Conceived and designed the experiments: JC SFE AG. Performed the experiments: JC. Analyzed the data: JC AJ. Wrote the paper: JC SFE AG. Generated and phenotyped the RILs: RFM. Generated the volatile dataset: AFdC JLR. Generated the transcript dataset: CP.
The authors have declared that no competing interests exist.
Considering cells as biofactories, we aimed to optimize its internal processes by using the same engineering principles that large industries are implementing nowadays: lean manufacturing. We have applied reverse engineering computational methods to transcriptomic, metabolomic and phenomic data obtained from a collection of tomato recombinant inbreed lines to formulate a kinetic and constraint-based model that efficiently describes the cellular metabolism from expression of a minimal core of genes. Based on predicted metabolic profiles, a close association with agronomic and organoleptic properties of the ripe fruit was revealed with high statistical confidence. Inspired in a synthetic biology approach, the model was used for exploring the landscape of all possible local transcriptional changes with the aim of engineering tomato fruits with fine-tuned biotechnological properties. The method was validated by the ability of the proposed genomes, engineered for modified desired agronomic traits, to recapitulate experimental correlations between associated metabolites.
Considering cells as biofactories, we aimed to optimize their internal processes by using existing design principles acquired from engineering. Herein, we present a synthetic biology approach based on experimental and computational methodology that integrates genomic, transcriptomic, metabolomic and phenomic data to formulate a kinetic and constraint based model of tomato agronomic and fruit quality characteristics. The model has been used for exploring the landscape of all possible local transcriptional changes with the aim of engineering tomato fruits with improved biotechnological properties. The methodology was validated by the ability of the proposed engineered genomes with modified desired agronomic traits, to recapitulate correlations between associated metabolites that are found experimentally in a number of examples.
Considering a cell as a DNA-based molecular factory
Previous works have considered modeling the global metabolism
Finally, in order to close the design cycle imposed by LM, the genetic modifications suggested by our computational approach were experimentally verified. This was done by demonstrating the predicted ability of the
We have extended our recently developed inference methodology,
From omic data (transcriptomics, metabolomics and phenomics), a quantitative global model was constructed using reverse engineering methods. The predictive model was used to propose genome perturbations, to improve desired phenotypes with relevant biotechnological applications. The genome perturbations were guided by an
Transcriptomic and metabolomic data from these 50 RILs were normalized by the LOWESS method
(A) Prediction of the agronomic properties experimentally measured over the 169 RILs. The straight line represents the exact prediction. (B) Distance between distributions of Pearson correlations for the fruit agronomic properties, metabolites and genes (green, red and blue points, respectively) over training sets and in random permutations of them with different noise levels. (C, D) Histogram of Pearson correlations between the measured and predicted metabolite and gene levels over their training sets (blue bars) and over sets with a 10- and 5-fold cross validation tests (red bars), respectively.
The next step was to construct an effective gene regulatory model able to predict autonomously the transcriptional processes that, by means of the model previously described, would generate a quantitative metabolic response. In this way changes at the transcriptional level resulting from the proposed genetic perturbations could be translated and predicted effectively into metabolic changes. For doing that, we used the microarray data obtained from fruits of 50 of the RILs to infer a network of gene-gene interactions. The CLR method provided the first sets (
We addressed the question of whether the agronomic/phenotypic properties of the tomato fruit could be determined by their metabolite composition. For that, we studied the relationship between agronomic properties and metabolic composition across 169 tomato RILs. We applied LASSO method to select a set of metabolites that may act as predictors for each agronomic property (
Next, to test the specificity of the inferred model parameters, we perturbed the target phenotypic profile for each RIL adding different levels of noise.
Here, our main goal is to redesign the genome of tomato to generate an engineered surrogate that, if viable, would be easier to study and of greater potential biotechnological interest. Our design approach was inspired by the practice of
Hence, mimicking the optimization patterns typical from LM, the landscape of desired agronomic properties of tomato fruit was exhaustively explored perturbing its effective transcriptional regulatory network (TRN) with single-gene alterations.
(A) Agronomic properties improved by perturbing a single gene as function of efficiency reached by that transcriptional perturbation with respect to the wild-type scenario; only perturbations causing positive mean efficiencies are plotted. Both agronomic properties and efficiencies of a single perturbation are tested on the 169 RILs and error bars represent their minimum and maximum values in both axis. (B) Relationship between agronomic properties in the wild-type genome and the average of the agronomic properties resulting of all single perturbations in the wild-type TRN for each RIL; vertical error bars represent the best and worst optimized re-engineered TRN for a given RIL. (C) Average number of single gene perturbations that overcome a given efficiency threshold in the 169 RILs (light bars; error bars represent standard deviation for the 169 RILs) and average probability of selecting the same gene-perturbation in a set of RILs (dark bars; error bars show standard deviation for all genes of the TRN). Left and right columns represent perturbations of single gene in case of knockout or over-expression, respectively. (A, B) show fitness as related to the acceptability of tomato fruit (blue) and production vs. quality (red); (C) and fitness values associated to maximize only fruit quality (green). Agronomic properties are plotted in arbitrary units.
After this, we ranked the list of knockout/over-expressed genes of the TRN according to two criteria directed to maximize: (i) the mean efficiency across all lineages in the case of goals such as acceptability and quality vs production; and (ii) the average of the maximum agronomic property reached by all possible TRN reconfigurations in the case of fruit quality (
Gene | Gene Annotation | Efficiency |
RIL |
|
|
||||
LE24K20 | Eukaryotic translation initiation factor 2 gamma subunit, putative | 2.91 | 103 | 1 |
LE18G02 | Heat shock protein, putative | 1.81 | 103 | 1 |
LE30E17 | Amino acid binding protein, putative | 1.79 | 103 | 1 |
LE21B20 | Chaperone GrpE type 2 | 1.68 | 103 | 1 |
LE11F03 | GATA transcription factor, putative | 1.45 | 103 | 1 |
LE13M10 | Ribosomal protein L30e | 8.84 | 103 | 1 |
LE32K06 | LEXYL2 | 5.87 | 103 | 1 |
LE14B20 | Clathrin adaptor complexes medium subunit family protein | 1.08 | 103 | 0.98 |
LE33M04 | Splicing factor 3B subunit, putative | 3.46 | 103 | 1 |
LE3H15 | Non-cell-autonomous protein pathway1, plasmodesmal receptor | 0.48 | 103 | 0.95 |
|
||||
LE24K20 | Eukaryotic translation initiation factor 2 gamma subunit, putative | 43.34 | 142 | 0.30 |
LE18G02 | Heat shock protein, putative | 39.06 | 142 | 0.31 |
LE25A03 | Ribosomal protein S27-like protein | 11.98 | 142 | 0.13 |
LE14J12 | 40S ribosomal protein S3a-like | 11.82 | 142 | 0.13 |
LE33G09 | Predicted protein from Populus trichocarpa | 11.75 | 142 | 0.31 |
LE15D07 | Polynucleotide kinase- 3′-phosphatase, putative | 227.31 | 142 | 0.31 |
LE27C02 | Phytoene dehydrogenase, chloroplastic/chromoplastic | 186.12 | 142 | 0.31 |
LE8A19 | Putative glycerophosphoryl diester phosphodiesterase family protein | 169.35 | 142 | 0.31 |
LE3H15 | Non-cell-autonomous protein pathway1, plasmodesmal receptor | 143.53 | 142 | 0.31 |
LE14B20 | Clathrin adaptor complexes medium subunit family protein | 135.47 | 142 | 0.31 |
|
||||
LE13F23 | Chloroplast phosphate transporter precursor | 15.32 | 135 | 0.63 |
LE15L08 | Putative rac protein | 12.32 | 135 | 1 |
LE1P20 | Glycyl-tRNA synthetase 2, chloroplast/mitochondrial precursor, putative | 12.00 | 135 | 1 |
LE22K20 | Ubiquitin-conjugating enzyme E2, putative | 11.27 | 135 | 0.93 |
LE26N09 | 6-phosphogluconolactonase-like protein | 10.23 | 135 | 0.99 |
LE14B20 | Clathrin adaptor complexes medium subunit family protein | 35.94 | 135 | 1 |
LE16L04 | Ureide permease, putative | 28.05 | 135 | 0.98 |
LE3H15 | Non-cell-autonomous protein pathway1, plasmodesmal receptor | 23.04 | 135 | 1 |
LE15D07 | Polynucleotide kinase-3′-phosphatase, putative | 20.22 | 135 | 1 |
LE8A19 | Putative glycerophosphoryl diester phosphodiesterase family protein | 16.59 | 135 | 0.63 |
Gene | Gene Annotation | Efficiency |
RIL |
|
|
||||
LE24K20 | Eukaryotic translation initiation factor 2 gamma subunit, putative | 2.91 | 103 | 1 |
LE18G02 | Heat shock protein, putative | 1.81 | 103 | 1 |
LE30E17 | Amino acid binding protein, putative | 1.79 | 103 | 1 |
LE21B20 | Chaperone GrpE type 2 | 1.68 | 103 | 1 |
LE11F03 | GATA transcription factor, putative | 1.45 | 103 | 1 |
LE13M10 | Ribosomal protein L30e | 8.84 | 103 | 1 |
LE32K06 | LEXYL2 | 5.87 | 103 | 1 |
LE14B20 | Clathrin adaptor complexes medium subunit family protein | 1.08 | 103 | 0.98 |
LE33M04 | Splicing factor 3B subunit, putative | 3.46 | 103 | 1 |
LE3H15 | Non-cell-autonomous protein pathway1, plasmodesmal receptor | 0.48 | 103 | 0.95 |
|
||||
LE24K20 | Eukaryotic translation initiation factor 2 gamma subunit, putative | 43.34 | 142 | 0.30 |
LE18G02 | Heat shock protein, putative | 39.06 | 142 | 0.31 |
LE25A03 | Ribosomal protein S27-like protein | 11.98 | 142 | 0.13 |
LE14J12 | 40S ribosomal protein S3a-like | 11.82 | 142 | 0.13 |
LE33G09 | Predicted protein from Populus trichocarpa | 11.75 | 142 | 0.31 |
LE15D07 | Polynucleotide kinase- 3′-phosphatase, putative | 227.31 | 142 | 0.31 |
LE27C02 | Phytoene dehydrogenase, chloroplastic/chromoplastic | 186.12 | 142 | 0.31 |
LE8A19 | Putative glycerophosphoryl diester phosphodiesterase family protein | 169.35 | 142 | 0.31 |
LE3H15 | Non-cell-autonomous protein pathway1, plasmodesmal receptor | 143.53 | 142 | 0.31 |
LE14B20 | Clathrin adaptor complexes medium subunit family protein | 135.47 | 142 | 0.31 |
|
||||
LE13F23 | Chloroplast phosphate transporter precursor | 15.32 | 135 | 0.63 |
LE15L08 | Putative rac protein | 12.32 | 135 | 1 |
LE1P20 | Glycyl-tRNA synthetase 2, chloroplast/mitochondrial precursor, putative | 12.00 | 135 | 1 |
LE22K20 | Ubiquitin-conjugating enzyme E2, putative | 11.27 | 135 | 0.93 |
LE26N09 | 6-phosphogluconolactonase-like protein | 10.23 | 135 | 0.99 |
LE14B20 | Clathrin adaptor complexes medium subunit family protein | 35.94 | 135 | 1 |
LE16L04 | Ureide permease, putative | 28.05 | 135 | 0.98 |
LE3H15 | Non-cell-autonomous protein pathway1, plasmodesmal receptor | 23.04 | 135 | 1 |
LE15D07 | Polynucleotide kinase-3′-phosphatase, putative | 20.22 | 135 | 1 |
LE8A19 | Putative glycerophosphoryl diester phosphodiesterase family protein | 16.59 | 135 | 0.63 |
Notice that the first five genes is the top 5 of single-gene knockouts and the following five is the top 5 in over-expression.
Efficiencies were selected in the RIL where the perturbation maximizes the fitness.
Probability of selecting the given perturbation across the set of RILs at the maximum level of efficiencies.
Lineages exhibited variability in their resistance to be optimized and this resistance changed with each target agronomic property.
We computed the average number of single-gene perturbations to overcome an efficiency threshold given in the 169 RILs and the average probability of selecting the same gene-perturbation commonly for the whole set of RILs. The right panel in
The next step in our study was to propose new genome re-designs including multiple perturbations. To do this, we sampled widely the landscape of the acceptability, quality and quality vs production of tomato fruits by introducing two-gene perturbations either by knockouts and over-expressions (
Heuristic exploration (A) and statistical significance (B) of the landscape of multiple desired agronomic properties of tomato fruit perturbing its effective TRN adding multiple genetic changes and, predictive power (C–F) for optimizing the levels of volatile compounds and identifying compounds in closed metabolic pathways. (A) Median efficiencies reached by transcriptional perturbation based in gene knockouts or over-expression to improve agronomic properties. (B) Average number of single gene perturbations that overcome an efficiency threshold in the top 5 RILs scored by single perturbation (light bars; error bars represent standard deviation for the selected RILs) and average probability of selecting the same multiple-perturbation commonly in a set of RILs (dark bars; error bars show standard deviation for all genes of the TRN). Precision, recall and F-score (green, red and blue lines, respectively) compare observed experimentally volatile compound correlations vs inferred set of potential genetic perturbations (gene knockout (C, D) or over-expression (E, F)) shared to optimize each compound independently. Note that experimental metabolite correlations r<0.5 were not considered in (D, F).
Gene | Gene Annotation | Efficiency |
|
||
|
|
16.54 |
|
|
16.54 |
|
|
16.40 |
|
|
16.37 |
|
|
16.07 |
|
|
15.90 |
|
|
15.87 |
|
|
15.85 |
|
|
15.84 |
|
|
15.78 |
|
||
LE27F15; LE29L05 | Protein kinase family protein; branched-chain amino acid aminotransferase | 422.60 |
LE16D08; LE6G08 | Similar to 60S ribosomal protein L35; sucrose phosphate synthase | 360.63 |
LE9A08; LE15E23 | GRAM domain-containing protein/ABA-responsive protein-related; putative threonyl-tRNA synthetase | 303.04 |
LE18E13; LE8A19 | MYB transcription factor; putative glycerophosphoryl diester phosphodiesterase family protein | 263.91 |
LE32B05; LE4D06 | YABBY2-like transcription factor YAB2; tRNA-dihydrouridine synthase A, putative | 253.33 |
244.32 | ||
LE29E13; LE13F06 | Fyve finger-containing phosphoinositide kinase, fyv1, putative; transmembrane protein, putative | 242.56 |
LE13F06; LE15J03 | Transmembrane protein, putative; ankyrin-like protein | 240.19 |
LE17G02; LE15D07 | Pantothenate kinase, putative; polynucleotide kinase- 3′-phosphatase, putative | 239.10 |
LE15D07; LE20I03 | Polynucleotide kinase- 3′-phosphatase, putative; DEX1, calcium ion binding | 239.03 |
|
||
LE13F06; LE15J03 | Transmembrane protein, putative; ankyrin-like protein | 49.79 |
LE12O13; LE33G22 | Prefoldin subunit, putative; adenylate kinase, putative | 49.16 |
LE2C24; LE29J02 | ATAB2; RNA binding; GTP-binding protein LepA homolog | 49.15 |
LE12P11; LE2C24 | Not found; ATAB2; RNA binding | 48.81 |
LE2C24; LE21J01 | ATAB2; RNA binding; Dolichyl-phosphate beta-glucosyltransferase, putative | 48.28 |
LE12O13; LE25M06 | Prefoldin subunit, putative; Pre-mRNA-processing protein prp39, putative | 46.63 |
LE12O13; LE14B20 | Prefoldin subunit, putative; clathrin adaptor complexes medium subunit family protein | 46.18 |
LE14B20; LE21J01 | Clathrin adaptor complexes medium subunit family protein; dolichyl-phosphate beta-glucosyltransferase, putative | 44.86 |
LE33B09; LE2C24 | Not found; ATAB2; RNA binding | 44.64 |
44.05 |
Efficiencies were selected in the RIL where the perturbation maximizes the fruit acceptability, quality and, quality
Knockout genes were showed in bold type and the others were gene over-expressed.
After generating our predictive model for the TRN and metabolism of tomato fruit, we use it to automatically design tomato genomes with extreme alterations for each of the 56 volatile compounds by introducing a set of genetic perturbations. We compared sets of genetic perturbations for all pairs of volatile compounds and then inferred their levels of correlations (see Methods). Hence, these predicted correlations were compared to the levels of correlations obtained from the experimental values for each volatile pair that often reflects their belonging or not to the same metabolic/regulatory pathway or to be or not structurally related.
To give further support to our model we constructed experimentally two inbred lines (ILs) derived from another interspecific cross whose transcriptome and metabolome were also experimentally measured. Parents of these ILs are a different cultivar of tomato M82 and a
LM is a methodology that is being implemented by large industries to optimize their production. In the process of decision making applied to the redesign of production systems, firstly, engineers evaluate systematically the addition or elimination of resources in each of the participating single processes; afterwards, multiple changes are considered trying to achieve maximum quality and production
We have proposed several re-engineered genomes that improve desired agronomic properties of the fruit by targeting single or multiple genetic modifications. It has been previously reported that single under-/over-expressed of certain genes may affect fruit quality traits, being these key genes involved in the biosynthesis of a product of fruit metabolism or to a general ripening regulators (i.e., carotenoids
Although it is not the objective of this paper, it does not escape our attention that some of the perturbations proposed are consistent with the biological processes associated to the trait and therefore the model could be used to reveal the molecular underpinnings of quality traits (see experimental evidences of each gene perturbation proposed by the model in the
The ability to target redesign crops for enhanced content of metabolites of interest has been experimentally achieved in a number of cases (for instance vitamins C
The construction of the tomato RILs used in this study has been described elsewhere
An effective linear model based on ODEs each providing the steady states of tomato fruit mRNA was used to describe transcriptional gene regulations
Our global model consists of three blocks of algebraic equations covering respectively from gene expression, through metabolic profile until agronomic properties, and in all three cases the same methodology was applied. The inference procedure consisted of two nested steps. Firstly, the network connectivity was inferred by using the
For the construction of the effective TRN model and its later integration with the metabolism, we used steady-state mRNA expression profiles derived from RILs transcriptionally and metabolically characterized. The dataset contains pre-processed expression data from 50×3 = 150 hybridization experiments using an array with 11876 probe sets spotted, and data for levels of 67 metabolites that were quantified over the same sample set. For this study, we only considered the 5592 genes whose expression values could be consistently found in more than 80% of the microarrays. We found 1057 TFs and 1962 genes with enzymatic activity after searching for the motifs transcription regulator and enzyme activity respectively in the functionally annotated tomato genome (
Three plain text files containing the transcriptional, metabolic and phenotypic model for tomato were constructed and are available in
Our algorithm searches possible reconfigurations of the global effective transcription regulatory network of tomato such as that the specified agronomic properties are improved (maximized or minimized) with respect to the properties of interest obtained in a given RIL. Different properties of interest have been optimized, ranging from single metabolites defining the sweetness or sourness of the fruit, to linear combinations of a set of metabolites determining the quality in terms of flavor and taste and even further to include objective functions that try to integrate two of those goals with a trade-off and balanced weighting factors such as fruit quality and yield.
We have addressed this optimization problem using two approaches. Firstly, we exhaustively enumerated all possible single gene knockouts and over-expression for each case to be optimized under a given selective pressure of interest. Second, we ranked all possible perturbations according to the new agronomic properties they would generate. The third step was to suggest genome reconfigurations that include multiple actions: gene knockouts, over-expressed genes, or both, in order to enlarge the combinatorial space of perturbed genomes. To do that, we have used an exhaustive method aimed at finding the global optimum in the space of all possible synthetic TRN. We started from the inferred model (see Mathematical model above) and applied an optimization scheme. At each step of the optimization process, we selected each gene among the ones involved in the transcriptomic model to evaluate the effect of three possible approaches (knockout, over-expression or wild-type scenario); we updated the model with the genetic perturbation that provided the best score. Note that to simulate knockout or over-expression in the gene
We computed the sets of single-gene perturbations, ?, by gene knockout or over-expression that alter significantly the levels of the 56 volatile metabolites representing the volatile compounds taking into account the global model. For the sake of the model we considered only those gene perturbations that would cause significant changes in the metabolite concentration higher than 1% (
The performance of the inferred metabolite correlations was evaluated using as a reference a set of empirical correlations previously obtained among these metabolites. We used different cut-offs, k, to identify metabolite correlations (
To estimate the range of
Transcriptional, metabolic and phenotypic models of tomato fruit.
(XLS)
Single knockout and over-expressed genes to improve desired agronomic properties (acceptability, quality and quality vs production of tomato fruits; four volatile compounds; vitamin C, and different types of sugars and acids) and functional categorization of genes that induced high degree of improvement in those agronomic properties; notice that functional enrichment of all genes involved in the TRN was included. Gene ontology enrichment analyses were performed using the TFGD tool [TFGD]. It is also showed the functional categories significantly represented among those genes that were selected to describe the TRN of tomato fruit. A total of 19 cellular processes and 45 biological components were represented. Among these, genes related to cellular metabolic processes were the most abundant (p<0.0001), what makes sense since they were selected to predict cellular metabolism; whereas genes related to response to nutrient stimulus were present but the least common (p<0.1).
(XLS)
Multiple combinations of knockout and over-expressed gene sets to improve desired agronomic properties (acceptability, quality and quality vs production of tomato fruits).
(XLS)
Experimental evidences of each gene perturbation proposed by the model to optimize the different scoring function used.
(XLS)
Synthetic biology of tomato fruit
(PDF)
From data to global models to redesign using an approach based on synthetic biology.
(PDF)
Phenotype prediction (number of fruits per plant, fruit harvested, average fruit weight and pH) by using the genotype described in the 50 RILs in which transcript levels were measured. Pearson coefficient correlation (A,C) between the predicted and measured phenotypic profile and number of genes (B,D) selected by LASSO method as predictors for different thresholds of the fitting parameter (tLASSO). Note that we used two different z-score levels (z = 2, (A,B); and z = 3 (C,D)) to included genes as possible predictors to be selected by LASSO. The dashed line plotted in (A,B) shows the parameter, tLASSO, and the level of z-score used to constructed the relationship between phenotype and metabolome.
(PDF)
Exhaustive exploration and statistical significance of the landscape of single desired agronomic properties of tomato fruit (vitamin C, blue; fructose and glucose, red; and citric and malic acids, green) perturbing its effective TRN locally. (A) Agronomic properties improved by perturbing a single gene as function of efficiency reached by that transcriptional perturbation with respect to the wild-type scenario; notice that only perturbations with positive mean efficiencies are plotted. Both agronomic properties and efficiencies of a single perturbation are average variables tested on the 169 RILs and error bars represent their minimum and maximum values in both axis. (B) Dependence between agronomic properties in the wild-type genome and the average of the agronomic properties resulting of all single perturbations in the wild-type TRN for each RIL; vertical error bars represent the best and worst optimized re-engineered TRN for a given RIL. (C–D) Average number of single gene perturbations that overcome an efficiency threshold in the 169 RILS (light bars; error bars represent standard deviation for the 169 RILs) and average probability of selecting the same gene-perturbation commonly in a set of RILs (dark bars; error bars show standard deviation for all genes of the TRN). Left and right columns represent perturbations in terms of single gene knockout or overexpression, respectively.
(PDF)
(A) Dendogram of the volatile compound correlations observed experimentally. (B, C) Dendograms inferred by the model defining the distance between volatile compound as the number of common genetic perturbations predicted to optimize the levels of each volatile compound.
(PDF)
Percentage of altered genes (via gene knockout or over-expression; blue bars) proposed by the model to minimize the levels of volatile compounds (linalool (A) or, 1-nitro-2-phenylethane, 2-isobutylthiazole and benzylnitrile (B)) that were found significantly over-/under-expressed in the transcriptome of two ILs characterized experimentally with extremely low levels of those volatile compounds. The cut-off of the coefficient of variation between replicates was 75%. The Mann-Withney's
(PDF)
Correlations observed between agronomic variables and metabolites of different fruit genotypes generated by simulating all possible single gene knockout (A–E) or over-expression (F–J) in the wild-type genome model of the tomato fruit. Standard deviations of all metabolites or agronomic variables show the diversity generated by implementing each genetic perturbation in the 169 RILs. Note that we only plotted re-engineered genomes whose transcriptome predicted showed errors lower than 1% (241 d.f. and 25 d.f. for knockout and over-expressed genes, respectively).
(PDF)
The top 5 single-gene knockouts and over-expressions that maximize the agronomic properties of the tomato fruit based on improve only one objective.
(PDF)
The top 5 single-gene knockouts and over-expressions that minimize the agronomic properties of the tomato fruit based on improve only one objective.
(PDF)
Genome design based on single perturbations to fine-tuning phenotypes with biotechnological interests. Model validation: fine-tuning tomato phenotype of two experimental inbred lines by computational genome design. Prediction of phenotypic correlations in re-engineered tomato fruits.
(PDF)
We thank Sophie Mirabel for excellent technical skills in microarray hybridization, J. Forment for help with computer resources and, G. Rodrigo and F. Heras for his fruitful comments.