human protein interactions]. When used with multiple transcripts as input, we find that TEPSS can predict non-obvious members of the cytosolic ribosome. We used TEPSS to predict S-nitrosylation (SNO) protein targets from a set of brain proteins that undergo SNO upon exposure to physiological levels of S-nitrosoglutathione in vitro. While some of the top TEPSS predictions have been validated independently, several of the strongest SNO TEPSS predictions await experimental validation. Our data indicate that TEPSS is an effective and flexible approach to functional prediction. Since the approach does not use sequence similarity, we expect that TEPSS will be useful for various gene discovery applications. TEPSS programs and data are distributed at http://icb.med.cornell.edu/crt/tepss/index.xml.

Figure 1.

Overview of the TEPSS approach. EST counts are obtained from dbEST using TissueInfo for each transcript of an organism (TiDumpCount box). The program TiSimilarity performs TEPSS for a query transcript, with a scorer and count information. The result is a ranked list of transcripts ordered by tissue expression profile similarity to the query transcript.

Figure 2.

Minimum evidence scorer. This plot illustrates how scores are calculated by the minimum evidence scorer. This example shows two transcripts with EST counts in six tissues (in practice, the TEPSS scorer uses more than 100 tissues). Tissues are denoted by the index τ. Values E(t1, t2, τ) are the evidence scores defined in Supplementary Equation 3, i.e. E(t1,t2, τ) = E(counts(tcd, t1, τ), counts(tcd,t2, τ)). The example on the left shows two transcripts that yield a negative TEPSS score (−10), while the example on the right shows two transcripts with a positive score (+54).

Figure 3.

TEPSS scores for interacting and non-interacting pairs. Scores are calculated with the confidence scorer (see Supplementary Material section). (a) Distribution of scores for interacting protein pairs in the complete human PIN. (b) Distribution of scores for interacting protein pairs in the human PIN supported by four or more experimental methods. The difference between the distribution of the score for reported interacting protein pairs and non-interacting protein pairs is statistically significant in both cases (Wilcoxon rank sum test, two tailed). The difference between the distribution of the TEPSS confidence scores for interacting protein pairs and the distribution of TEPSS confidence scores of randomly shuffled counts is also significant in both cases (Wilcoxon rank sum test, two tailed).

Figure 4.

TEPSS scores for metabolic pathways. Distribution of TEPSS scores for 5000 pairs of genes coding for enzymes contributing to the same metabolic pathway in human. The differences with respect to that of enzymes contributing to different pathways is significant (Wilcoxon rank sum test, two tailed). The difference with respect to the distribution of scores of randomly shuffled counts is also significant.

Figure 5.

Screening for ribosomal transcripts. We used ribosomal transcripts as input to a whole human transcriptome TEPSS search. The plot shows a lift curve, constructed by leaving out one ribosomal gene of the input, and searching the genome with the rest of the ribosomal transcripts. The x-axis indicates at what relative rank in the genome the gene that was left out was found in the TEPSS output (fraction of total human transcripts represented in TissueInfo, or the proportion of the genome that needs to be inspected to find the left out gene). The y-axis indicates the proportion of ribosomal transcripts that would be found at the rank observed for the transcript (% reported_predictions). The diagonal of the plot represents the expected rate of random prediction. Dotted lines illustrate results obtained when random sets of transcripts are used as input with different scorers.

Figure 6.

Screening for SNO protein targets. This lift curve illustrates that the TEPSS approach can effectively predict SNO targets in the human transcriptome.