Method of predicting Splice Sites based on signal interactions

Splicing signals

Specificity in the splicing process derives partly from sequences other than SS signals, including Exonic Splicing Enhancer (ESE) and Exonic Splicing Silencer (ESS) signals [8, 9]. ESE signals are required for a constitutive exon definition and for an efficient splicing of weak alternatively spliced exons [10] (while ESS signals suppress the removal of adjacent introns [9, 11]), which may lead to exon skipping. There are 10 serine/arginine-rich (SR) Splicing Enhancer proteins known today (SRp20, SC35, SRp46, SRp54, SRp30c, SF2/ASF, SRp40, SRp55, SRp75, 9G8 [12]) and approximately 20 hnRNP Splicing Silencing factors [13], among them the most studied hnRNP A1 complex [11]. Tra2β is reported to be the SR splicing regulator [12]. All the SR proteins have two structural motifs: the RNA Recognition Motif (RRM) binding to certain motifs in RNA; and the arginine/serine-rich (RS) domain responsible for Protein-Protein interactions within splicing complex [12].

Together with inefficient SS signals, the appropriate balance of ESE and ESS elements somehow allows fine tuning of the splicing mechanism [9]. Both 5' U1 snRNP and 3' U2AF⁶⁵·U2AF³⁵ were shown to interact with ESEs [12]. Cross-intron bridging may happen through hnRNP complexes [14]. Experiments show that the 3' end definition is not affected by intron bridging, but is defined solely by the strength of the acceptor site polypyrimidine tract and the position of splicing enhancers and silencers [10, 15].

The silencing process is still poorly understood [16]. However, there are several models explaining the observed antagonism between hnRNP complexes and SR proteins [17]. For example, hnRNP A1 binds to the ESS and hinders binding of SR proteins to a weak ESE located just downstream of ESS [9]. Several rules have been identified for interaction of ESE/ESS factors with spliceosomal assembly:

• ESE and ESS elements are frequently located in downstream exons [18];

• The precise mechanism by which hnRNP A1 binds the ESS in the upstream exon and represses splicing of the upstream intron remains unknown, although the 3' site is a likely target for repression [9] as shown in Figure 2(c);

• Most splicing enhancers are located within 100 nucleotides of the 3' SS and are not active further away [15];

• Each enhancer complex assembles independently for 3' and 5' sites, and there is a minor interaction across an intron [19], as shown in Figure 2(a);

• Based on current views of exon definition, each exon should be recognized by the splicing machinery as an independent unit [3, 19], as shown in Figure 2(b);

• Analysis of the experimental data revealed that the splicing efficiency is directly proportional to the calculated probability of a direct interaction between the enhancer complex and the 3' SS:

- Strong natural enhancers function at a greater distance from the intron than weak natural enhancers do [18];

- The closer an ESE is to a SS, the more efficient it is [15];

- Multiple enhancer sites increase the probability of splicing activation [15];

- Strong ESS sites may suppress an effect from ESE(s) located upstream [9].

Bayesian sensor performance

We use the Receiver Operating Characteristic (ROC) to compare performance of different sensors. A ROC curve is a graphical representation of the tradeoff between the false negative and false positive rates for every possible cutoff. By tradition, the plot shows the false positive rate (1 - Specificity) on the x axis and the false negative rate (Sensitivity) on the y axis.

Sensitivity (Sn) and Specificity (Sp) were calculated according to the formulas

Here TE is the number of accurately predicted exon boundaries, AE is the number of annotated exon boundaries in the test set and PE is the number of predicted exon boundaries.

The accuracy of a test (i.e. the ability of the test to correctly classify cases with a certain condition and cases without the condition) is measured by the area under the ROC curve. An area of 1 represents a perfect test. The closer the curve follows the left-hand and top borders of the ROC space, the more accurate the test; i.e. the true positive rate is high and the false positive rate is low. Statistically, more area under the curve means that it is identifying more true positives while minimizing the number of false positives.

To evaluate Bayesian sensor performance we compiled three test sets:

1. 250 first multiexonic all-canonical SS genes picked from the top of our GIGOgene annotation. This test set includes 2,482 donor and 347,402 donor-like signals in addition to 2,482 acceptor and 465,000 acceptor-like signals.

2. 1,072 human 5' UTR gene-annotated fragments, including the first 50 nt from the CDS region. We picked only GIGOgene annotations containing at least one intron with all canonical SSs. Our 5' UTR test set includes 1,869 donor and 734,744 donor-like signals in addition to 1,846 acceptor and 925,464 acceptor-like signals.

3. 183 rat multiexonic all-canonical SS genes we were able to annotate with GIGOgene. This test set includes 1,405 donor and 240,539 donor-like signals in addition to 1,405 acceptor and 295,640 acceptor-like signals. The test set is specifically included to evaluate cross-specie sensor performance.

For experimental purposes on human test sets we removed cross-correlating gene-annotated fragments from the learning set. In experiments with human test sets we BLAST-aligned the test set to the learning set and removed all homologous fragments, both human and mouse, with BLASTN hit expected value less than 10^-10 and bitscore more than 75 bits. The experimental sensor performance study is shown in Figure 4. ROC curve irregularities could be attributed to multimodal score distribution of splice and splice-like signal, as could be seen in Figure 3. We did not removed cross-correlation between the learning set and the test set of 183 rat genes.

Our Bayesian 5' SS sensor outperforms the recently introduced Maximum Entropy SS sensor [22] and MDD sensor used in GenScan tool [3] as could be seen in Figures 4(a), 4(c) and 4(e). Performance of the Bayesian 3' SS sensor is similar to that of the Maximum Entropy SS sensor [22].

Bayesian 3' SS sensor design seems to favor cross-correlation between the learning set and test sequences; cross-correlation removal worsened sensor ROC characteristics [see Subsection Learning and test sets cross-correlation degree study]. In reality, sensor performance should be as good or better than reported because we used an extensive set of genes in human and mouse genomes for learning, covering most gene families. Chances are high that a sequence in question substantially cross-correlates with the learning set we use. Test example with 183 rat genes confirms our expectations, as could be seen in Figure 4(f), where ROC of the Bayesian sensor prevails over all other sensor designs.

Learning and test sets cross-correlation degree study

We study Bayesian sensor performance depending on the degree of cross-correlation between the learning and test sets. Test performance appears to be the best for experiments with intersecting test and learning sets, denoted as "Bayesian" in Figure 5. Removal of test set sequences from the learning set corresponds to curves denoted as "Bayesian (cross-validation)" and further elimination of cross-correlating sequences [see Subsection Bayesian sensor performance] corresponds to "Bayesian (no cross-correlation)" curves in Figure 5. Cross-correlation removal has a dramatic effect on the performance of the 3' SS sensor as shown in Figure 5(b). On the other hand, cross-correlation has practically no effect on the performance of the 5' SS sensor as shown in Figure 5(a).

Learning set size study

Our experiments with learning set size indicate, that 5' SS performance tolerates substantial decimation of the learning set without apparent quality loss, as shown in Figure 6(a). Decreased size of learning set causes substantial performance loss for 3' SS sensor, as shown in Figure 6(b).

The learning set we collected [see Subsection Splice Sites sensor] is not sufficiently vast for 3' SS sensor to avoid performance loss in case of removed cross-correlation. Ideal 3' SS ROC curve should be similar to the "Bayesian" as shown in Figure 5(b).

Detection of ISE signals

ISE motifs are essential components for understanding splicing events. In order to predict ISE motifs located in vicinity of 3' and 5' SS, we compiled two sample sets of 2,000 pre-donor and post-acceptor 150 nt fragments known to be real with a high degree of confidence. For fragment extraction we parsed results of GIGOgene [23] spliced alignment of human RefSeq against the phase III human DNA database from NCBI, picking only canonical SSs from predicted multiexonic structures.

We applied MHHMotif to these sample sets of sequences, and recovered motifs shown in Figure 7, where representative motifs were obtained from HMM components in "generative mode" [see Additional file 1].

Intronic conservation study

We adopt conservation criteria to evaluate significance of the putative ISE elements found [see Additional file 1]. We subdivided set of all possible 8-mer oligonucleotides in two subsets: hypothesis subset, i.e. all detected putative ISE elements, and null-hypothesis subset – the rest of the possible oligos. To study evolutionary conservation we took 3,005 mouse-rat intronic alignments. In our conservation study we considered only first 100 nt and last 100 nt of each alignment, excluding very first 5 nt and very last 5 nt for they playing role in SS highly conserved consensus signals.

We used sliding window of size 8 nucleotides to detect positions matching the set of detected putative ISEs. Ratios of conserved nucleotide positions vs. non-conserved nucleotide positions were calculated for our hypothesis set and null-hypothesis set elements. We removed test pairs where either ISEs were not detected or non of the nucleotides within putative ISEs changed. Alignments too short to extract statistics regions were also removed.

Detecting ESE signals

To detect putative ESE signals we applied the MHMMotif tool to the set of 2,000 distinct exons as we parsed the human genome annotation of our GIGOgene tool. In our experiments, motifs in Figures 8(a), 8(b), 8(c), 8(f), and 8(h) converged in two families with similar ESE signal signature but different convolution patterns supporting either 5' or 3' exonic ends. We present putative motifs [see Additional file 1]. Our set of putative ESE signals substantially overlaps with ESEs suggested by Burge and co-workers [24]. Among 202 detected putative ESE elements, 42 are present in this previously reported set of 238 ESEs, which exceeds randomly expected overlap by 3.5 times. Strong evolutionary conservation was found for these ESE signals located near splicing signals [25].

SpliceScan performance

We use test sets [see Subsection Bayesian sensor performance] for the SpliceScan performance estimates. We chose the second test set of non-coding 5' UTRs because it was suggested that at least some introns in 5' UTRs may attenuate translation at the initiation stage [26, 27]. Thus, a reliable prediction of introns and strength of splicing signals in 5' UTRs is essential for 5' UTR studies.

Performance of SpliceScan appears to be the best for short non-coding 5' UTR gene fragments test set, as presented in Figures 9(e) and 9(f). For the test set of 250 human genes and 183 rat genes our tool predicts SSs better than SpliceView [5], GeneSplicer [28], NNSplice [29], Genio [30] and NetUTR [6].

Splice Sites sensor

SSs are known to contain complex interdependencies between nonadjacent positions that could be attributed to evolutionary pressure from very complex biological splicing machinery. Recovery of these interdependencies is difficult problem in machine learning, but may lead to better understanding of the splicing process and improved sensor design.

Numerous approaches have been taken towards effective detection of SS [38–50]. The earliest SS sensors were the Weight Matrix Model (WMM) [51], Weight Array Model (WAM) [52], and Windowed second-order WAM model (WWAM) [3]. The Maximal Dependence Decomposition (MDD) sensor mentioned in [3] outperformed previously known 5' SS sensors. It explores long-range dependencies in the donor 5' motif by iterative subdivision at each stage splitting on the most dependent position, suitably denned. Leaves of the resulting bifurcation tree appear to be simple WMM models. Another approach, published in parallel and implemented in the SpliceView program [5], explores the same idea of creating WMM motif families with a clustering algorithm, which leads to better performance compared to simple WMM and WAM.

Various sensors were built later or in parallel based on Bayesian Networks [53, 54], Neural Networks [29, 55] and Boltzmann machine with Bahadur expansion [56]. Also there was a recent study based on Support Vector Machine (SVM) [57]. None of these methods were shown to outperform MDD for 5' SSs. A new Maximum Entropy sensor [22], Permuted Markov Models [58] and approach based on dependency graphs and their expanded Bayesian network [59] outperformed MDD on 5' SSs.

Methods that use subdivision of SSs into families of consensus motifs, such as MDD, SpliceView, Maximum Entropy sensor and Permuted Markov Models, appear to have the highest performance as they potentially can reveal distant correlations by putting similar-looking motifs into the same probabilistic class with its own consensus.

We use a simpler sensor design based on 7-mer oligonucleotide counting (16,384 possible oligos) in splice and splice-like signals. We place 7-mer blocks within SS consensus signals similar to Maximum Entropy Sensor [22, 60], as shown in Figure 10.

We used our GIGOgene [61] tool to collect an extensive learning set of predicted human and mouse gene structures from which we extracted 179,079 donor and 34,258,282 donor-like signals (surrounding GT dinucleotide) plus 179,076 acceptor and 44,353,884 acceptor-like signals (surrounding AG dinucleotide). Based on collected oligonucleotide frequencies, we can evaluate the probability of a 5' SS given an oligonucleotide (GT surrounded by a context)

where P(ss) – prior probability of an oligonucleotide to be 5' SS, P(¬ss) – prior probability of an oligonucleotide to be donor-like signal, P(oligo|ss) – likelihood of oligonucleotide in case of 5' SS, P(oligo|¬ss) – likelihood of oligonucleotide in case of donor-like signal. With the donor sensor we output logarithm of the block 0, given certain oligonucleotide, while in case of acceptor sensor we output sum of probability logarithms for all blocks that are calculated with formula similar to (3) under 3' SS condition. An advantage of our sensor design, compared to other projects, is in use of an extensive learning set. Because of the simple sensor structure we can combine very large numbers of splice and splice-like signals in the learning set without a dramatic learning time increase, which substantially improves sensor performance [see Subsection Bayesian sensor performance]. Close homologs or repeating gene structures in the learning set do not affect the resulting Bayesian SS sensor prediction quality; repeating domains will equally contribute scores to nominator and denominator in (3) with the resulting Bayesian probability staying the same. Our learning set is well-balanced, i.e. we use naturally occurring signals to calculate proper signal/noise ratios.

De novo motifs detection

Detection and explanation of subtle motifs in human genes is very important, since many mutations disrupting these elements may affect gene transcription, splicing or translation with possible severe consequences [62]. The problem could be formulated as a standard missing-value inference and model parameter estimation. Many approaches of motif detection are based on standard machine learning techniques, such as Gibbs sampling and Expectation Maximization (EM) algorithms. Among the tools using these algorithms are MEME [63], AlignACE [64], LOGOS [65], BioProspector [66], Gibbs Motif Sampler [67] and many others. Different approaches explore the idea of word enumeration, dictionaries and string clustering, for example RESCUE-ESE technique [24] and a recent method based on probabilistic suffix trees [68]. An interesting method of using prior knowledge in motif finding process has been presented in the LOGOS framework [65], where specific knowledge of DNA-specific bell- and U-shaped motifs signatures has been incorporated. The RESCUE-ESE method [69] is based on a prior belief that ESEs preferentially support weak SSs.

In design of our application we were primarily interested in use of prior knowledge to detect constitutive splicing enhancing elements. MHMMotif is designed to search for motifs in pre-mRNA, the single-stranded molecule. Only a fraction of sequences are assumed to have motifs of certain type within them. In our experiments we saw no apparent correlation between mRNA factors, so we assume that motifs are colocalized independently at a certain distance from target sites (e.g. TSS or SS). We also assume that motifs come in a localized family, and many of them are highly degenerate. Furthermore, motif families could be subdivided into subfamilies, with overall prediction quality improving.

As an example of preferential motif location, we present in Figure 11 Logarithm of Odds (LOD) diagrams we measured in the vicinity of SSs for two known ISE and ESE motifs, GGG [70] and (GAC) ⁿ /ACGAGAGAY/WGGACRA (9G8 binding site) [62]. We calculate LOD diagrams as the logarithm of the ratio of signal concentration near SS to the signal concentration near splice-like signal [see Section Using detected signals to improve SS prediction]. The signals have distinct bell-shaped concentration increase or decrease once we are getting closer to SS.

Our signal-detecting framework is purely based on Hidden Markov Models (HMM). HMM has many attractive properties in the context of motif detection, such as a tractable probabilistic inference and learning procedure, flexible topology, and ability to incorporate prior knowledge.

Using the mixture of Hidden Markov Models

Our strategy of motif detection is based on clustering with Mixture of Hidden Markov Models (MHMM). MHMM has a long record of successful implementations that started in speech recognition [71] and later were used for clustering protein families [72] and sequences [73]. To simulate location constraints of motifs within mRNA we use convolution of geometric states, that is described with bell-shaped negative binomial distribution, characterized by parameters p (probability to stay in the same state) and n (number of consecutive states in convolution) [74]. To detect motifs we fit our MHMM model, shown in Figure 12, to the set of sample sequences, using the Baum-Welch algorithm described in [75, 76]. The mixture component, shown in Figure 12, is a Hierarchical HMM (HHMM) with stack transformation to a plain HMM, as described in [77].

Many contemporary motif finders [63, 67] use Product Multinomial (PM) model [65]. PM model corresponds to the widely known binding motif consensus, which could be easily visualized with logos [78]. The HMM motif model is strictly more general than the standard PM model, since the HMM model is capable of catching dependencies between adjacent motif positions, which used to be a major improvement to weight matrices (PM model) [52]. Furthermore, a mixture of HMMs is potentially able to recognize dependencies between nonadjacent positions, by subdividing motif families into several subfamilies, each catching a specific group of dependencies between local positions.

By implementing the MHMM topology, as shown in Figure 13, we simultaneously learn competing HMM components of the mixture, so that they pick various competing motifs. Our application MHMMotif picks signals even if their observation frequency is low compared to simple background, i.e. a statistically significant concentration of a certain signal is not a primary detection criteria. Internal signal structure and its preferential location are the main conditions for our tool. This is another advantage of our approach compared to existing methods.

Since prediction of SS enhancing elements is still an active research topic, we decided to supplement the known enhancing motifs [see Subsection Splicing signals] with newly discovered putative enhancers. Our initial assumptions of motif finding capability of MHMM were supported by successful preliminary test results on artificial data sets, where known SF2/ASF [79] and SC35 [80] elements were detected against random background. Another experiment on the sample set of 1,871 human promoter segments (-150...+10 bps relative to initial codon) from The Eukaryotic Promoter Database [81], clearly identified known landmarks of this area, such as OCT-1, NF-1 and AP-2 factors in addition to TATA-box, CAT-box, GC-box and TATA-like A-box factor. Having preliminary-encouraging results, we applied MHMMotif tool to the splicing data sets [see Subsections Detection of ISE signals and Detecting ESE signals].

Using detected signals to improve SS prediction

Found ESE and ISE motifs have been evaluated for the ability to improve SS prediction with our new splicing simulator SpliceScan. Our splicing model is based on various-strength SS interaction with signals, such as SS themselves, and Enhancing/Silencing motifs located nearby.

SS classification enhancement in our system follows Bayesian rule in terms of Logarithm of ODds ratio (LOD) [82]

The quantity Prob(D|H) is called the likelihood of the data D (in our case ISE, ISS, ESE, ESS and competing SS signals) under hypothesis H. The last term of (4) is the LOD ratio of the prior probabilities of the Splice versus Splice-like signals, obtained with the Bayesian sensor. The first sum term in (4) takes into account the evidence provided by the data and comes up with a valid posterior LOD ratio. In order to make a noise-tolerant conversion of donor and acceptor probabilistic sensor scores into prior LOD, we approximate real signal score distribution histograms with a mixture of Beta Probability Density Functions (PDFs). The PDF of the Beta distribution is

and the mixture of n components is

. Using the Expectation Maximization algorithm for mixture learning, as explained in [83], we fitted the beta mixture (5) to the donor and acceptor SS score histograms as shown in Figure 14.

The donor histogram (shown in Figure 14(a)) was fitted in the range from -25.1 to -0.38 with the mixture

Mix _donor (p) = 0.006 × Beta(p, 0.06, 0.91) + 0.994 × Beta(p, 14.27, 1.80)

and the acceptor histogram (shown in Figure 14(c)) was fitted in the range from -111.01 to -19.19 with the mixture

Mix _acceptor (p) = 0.83 × Beta(p, 48.42, 6.37) + 0.17 × Beta(p, 7.18, 1.85)

where we assume probability argument in the formula

We built LOD diagram for each of the enhancing motif interacting with SS of different strengths (in range from 1 to 10), similar to shown in Figure 15. First, we measure normalized signal concentrations around SS, as shown in Figure 15(a). Using Matlab^® polynomial interpolation we approximated characteristics as could be seen in Figure 15(a). In order to find LOD characteristic, shown in Figure 15(b), we calculate , where Prob(D|H _SS ) is normalized signal concentration at certain location next to a SS and Prob(D|H_¬SS) is normalized signal concentration at certain location next to a splice-like signal. Signal LOD scoring happens as schematically shown in Figure 15(c).

We were not able to use all the available enhancing/silencing signals for SS prediction. Part of the problem appears that the linear LOD sum accumulates noise and can only include limited number of factors. The signals we used for donor enhancement are all ISE elements supporting 5' SS shown in 7(h)-7(m), plus ESE element shown in Figure 8(a), 8(d), 8(g) and 8(j) combined with hnRNP A1, poly A and srp20 signals [see Subsection Splicing signals]. Surprisingly, polyadenylation signal use contributed substantial enhancement to the ROC diagram. We were not able to use SF2/ASF and SC35 enhancing signals for having LOD characteristics with indistinct enhancement profile. For acceptor enhancement we used ISE elements supporting 3' SS shown in Figures 7(b), 7(c) and 7(e) plus ESE element shown in Figure 8(a), 8(d) and 8(g) combined with hnRNPAl, polyA and srp20 signals.

SS/SS interaction LOD diagrams are the key ingredient of our prediction algorithm as they contribute most to SS prediction enhancement. Figure 16 shows various interaction diagrams. X-axis is the SS vicinity position, where SS of strength 1 located at 0, Y-axis is the strength of interacting signal and Z-axis is the LOD characteristic. The main conclusion we make based in Figures 16(a) and 16(b) is that weak signal preferentially avoids strong competitors nearby, especially inside exon, as they can redefine exonic boundary. Figures 16(c) and 16(d) indicate that weak signal (donor or acceptor) has preferential need for strong complementing exonic boundary.

Reviewer's report 1

Manyuan Long Department of Ecology and Evolution, The University of Chicago Chicago, United States, Email: mlong@uchicago.edu

The authors attempted to develop a simple sensor to detect splice and splice- site signals. A 7-mer was designed to scan a sequence. The ROC diagrams (Fig. 3) showed its obvious advantage, significantly higher specificity and sensitivity than other methods. In addition, the authors also used MHMM to detect ISE and ESE signals and used found signals to improve SS prediction. I think that the authors developed useful new methods for SS detection and I favor its publication in Biology Direct. However, I also have following minor concerns and hope them get fixed in revision.

Page 1: "Figure 1...": is not original, some sources should be cited (for example, the early work of Tom Schnider of NCI in 1992?". Originality is something that a paper in bioinformatics wants to emphasize.

Reviewer's report 2

Arcady Mushegian, Stowers Institute, Kansas City, United States

Email:ARM@Stowers-Institute.org

Good: 1. Part 2: the main idea appears to be to trade a more complex model for a larger training set. This seems to improve specificity of the splicing site detection.

Two relevant issues that are not discussed but should be: a. In Figure 3, all ROC curves are still below the non-discrimination line – is this acceptable.

Reviewer's report 3

Mikhail Gelfand, Institute of Information Transfer Problems, Moscow, Russian Federation Email: gelfand@iitp.ru

On "Method of predicting splice sites based on signal interactions" by A.Tchourbanov et al., submitted to "Biology Direct"

The problem of identification of donor and acceptor splicing sites is not new, but far from solved, whereas identification of sites regulating splicing (exonic/intronic splicing enhancers/silencers) has emerged relatively recently. Given the importance of both these problems for gene recognition and understanding alternative splicing, any progress in this area is most welcome.

The authors attempt to address both problems in one framework of Bayesian analysis. They apply Bayesian sensors to detection of donor and acceptor splicing sites. The exposition in this part (section 2, pp. 2–3) contains several gaps. It is not clear how well the described approach of 7-mer counting with subsequent Bayesian weighting generalizes; in particular, it seems that the sensors will not accept a completely new 7-mer as a site. If the authors implicitly claim that all possible 7-mers have already been observed in the training set, and the only problem is proper weighting, this needs to be substantiated. A helpful piece of data would be the rank distribution of 7-mers in the positive and negative sets. How many 7-mers have been observed only once in the positive set (and would be missed if only half of that set were used for training)?

Author response

With cross correlation removed between learning set and the test set, when testing on the set of 250 human genes, we had miss rate of 0.52% for our 5'SS sensor, which is acceptably low value. For 3'SS sensor overall miss rate is negligibly low, since sensor topology is composite of several blocks. We show top 40 ranking 5'SS nonamers in Table 2. We added discussion on sensor performance related to learning set size [see Subsection Learning set size study/, where we show that Bayesian sensor has preference for the large learning sets. For example, the sensor could be successfully applied to recognition of the Translation Initiation Site (TIS) against upstream A UGs, where we can collect large learning set. In the TIS sensor design we used three strategically located heptamers, so that they can catch both long-range dependencies and initial codon bias, as shown below:

We used 42,883 TIS and 77,140 TIS-like signals from human, mouse and rat RefSeq databases to learn our TIS Bayesian sensor, which demonstrated, in our preliminary experiments, superior performance as compared to simple Kozak's consensus rule (GCC)GCCRCC AUG G (where R = G or A) [84] and corresponding weight matrix. However, the sensor design does not generalize well to recognition of other signal types, such as transcription factors, with very thin learning sets.

Another missing part of exposition is a formula for combining several sensors for acceptor site analysis. Is the final score (probability) obtained by multiplying probabilities assigned by the sensors?