Access to automated DNA sequencing technology has made possible the rapid generation and analysis of gene transcripts expressed in organisms via expressed sequence tags (ESTs) [1,2,3,4,5]. This information has helped to identify those genes expressed in particular stages of development and in specialized tissues or organs [6,7,8,9,10]. Novel gene products and target leads for therapeutic intervention can also be gleaned rapidly from ESTs [1,2,11]. A more detailed understanding of the molecular interactions between symbionts, whether pathogenic or mutualistic [12], is also possible with this approach [13,14,15,16,17].

For a sequence isolated from interacting symbionts, determining its cellular role (or roles) is complicated by not knowing which species expressed the sequence [18]. We refer to this challenge as 'the problem': given a sequence x expressed in an interaction between species A and B, did x originate from A or B? Various solutions are readily conceived, each with merits and faults. Here, we show that a comparative lexical analysis of word counts (specifically, hexamer frequencies), previously used to detect library contamination in sequencing projects [19], provides a powerful computational basis to infer a transcript's species of origin.

Experimentally, one can attempt to solve the problem by hybridizing a clone (as probe) to genomic DNA (target) from both species and determining to which target the probe hybridizes. This approach can produce very reliable results. However, if a sequence is highly conserved in the two taxa, hybridization stringency conditions can influence the outcome considerably. For high-throughput EST sequence analysis, source verification by hybridization is impractical in terms of time and reagents. As an alternative to in vitro hybridization, several computational solutions are possible.

Were the genome sequence of both species completely determined, one could simply use sequence similarity searching [20,21,22]. However, most plant hosts and their microbial symbionts have little or no genomic sequence data available, which makes this approach very unreliable. Strong similarity to a sequence from one organism does not preclude the possibility that a similar sequence is present in the other species. Conclusions based upon such partial knowledge have been informative, but are potentially misleading [18,23].

Codon usage varies across taxa [24,25,26]. Exploiting this fact may seem a viable solution to the problem, as it has proven suitable for predicting the presence of introns among exons in genomic DNA. However, it really is not practical, because of the need to know the reading frame for translation of a messenger RNA into an amino acid. EST data are of notoriously unreliable quality, sometimes having a large proportion of ambiguous bases, and sometimes having single base-pair insertions or deletions, which disrupt a reading frame. Word counting is less prone to these sources of error, and uses information intrinsic to biases in codon usage by counting codon pairs as hexamers in a sliding window, whereas codons are read in non-overlapping, tiled windows.

An intuitive approach to the problem that examines sequence composition is to compare the guanine and cytosine (GC) base content of a sequence with other sequences from the species being studied. When two species' genomes have different GC content, this method can be very useful. In a recent investigation, for instance, sequences from the stramenopile plant pathogen Phytophthora sojae and its soybean (Glycine max) host showed a 20% difference in mean GC content [18]. The origin of a number of sequences could readily be identified this way, but a large proportion could not, because of considerable overlap in the distributions' tails. Counting frequencies of GC is simple word counting, where the word size k is 1/2: only two semi-words, G/C and A/T are counted.

An alternative approach to determining the origin of a sequence is suggested by previous work on analysis of word counts, or k-tuple frequencies, which was intended as a means of evaluating a library for contamination when sequencing from a single model organism [19]. The word-counting method provides distinct advantages over other computational methods. Unlike sequence-similarity searching, it does not require that the full protein-coding content of both genomes be known for reasonable inferences to be made. Further, word counting is sensitive to biases in codon usage and GC content commonly observed when comparing taxa, but does not require knowledge of the reading frame for amino-acid translation. That is, the underlying differences between the two organisms that result in base composition or codon usage biases can also be detected by counting words. Unlike GC analysis, lexical analysis establishes a clear threshold above or below which we can infer the species of origin, and a confidence level for an inference can readily be assigned. Dunning's likelihood-ratio test of word dissimilarities [27] also has the appealing property of being non-parametric, having no assumption of normality for the underlying frequency distribution, which makes it statistically powerful [28]. Dunning [27] demonstrated that unreliable results can be obtained from parametric tests, such as Χ², particularly in such cases as lexical analysis.

In the experiments detailed below, we first validate the word-counting method on sequences whose origin and function are known, then compare it with ability to diagnose the origin of sequences with distributions of GC content. We examine sequences from pathogenic interactions between species from the genus Phytophthora and the plant hosts G. max and Medicago truncatula, then apply the word-counting approach to sequences from two microbial mutualists in association with M. truncatula, the arbuscular mycorrhizal zygomycete Glomus versiforme, and the nitrogen-fixing bacterium Sinorhizobium meliloti.

Validation sequence accession numbers, gene names, and comparison results appear in Table 1. Incorrect inferences are underlined. The word-counting method was generally quite reliable when tested against sequences of known origin, being wrong in 3 cases out of 50; a phosphate transporter from G. versiforme and two in planta-induced genes from Phytophthora infestans were misidentified as plant sequences. This indicates a failure rate of 6% - all false negatives under the null hypothesis that a transcript originates from the plant host. Performance of the method was not influenced by whether the isolated source of a sequence was an mRNA or DNA molecule, as indicated by the column labeled 'mRNA?'.

Distributions of GC content are approximately normal in two of three cases studied, those of axenic P. sojae cultures (Figure 1). For sequences from infected plant cultures, a bimodal distribution is apparent. Roughly 25% of a total of 927 infected G. max sequences contain less than 50% GC; most of these are likely to be plant transcripts [18]. This is a considerably greater number than for axenic P. sojae cultures, in which fewer than 5% of mycelia and zoospore isolates contain less than 50% GC.

Several properties of cumulative distribution functions warrant comment, to help explain similar plots from word dissimilarity comparisons (Figures 1b,2a). The median of a distribution occurs where the function reaches a cumulative probability of 0.5. Medians from all three P. sojae libraries are similar, varying by less than 4% GC (Figure 1b). Other moments of the distributions are readily apparent; the variance is inversely related to the slope at the median value of the function. A useful property of cumulative distribution functions is that any point on the y axis gives the integrated area (cumulative probability) under the curve. We use this property to establish experiment-wide false-positive and false-negative rates (Figure 2a). In this case, α = 0.088 and β = 0.032.

Calibration curves from hexamer dissimilarity tests, shown in Figure 2b as solid black lines for plant and dashed black lines for stramenopile training sequences are approximately normal. The medians differ considerably, with only about 10% percent overlap in the two distributions' tails about the neutral t-value of zero. Superimposed are comparison curves from P. sojae test sets (Figure 2b), which parallel the GC content curves in Figure 1b but show slightly less variance. Axenic sequences are clearly more like stramenopiles (B₁) than plants (A₁) in hexamer composition, with all but a small percentage having positive t values. Plant-like sequences are as abundant in the mixed library as detected by GC content, about 23%. As expected, the two methods agree, having positively correlated values for GC and t (r² = 0.852, P < 10^-16, v = 2,641).

Looking in more detail at the paired dissimilarity values (Figure 3), we can see which individual sequences are more or less like plant and pathogen. The magnitudes of dissimilarity are also apparent, with longer sequences having larger dissimilarity values. BLASTX similarity searches against the protein sequences in nr, a non-redundant library of proteins [29,30,31] revealed that none of the 12 plant-like mycelial transcripts significantly resemble known proteins (E > 10^-4). Among the top ten most plant-like transcripts from the infected G. max library, three had no significant matches, four matched putative Arabidopsis thaliana proteins, and three matched known G. max proteins: cytochrome P450 (accession AF022460, E < 10^-34), methylglyoxalase (accession P46417, E < 10^-34), and a ripening related protein (accession AF127110, E < 10^-71). Thus, the majority of the most plant-like transcripts in the infected soybean library strongly resemble characterized plant sequences. Analysis results from all P. sojae and mixed-culture transcripts are available as additional data files, grouped by the library from which transcripts were sequenced.

Figure 4 shows that calibration curves from comparing plant and microbial symbiont training sets have good separation and minimal overlap (about 10%) in two of three cases, but not for training set B₂, comprised of zygomycetes and chytridiomycetes, which overlaps considerably with plants (Figure 4b). The associated error rates are α = 0.126 and β = 0.207. When comparing between plants and bacteria, the error rates are α = 0.052 and β = 0.084, much lower than when comparing plants (A₂, Medicago) with fungi (B₂, zygomycetes and chytridiomycetes). Error rates for comparing stramenopiles and P. infestans ESTs with plants are as in Figure 2 (α = 0.088, β = 0.032).

Also shown in Figure 4 are cumulative distributions from comparisons with M. truncatula and microbial symbionts. All resemble calibration curves from plant sequences, having similar medians and slightly less variance than the plant calibration curves. Comparison curves show that the great majority of test sequences are more plant-like than otherwise, with 20% or less resembling microbial symbionts more closely than plants. A greater proportion of microbial sequences is present in the M. truncatula-G. versiforme interaction library (20%, Figure 4b) than in the P. medicaginis-infected M. truncatula library (5%, Figure 4a). However, Long's root-hair enriched library (MtRHE) [6] had a greater proportion of putative microbial sequences present (7% and 25%) than any of the libraries isolated from symbiont-associated cultures. The axenic and nodulating root libraries had the smallest portion of putative microbial transcripts (< 2%, Figure 4c), with the axenic library closely resembling nodulating root libraries. The method of preparing a library can affect the proportion of plant and non-plant sequences, as discussed below.

Paired dissimilarity values in Figure 5 show in greater detail which sequences are more or less like plant and symbiont. Sequences from an interaction library and pure plant root cultures appear together for comparison. Considerable variation in the degree of dissimilarity to both training sets is clear, largely due to variation in the length of sequences within test sets. Consistent with the cumulative distributions of D(A) – D(B) in Figure 4, most sequences lie above the identity function, and resemble the plant host more closely than the microbial symbiont. Mycorrhizal test sequences are more difficult to differentiate than sequences from the rhizobacterial or pathogenic associations, as seen by the diminished variation about the identity function in mycorrhizal comparisons (Figure 5b), contrasted with comparisons from pathogen-infected and nodulating root libraries (Figures 5a and c, respectively). Analysis results from all M. truncatula and mixed-culture transcripts are available as additional data files, grouped by the library from which transcripts were sequenced, and sorted from the least plant-like transcripts to the most plant-like.

Clearly, the word-counting approach provides a reliable solution to the problem of source identification with known confidence, and has several significant advantages. The reliability of the method is best justified in terms of the favorable validation test results, and is further corroborated by agreement with an analysis of GC content. In test cases where the correct answer is known a priori, results were correct within error rates expected from overlap in training sets. (Recall that α = 0.088 for comparisons between plants and stramenopiles, and α = 0.052 for comparisons between plants and bacteria.) Unlike GC content, the problem is clearly resolved by word counting with a threshold value of t = 0, and with statistical rigor, because false-positive and false-negative rates for a set of comparisons are readily computed from cumulative distributions of dissimilarity between two training sets. Optimal statistical power (minimal false-negative rate) is ensured when using a likelihood-ratio test statistic, as demonstrated by the Pearson-Neyman Theorem [28]. Further, word counting need not be trained only for the species being compared. Rather, it is sufficient that the training set be related to, but not necessarily congeners of, the species from which sequences are being compared. Sequences from several species of the genus Phytophthora were correctly distinguished from plant and bacterial sequences, and three genes from Agrobacterium tumefaciens were correctly identified as representing a bacterial sequence.

However, several caveats warrant prudence. Transcribed sequences that do not encode proteins, but rather catalytic single-stranded RNAs such as transfer and ribosomal RNAs [32], should be treated independently because they are more highly conserved across taxa than messenger RNAs. Also, filtering or trimming of low-complexity repeat regions, such as poly(A) or poly(T) tracts, is helpful because comparison results can be influenced by the abundance of a single hexamer. Early in our investigations, using one set of training sequences obtained from directionally cloned P. infestans cDNAs produced results that were difficult to interpret. It eventually became clear that, as the P. infestans sequences were all single-pass reads from the 5' end of a clone generated with the T3 primer, few sequences complementary to the 3' end of the mRNA sequence were present in the training set. This meant that the hexamer AAAAAA was common, but the hexamer TTTTTT scarce. Large amounts of the poly(T) hexamer would be expected when sequencing reverse complements of mRNAs obtained from 3' sequences generated with the T7 primer. Both poly(A) and poly(T) regions were present among plant training sequences. As a result, any sequence that contained a poly(T) tract tended to resemble the plant sequences. Further, because the error rates for an inference depend on the degree to which calibration curves overlap, the best results are obtained where overlap is minimal. Despite these caveats, word counting presents a viable solution to the problem.

The P. sojae-infected G. max library provides a clear example of contrast in both hexamer composition and GC content, resulting in readily diagnosed origins. Not every case is this simple. For clear separation between the two species to appear, the two must differ in composition and a detectable proportion of transcripts from each species must be present in the library. To be detectable, the proportion of transcripts present from a particular species must be greater than the error rate obtained from calibration curves.

Though these criteria are true for the infected G. max library (t < 0 for <25% of 927 transcripts), they do not appear to be true for the M. truncatula libraries we analyzed (t < 0 for 80–99% of 890–3,017 transcripts). In the P. medicaginis interaction library, we might expect the same bimodal distribution as seen with P. sojae. However, the two libraries were prepared in different ways. The P. sojae-infected library was prepared two days after infection, using a susceptible plant host strain, so as to maximize the number of pathogen transcripts present in the host tissue [18]. Further, G. max hypocotyl tissues were infected directly with a zoospore suspension. In contrast, the P. medicaginis-infected library was prepared ten days after infection and individual plants varied in their degree of susceptibility (C. Vance, unpublished data). Plants were also inoculated in a different manner: ground mycelia were dissolved in sterile water and incubated, and the resulting inoculum was pipetted onto the soil surface, rather than the plant. These differences in how tissues were cultured prior to library preparation could have produced the disparate abundance of plant transcripts, though both libraries were prepared from plant tissues infected with Phytophthora.

For mycorrhizal root libraries, we might explain the relative lack of symbiont sequences as resulting simply from a relative lack of transcripts in the host tissue. Most of the biomass in mycorrhizal roots is plant biomass [33]. We might therefore expect that most of the transcripts therein originate from the plant host. Confounding this result, the error rates in this comparison are the greatest among all the comparisons we performed, most likely because the evolutionary distance between fungi (zygomycetes and chytridiomycetes) and plants is the least among comparisons [34]. Also, zygomycete protein-coding sequences are rare in GenBank, which resulted in a small training set for these fungi, and may have amplified any biases. The high false-negative rate probably led to a failure to detect some symbiont transcripts.

In nodulating root libraries, we do not expect to observe an abundance of bacterial transcripts, because bacteria generally do not form polyadenylated mRNAs [35]. As the protocols used to extract and purify mRNAs from tissue lysate for the libraries cited in this study all relied on the presence of polyadenylation sites, we generally do not expect to find bacterial transcripts.

The abundance of putative microbial symbiont transcripts among sequences from a pure plant root library is difficult to interpret. The predicted portion of microbial transcripts was greater in the axenic root-hair enriched library than in mixed cultures. Error rates were greatest for comparisons between training sets from plant and pooled zygomycete and chytridiomycete sequences. Other than providing an 87% confidence level, the 13% false-positive rate does not completely explain why about 15% of root-hair enriched transcripts resemble fungal hexamer composition more closely than plants, and warrants further study.

Care had been taken to avoid contaminating plant tissue cultures by culturing seedlings in covered plates. Because of concern that ethylene accumulation in covered plates could improperly stimulate nodulation-related gene expression, seedlings were treated with Ag₂SO₄, an inhibitor of the plants' response to ethylene [6]. Inhibition of the ethylene response could have resulted in synthesis of transcripts that are uncharacteristic of plant roots. Analysis of another axenic root-hair enriched library, particularly one provided a carbon source to identify potential contaminants, and not treated with an inhibitor of ethylene response, would be an informative test.

These observations warrant further experimental scrutiny. The transcripts identified as most and least like plant or symbiont might also be studied in more detail as candidate participants in symbiosis. Symbiotic interactions, whether pathogenic or mutualistic, present novel challenges to both plant hosts and the biologists who study them. Computational approaches, in concert with experimental verification, can help resolve these challenges.

The following files are available for download:

countGC.pl: PERL script used to compute GC content of sequences analyzed.

hyb2dis.txt: patch file that converts White's hybridize program to compute individual dissimilarity values.

Training sets (GlycineMedicago.txt,Rhizobia.txt, Stramenopiles.txt, ZygoChytrid.txt): FASTA-formatted text files that contain the sequences used for calibration and comparison.

Test sets (PsojaeHA.txt, PsojaeMY.txt, PsojaeZO.txt, MtRHE.txt, DSIR.txt, MHAM.txt, KV0.txt, KV2.txt, KV3.txt): FASTA-formatted text files containing transcripts analyzed, edited for quality.

Test results (PsojaeHA.dat, PsojaeMY.dat, PsojaeZO.dat, MtRHE-A1B1.dat, MtRHE-A2B2.dat, DSIR.dat, MHAM.dat, KV0.dat, KV2.dat, KV3.dat): text files that contain transcript analysis results, sorted from least to most plant-like.