The overall strategy used to isolate and sequence thousands of Fundulus cDNAs was (1) generate a high quality unidirectional cDNA library, (2) normalize the library, (3) randomly pick colonies and amplify by PCR the cDNA within the vector, (4) sequence and identify PCR products, and (5) after approximately every 1,000 clones, subtract these from the normalized library and repeat steps 3–5. Details for all protocols are provided at http://crawford.rsmas.miami.edu/ and were used in the Comparative Functional Genomic course at Mount Desert Island, ME 2000.

We sequenced 46,433 ESTs and 40,043 of these are available in the dbEST database at NCBI (dbEST identification numbers: 23,480,307 to 23,515,306; 23,520,047 to 23,525,409 and 24,320,128 to 24,320,184) as of June 26, 2004. Sequences in FunnyBase are identified by a number series: unique sequence number, array number, plate number and well identifier (example: 23434_125_001_H04). The remaining 6,966 un-submitted sequences failed to meet one of the sequence quality parameters. Two criteria are used for defining "good" sequences: 1) >100 bp of sequence with Phred score >20 or 2) form an overlapping cluster with other sequences. Of the 19,937 sequences processed with the current version of FunnyBase, 17,893 (90%) passed one of these quality parameters. In earlier iterations of FunnyBase, visual inspection and later a sequencer-specific quality measure equivalent to a Phred score of 15 were used as filters resulting in 3,603 of the first 5,760 sequences (63%) and 13,922 of the next 15,168 sequences (92%) meeting quality standards, respectively. Re-sequences account for 5,668 sequences, and 4,625 of these were submitted.

One of the most important steps for producing microarrays from cDNA libraries is being able to associate the bacteria containing the cDNAs of interest with the EST annotation. High-throughput procedures are highly prone to tracking errors including: loading plates into an automatic sequencer in the wrong order, orienting symmetric plates in the wrong direction, or mislabeling of plates. The ability to identify these types of mistakes requires controls for identifying plates and plate orientation. The FunnyBase system has a number of integrated quality control steps. First, a Ctenophore cDNA (NCBI: accession number: CN992733) that is unlike anything else in GenBank is used as a control. Controls are placed in 96-well plates in wells corresponding to the plate number and two orientation wells (A5 and F9). Sequences from 96-well plates are automatically scanned for these controls so that the identity and orientation are confirmed and a report is generated for manual review. Secondly, all clones used for microarray production are re-sequenced. This is necessary because individual cDNAs are cherry picked, re-grown and re-amplified, and each of these steps has the potential to introduce or magnify an error. For example, for a 6,000 gene array, a 5% error rate would result in 300 incorrect clones. Re-sequenced array plates are compared using pair-wise BLAST 42 against previous sequencing results so that the identity of printed microarray spots are verified.

EST projects generate a number of redundant sequences due to the random selection of cDNAs from tissue libraries (Table 1). Clustering redundant sequences is a critical first step of analysis in order to identify genes to target for subtraction. The program CAP3 by Xiaong Huang 43 was used to cluster EST sequences with a 30 bp overlap and 75 percent similarity.

FunnyBase contains a total of 40,043 EST sequences from F. heteroclitus heart and liver. Clustering with CAP3 yields 3,776 clusters that contain 30,688 ESTs (77%). The remaining 8,991 ESTs (23%) are singletons. By storing the results of clustering with annotation, FunnyBase easily identifies these genes and aids in the selection of genes to be used for library subtraction with the goal of picking less common transcripts.

The 10 annotated clusters with the most sequences are listed in Table 1. In microarray experiments these genes tend to be highly expressed and the fluorescent signal tends to saturate the photomultiplier tube. These genes also serve to verify microarray printing because the predicted spots for these genes have the strongest signal.

The distribution of the number of clusters with two or more ESTs is depicted in Figure 2. Although most clusters (2,581 or 68%) in FunnyBase have two or three-to-four sequences (Fig. 2), a small number of highly expressed genes form clusters with a large number of ESTs. For example, there are three clusters that contain 512 to 1,024 ESTs and ten clusters that contain 256 to 512 ESTs (bottom two bars for killifish in Fig. 2). This distribution is similar to other teleost fish EST collections (Fig 2). Notice, as more ESTs are added, clusters tend to get larger (more ESTs per cluster) rather than new small clusters growing in frequency. Of the 3,779 killifish clusters, 14% have more than eight ESTs, yet of the 14,714 Medaka clusters, 48% have eight or more ESTs. These distributions suggest that adding more EST sequences has diminishing returns.

One of the objectives of EST projects is to isolate most, if not all, genes expressed in a tissue or organism. The increasing size of larger clusters with more sequencing efforts indicates that strategies to increase the probability of isolating new genes need to be employed. We used two strategies. First, we normalized the library to reduce the differences among expressed genes to less than 10-fold among rare and abundant mRNAs 4445. Using this technique we were able to reduce redundancy in annotated genes from 33% in the non-normalized library to 11% after normalization. Second, we targeted specific sequences for subtraction: annotated cDNAs with high frequencies were targeted in order to focus effort on picking new, rare sequences. Through subtraction we were able to increase the rate of discovery of new annotatable sequences from 24% to 36%. However, analysis of these results indicate that a set of highly expressed sequences, some of which were not subtracted because they were not in the set of annotated genes, still make up much of the EST library and should be the focus of future subtractions.

Of the 12,776 unique ESTs (3,776 clusters and 8,991 singletons), 3,877 (30%) were annotated. The distribution of e-values for these annotations is shown in Figure 3. Most (84%) of these ESTs have e-values less than 10^-5. Among the clusters, 2,265 of 3, 779 (60%) were annotated as compared to 333 of 1,131 (30%) of confirmed high quality singletons. The lower percent of annotated singletons suggests that these are either rare fish-specific genes, or represent otherwise divergent, likely non-coding (5' or 3' UTR), regions.

Annotations for ESTs are based on similarity using BLASTX or BLASTN 42 to sequences in one of six-public databases: Swiss-Prot, Human UniGene, Danio rerio UniGene, Oncorhynchus mykiss Unigene, Oryzias latipes Unigene, or GenBank NR. FunnyBase includes locally parsed copies of these databases in a relational format. Thus, all searches are done locally and annotation features beyond the FASTA description can be queried. Consensus sequences from the Fundulus EST clusters as well as high quality single unique sequences (singletons) are used as query sequences for BLAST searches against these public databases. The use of consensus sequences, when available, allows sequences that do not contain regions of significant similarity with known protein (e.g., 5' or 3' noncoding regions) to be annotated if they are members of an annotated cluster. All hits with e-value less than 0.001 and their associated alignments are stored in the database and tracked with any associated functional annotation. Users can specify a custom level of significance when assessing the validity of homology based annotation. This record, which goes beyond storing a certain number of 'best hits', is critical because in many cases additional results may have a negligibly lower alignment score, but provide much more useful functional data.

The use of multiple databases increases the total number of annotated ESTs (Fig. 4) as compared with any one source and provides opportunity to compare annotation between all three sources for 1,841 (47%) sequences. GenBank NR provided the most number of annotations, but these tend to be less informative (see systematic functional annotations below). Human Unigene provided an additional 311 (8% of total) annotations. SwissProt provided an additional 32 annotations (1% of total) with 743 fewer annotated sequences than the NR. However, SwissProt is uniformly well annotated as compared to NR where informative functional annotation can easily be buried by numerous uninformative hits at similar e-values. Besides increasing the number of annotations, comparing the annotations from multiple databases ensures that mistakes in the curation can be detected and information such as alternative gene names can be compiled from multiple sources.

In conjunction with performing similarity searches by BLAST, FunnyBase includes locally parsed representations of public databases such as SWISS-PROT in a relational database format. These databases provide additional information that cross-references other public resources such as GO, KEGG or OMIM 46 that are not available in the single FASTA description line returned by BLAST search.

KEGG 39 is a unique tool that represents metabolic and signal-transduction pathways both visually and computationally. FunnyBase links annotated genes to enzymes in KEGG pathways based on enzyme commission (EC) numbers. These pathway associations are stored and queries can readily identify genes from a given pathway that show specific patterns of expression. For visual inspection of the pathway, the web interface http://genomics.rsmas.miami.edu/funnybase/super_craw4/ links directly to the graphical KEGG pathways in which a gene occurs.

Of the 3,877 annotated ESTs in FunnyBase, 588 (14%) participate in one or more pathways defined by KEGG. These 588 ESTs represent 105 different pathways. Table 2 provides a breakdown of the number of ESTs in FunnyBase for the 10 pathways associated with the largest number of distinct sequences (contigs or singletons). The extent that a given pathway is represented in FunnyBase can be used to identify metabolic differences among tissues 47 or in different species.

The Gene Ontology project (GO) has produced a structured vocabulary in the form of an acyclic directed graph that biologists can use to annotate genes in a systematic manner 38. FunnyBase includes two non-trivial steps to make the best possible use of GO terms. First, many GO annotations are lost if only the single 'best hit' from a homology search is considered because GO annotation is applied most often to a few model species such as human that may not appear as the single 'best hit' in a list of BLAST results. FunnyBase identifies the gene name associated with the 'best hit' BLAST result and then uses all GO annotation associated with hits from the complete BLAST results that have the same gene name as the 'best hit' and an e-value of e < 10^-12. The goal of this approach is to identify annotation associated with a single 'best hit' gene based on results that may come from multiple species (orthologous genes) and therefore may have varying degrees of sequence similarity due to phylogenetic distance, but to avoid the problem of selecting an inconsistent set of GO terms arising from gene families that share regions of sequence similarity but may have different functions.

Secondly, GO annotation in public databases tends to annotate sequences with only the most specific GO term available, for example RNA polymerase II transcription factor activity, enhancer binding (GO:0003705) rather than the more general parent term transcription regulator activity (GO:0030528). However, in functional genomic analysis, significant patterns of expression may exist at the more general level of functional description. FunnyBase takes advantage of the connected parent-child relationship of GO terms provided by using the relational database version of GO available for download at http://www.geneontology.org to identify such relationships. These data are used to extract the tree of more general GO terms related to those provided by public databases. A FunnyBase script then re-annotates genes with this more complete set of GO terms.

Of the 3,877 annotated genes, 1,912 (54%) are assigned one or more GO terms with a total 6,728 GO assignments being made directly based on information in public databases such as SwissProt. Using parent-child GO term relationship backtracking, an additional 34,112 GO term assignments were made, resulting in a final count of GO assignments of 36,024 excluding the most general terms that divide GO into three categories. Thus, on average, 19 GO terms are assigned to each of 1,912 annotated genes.

Humans have approximately 30,000 expressed genes, yet there are over 1,000,000 human UniGenes (NCBI). Clearly, these clusters of cDNAs greatly overestimate the number of unique genes. Similarly, FunnyBase has multiple clusters for the same gene: 15 apolipoprotein I, 10 cytochrome oxidase I, and 53 vitellogenin clusters. To provide a more precise estimate of the number of unique genes, consensus sequences were queried against the 27,695 sequences from the Human RefSeq 48 database, then grouped by identical gene symbol. Of the 2,376 Fundulus clusters that were similar to a sequence in Human RefSeq (e-value < 10^-10), 1,818 (76%) had distinct gene annotations. This method of clustering clusters by similarity to well-annotated reference sequences provides a method to more accurately define the number of unique genes represented by an EST set.

As a case study in how functional annotation in FunnyBase can be integrated with microarray data in a rigorous manner, we used a data set based on a microarray of metabolic genes printed from ESTs annotated in FunnyBase 47. Statistical analysis of this set of 363 metabolic genes identified a set of 62 genes that showed statistically significant regression between gene expression levels and temperature along the Atlantic coast. That is, among individuals collected from different locations along the thermocline and then acclimated to common physiological conditions for at least nine months before analysis, 17% of the metabolic genes had a linear relationship between the amount of mRNA and the environmental temperature these animals evolved in. Our hypothesis was that this set of 59 genes represents a functionally different set than those genes that do not show regression with temperature. To test this hypothesis we examined the frequency of genes annotated with a given GO term in the statistically significant gene set versus the non-significant genes. Figure 5 shows the relevant proportions in each set for GO terms that are represented by 5 or more ESTs in the significant set. For example, the GO term Amine Metabolism (GO:0009308) is assigned to 14% of the 62 statistically significant genes but only 3% of the non-significant genes (those that do not show significant regression with temperature). A Fisher-exact test indicates these frequencies (14% vs. 3%) represent different underlying distributions (p < 0.001). Specifically, genes involved in amine metabolism are overrepresented in the set of genes that show regression with temperature as compared with the remaining sequences. This significant increase is found for two other non-mutually exclusive GO terms: amino acid and derivative metabolism, and amino acid metabolism (p < 0.05). Other GO terms show a reverse trend although none were statistically significant. For example, ion transport (p = 0.08) and cell growth (p = 0.16) had few genes with a clinal variation in expression. These data suggest that the functions of genes influence whether they are affected by ecologically interesting patterns of expression (Fig 5).

The web interface http://genomics.rsmas.miami.edu/funnybase/super_craw4/ provides public access to the FunnyBase system and dataset. Searches can query by keyword in annotation, gene name, GO term, metabolic pathway, clone or plate id, and BLAST homology search. All data including raw sequences, cluster memberships, cluster alignments, and alignments with homologous sequences are provided for the user to examine the source of annotations. Links associated with each annotation are made to external resources such as GO's AMIGO browser, KEGG pathways, SwissProt, and NCBI records.

FunnyBase was constructed to annotate sequences for the analysis of gene expression. It provides identification and annotation for genes in the Crawford laboratory with a primary goal of identifying clones useful for the construction of microarrays. As such, other Fundulus sequences in Genbank are not included. However, FunnyBase forms the basis of the TIGR Killifish gene index http://www.tigr.org/tigr-scripts/tgi/T_index.cgi?species=killifish that includes publicly available F. heteroclitus sequences.

The animal protocols used in the present study have been approved by the University of Miami Institute Animal Care and Use Committee. The teleost fish Fundulus heteroclitus used for ESTs were collected from two sites: Scorton Creek in Sandwich, MA, and Stone Harbor, NJ. These populations are in the central portion of the thermal cline and have relatively high levels of heterozygosity 32. These fish were subjected to the following environmental regime before tissues were harvested for mRNA extraction: kept in controlled temperature and aeration conditions, and acclimated to common conditions (20°C, 15 ppt salinity) in re-circulating aquaria for at least nine months before experiments. Following this common acclimation a subset of fish were subjected to one of several stresses: 4°C, 34°C, hypoxia, or a complex mix of hydrocarbons.

To effectively isolate and sequence thousands of cDNAs for the production of microarrays, a unidirectional cDNA library with few non-recombinants was required. We created four cDNA libraries: heart libraries from non-stressed and stressed fish and liver libraries from non-stressed and stressed fish. The non-stressed F. heteroclitus cardiac and liver libraries were provided by Drs. S. Karchner and M. Hahn, WHOI 58 and were constructed using the UniZap λ cDNA Gigpack Gold cloning kit (Stratagene, La Jolla, CA, USA). The cardiac library was produced from 27 fish hearts (both sexes) sampled from Scorton Creek in Sandwich, MA. The cDNAs in these libraries are oriented such that the 5'end of each cDNA is ligated to EcoR1 and 3' poly A is ligated to XhoI. These libraries had less than 1% non-recombinants, i.e. 2 of 300 random clones from a non-normalized library had no inserts. The stressed libraries included 4 fish subjected to the four stressors (above) and 4 non-stressed individuals. Unidirectional heart and liver libraries were constructed such that the 5'end of each cDNA is ligated to EcoR1 and 3' poly A is ligated to XhoI of the plasmid vector pSmart (Lucigen, Middleton, WI, USA). The pSmart-cDNA vector was designed for EST work. The vector expresses kanamycin-resistance and has a terminator on both sides of the cDNA insertion site preventing expression of cDNA. These two attributes (non-expression and Kan-resistance) increase the stability of different genes in the library versus cDNA libraries in Amp libraries with Lac promoters (Crawford, unpublished). These libraries had less than 1% non-recombinants.

Normalization of cDNA libraries reduces the differences among expressed genes to less than 10-fold among rare and abundant mRNAs 4445. Normalized libraries were produced by isolating cDNAs from approximately 10¹² plasmids. The cDNAs were isolated using PCR amplification with vector specific primers immediately 5' and 3' to the insertion site (EcoRI and XhoI sites). These PCR products (PCR-cDNAs) were denatured and hybridized to single stranded plasmids from the cardiac cDNA library. Taking advantage of Cot values, the most abundant cDNAs were annealed to the more abundant PCR products and were removed selectively by hydroxyapatite-column chromatography. The single-stranded plasmids in the flow-through were converted to double strands using the Sequenase DNA polymerase (Amersham, Piscataway, NJ, USA). DH10s E. coli (BRL) were transformed with these double-stranded plasmids by electroporation. The number of recovered plasmids and the resulting complexity of the normalized library depended on the duration of hybridization or Cot values. Two normalized libraries were made using either a 12 or 24 hour hybridization. The library from the 12-hour hybridization yielded 250,000 plasmids. The library from the 24 hour hybridization yielded 3,000 plasmids and had a greater representation of rare mRNAs and greater frequency of non-recombinants.

Characterization of cDNAs (growth of individual bacterial colonies containing plasmids, PCRs, purification of PCR products, sequencing reactions) used 96 well plates and octopipettes. To characterize cDNAs, 96 individual bacterial colonies from the normalized library were randomly chosen, and each was grown in 1.25 ml of Superbroth in 2 ml-96 well plates. After 18 hours of growth, two 250 ul bacterial glycerol stocks were made and stored in 96 well plates at -80°C. One microliter of these bacterial growths was used for PCR reactions using forward and reverse plasmid specific primers: (PucF = CGCCAGGGTTTTCCCAGTCACG, PucR = GAGCGGATAACAATTTCACACAGGAAA). PCR reactions had 0.2 mM dNTPs, 10 pmoles of each primer, 1 unit of Promega Taq (0.2 ul), and reaction buffer with detergents and DMSO (final concentrations: 50 mM Tris HCl, pH 9.2 (25°C), 16 mM (NH₄)2SO₄, 2.25 mM MgCl₂, 2% (v/v) DMSO, 0.1% (v/v) Tween 20). Two-step thermal cycle conditions were used (94°C for 10 seconds; then 32 cycles of 94°C for 30 seconds followed by 70°C for 5 minutes; then 72°C for 15 minutes). PCR products were purified manually in 96 well format using Sephadex G-50 in a deep well plate with a 0.2 microfilter (Millipore, Billerica, USA) or robotically using AmPure (Agencourt, Beverly, MA, USA) and EvolP³ 96 pipetting liquid handling system (PerkinElmer Life Sciences Inc., Boston, MA, USA).

PCR products were sequenced from the 5' end (relative to the mRNA) on an ABI 373 or ABI 3730 sequencer using ABI "Big Dye" reaction mix. We typically used 1/16 the amount of reaction mix, yielding 300 to 400 unambiguous bases. Sequences were purified using biotin primers and streptavidin coated magnetic beads (for the ABI 373) or Agencort CleanSeq (for the ABI 3730).

We used three procedures to verify that the correct sequence was associated with each cDNA. 1) Each 96-well plate had three wells with a "marker cDNA" (Ctenophore cDNA #5, a random cDNA with no similarity to any sequence in GenBank). Two wells (#40 and #67) always contained the marker cDNA, and thus any misloading or mislabeling of sequencing lanes was identifiable. The third marker cDNA was placed in a well that corresponds to the plate number (e.g., plate 2 had the marker in well 2). 2) After the production of 12 plates, one row (8 wells) from each plate was re-sequenced. Thus, 8/96 or ~8% of all sequences and their locations were confirmed. 3) cDNAs used for microarrays were re-sequenced. These measures are important to ensure that the correct and known cDNAs are printed.

The complexity of the normalized library was reduced by subtracting the characterized cDNAs previously isolated from the normalized library. Subtraction greatly reduced the probability of isolating the same cDNA and thus improved the efficiency of screening the library for unique clones. Subtraction used a 100-fold molar excess of biotin-labeled antisense cDNAs produced by PCR using all the characterized cDNAs as substrates and vector-specific primers in which the 3' primer was labeled with biotin. These PCR products were hybridized to the cDNA libraries in the presence of oligo-dA and vector-specific oligos (that prevented non-specific hybridization to oligo-dT or vector sequences). After a 24 hour hybridization, genes in the library that bound to these biotin-labeled PCR products were removed with the use of magnetized, streptavidin coated beads. DH10s E. coli were transformed with the subtracted library by electroporation.

Computational work was done on an Apple G5 dual 2 GHz processor system with 4 GB of RAM. Data are stored in a MySQL database, perl scripts were used extensively for parsing and loading data, and PHP was used on an APACHE web server to construct the user interface. Additional programs available from their authors are mentioned within context. Software and databases are described in Table 3.

Microarrays were printed using a select 384 cDNAs from F. heteroclitus cardiac library encoding essential proteins for cellular metabolism isolated from over 40,000 expressed sequences http://genomics.rsmas.miami.edu/FunnyBase/super_craw4/. These 384 cDNAs were amplified with amine-linked primers and printed on 3-D Link Activated slides (Surmodics Inc., Eden Prairie, MN, USA) using GeneMachine OminGrider, and blocked following slide manufacturer protocols. The suite of 384 amplified cDNAs was printed as a group in four spatially separated replicates. Four hybridization zones of these four replicate arrays were printed per slide, with each zone set separated by a hydrophobic barrier. Samples were hybridized twice; once with Cy3 and once with Cy5 resulting in overall technical replication of 8-fold per sample.

RNA was extracted from tissue homogenate in a chaotropic buffer using phenol/cholorform/isoamyl alcohol and RNA quality was assessed using the Agilent 2100 Bioanalyzer (Agilent Technologies, Palo Alto, CA, USA). RNA for hybridization was prepared by amplification using a modified Eberwine protocol 59 using the Ambion Amino Allyl MessageAmp aRNA Kit. Cy3 and Cy5 were hybridized to slides and incubated 12–18 hours at 42°C. Following hybridization, slides were scanned using the Packard Bioscience ScanArray Express microarray scanner (PerkinElmer Life Sciences Inc., Boston, MA, USA) and images processed using ImaGene (Biodiscovery Inc., Marina del Rey, CA, USA).