We adapted an improved micropurification procedure to obtain enriched mitochondrial fractions. The purity of these fractions in terms of presence or absence of mitochondria was routinely assessed by northern blots or by reverse transcriptase (RT)-PCR analysis using genes for mitochondrial, cytosolic and plasma membrane proteins as molecular probes (see Additional data files). However, the mitochondrial fractions thus obtained always included mRNAs from other intracellular membranes (data not shown). This is not surprising, as mitochondria and endoplasmic reticulum are known to be highly interconnected in S. cerevisiae 15. This was taken into account in the following analyses. After growth in galactose and RNA extraction, the amplified samples were treated according to the manufacturer's instructions. To assess mRNA quality, a quality-control was included in the protocol before amplification. Figure 1 summarizes the biological question and the experimental strategy.

A major difficulty in studying mRNA localization with microarrays is microarray normalization. All current methods of normalizing microarray data were developed to allow between-chip comparison in transcriptome expression experiments 1617. The hypotheses they rely on - the constant expression levels of most genes across experiments, for instance - do not hold true when different subcellular mRNA populations are compared. In a previous experiment, Cy5/Cy3 ratios were ranked into percentiles and averaged to compare different experiments 14. Diehn et al. 18 adapted a moving-average analysis to study membrane-bound polysomes. In both cases, the method of combining results from different experiments was not straightforward and the question of the reproducibility was not readily addressed.

In this study, a more adequate experimental design and a more complete dataset allowed us to design a more direct strategy, which is statistically more satisfactory. Three mRNA populations were isolated and hybridized to distinct chips: MP (mitochondrial and membrane bound polysomes); FP (free cytosolic polysomes); and TOT (total polysomes, that is, the whole transcriptome). This experiment was repeated twice, giving six results for each probe set. Pairwise quantile normalization was then applied (MP₁ and MP₂; FP₁ and FP₂; TOT₁ and TOT₂). The expression levels were then calculated (see Materials and methods for details). Spike mRNAs present on Affymetrix gene chips were used to check whether an additional normalization step was necessary to allow comparisons among the three normalized couples. As these exogenous control mRNAs, which were introduced before RNA amplification and chip hybridization, gave almost identical signals for the six chips, no further correction was applied 19. Of the probe sets 3% had at least one expression value under an arbitrary threshold. They were eliminated from subsequent analysis, to filter against uncertain extremely low intensities. A localization parameter called MFI (for membrane versus free index) was defined for each triplet of chips (MP₁, FP₁, TOT₁ and MP₂, FP₂, TOT₂) and each transcript by the formula:

MFI_i = (log(MP_i) - log(FP_i))/log(TOT_i) where (i = 1 or 2).

The MFI₁ and MFI₂ values were very close, indicating good reproducibility of the whole process. Hence, the MFI₁ and MFI₂ values were averaged. The methodology described above, in particular the normalization procedure and the utilization of the total mRNA population (TOT), can be easily generalized to efficiently analyze larger-scale experiments involving more subcellular mRNA populations and more chips.

The MFI values were compared to the 'Mitochondrial Localization of RNA' (MLR) values previously obtained from a different experimental approach. This comparison was limited to the available MLR data as the previous analysis covered only about half the genome. Despite quite different experimental designs, the comparison resulted in good correlation values, thereby validating our new approach. More precisely, previous localization values were calculated using data from one to six independent experiments. Figure 2 is a plot of the rank correlation coefficient between new values (MFI) and old values (MLR) as a function of the number of experiments included in the former. The graph includes 95% confidence intervals calculated with non-parametric bootstrap 20. The correlation increases quite sharply from one to four experiments, after which the curve is almost flat, with a correlation coefficient between ranks of above 0.6 from four to six experiments. This clearly demonstrates that MFIs more accurately quantify the localization than the previous MLRs.

The localization results for 97% of the probe sets, corresponding to 96% of the genes (95% of the genes encoding mitochondrial products) are available online at 20. Individual results or lists of results can be retrieved by ORF name, gene name or Affymetrix probe set identifier (affxid). To facilitate data interpretation, MFIs were grouped in 10 evenly divided gene classes. For instance, a mitochondrial gene belonging to MFI class 10 has its MFI among the genome's 10% highest, and hence its mRNA is much enriched in mitochondrial outer-membrane-bound polysomes. Figure 3 presents the genome-wide distribution of MFI values and highlights two lists of mitochondrial genes with extreme MFI values. More precisely, the mitochondrial genes listed on the right of the plot have MFI values among whole-genome highest 5% (the upper half of MFI class 10) and correspond to mitochondria-localized transcripts. In contrast, the list of mitochondrial genes on the left of the plot have MFI values among whole-genome lowest 5% (the lower half of MFI class 1) and correspond to transcripts that are translated on free cytosolic polysomes. The subcellular localization of many yeast proteins is reasonably well known. We used the public database MIPS CYGD 21 as a reference. We then plotted the empirical cumulative distribution functions of the MFIs for mRNAs belonging to three separate known subcellular protein localization categories (Figure 4). As expected, genes encoding cytosolic proteins have low MFIs, whereas genes encoding plasma membrane and endoplasmic reticulum proteins have higher MFIs. If mitochondrial proteins were all translated from free cytosolic polysomes, one would expect the corresponding curve to be quite similar to that of cytosolic proteins. In contrast, the curve for mRNAs encoding mitochondrial proteins is between the two reference curves for mRNAs translated on cytosolic and on membrane-bound polysome fractions, respectively. This is the most provocative aspect of these data because it shows that a large proportion of the mRNAs coding for mitochondrial proteins is actually localized to the vicinity of mitochondria. Thus, what had been previously observed by various experimental approaches is confirmed and quantified at a genome-wide level.

A critical point in such an approach is the procedure used to purify the mitochondria. The basic protocol used for isolation of mitochondria-associated polysomes is very similar to the previous procedure published by Butow's group 10. However, we previously observed 14 that the linear sucrose gradient step was not necessary to obtain reliable results. Accordingly our results are consistent with those previously obtained by Butow's group on a limited set of genes. However it should be emphasized that this purification protocol does not lead to pure mitochondrial fractions. Considerable contamination by diverse membrane fractions is generally considered unavoidable; nevertheless these contaminants do not affect the general conclusion drawn from the results of Figure 3: the nuclear genes known to code for mitochondrially localized products are represented in the two extreme values of the MFI curve.

Our extensive global study shows that more than 88 genes coding for mitochondrial products exhibit an MFI value over 0.15. This means that the corresponding mRNA are, in normal growth conditions in galactose, more prone to be in found in polysomes associated with mitochondria. We addressed the specific features shared by this important mRNA class. It should be kept in mind that for the six cases we examined, the 3' untranslated region (3' UTR) sequence of these mRNAs contains the mitochondrial addressing information that can target, in vivo, a tagged version of this RNA fragment 14. Significantly, this in vivo targeting process did not require any translational activity as it was observed with non-translatable RNAs produced by RNA polymerase III. If this observation might be extended to the whole class of mRNAs that are translocated to the vicinity of mitochondria, it does not explain the raison d'être of this targeting process. To address this question we systematically analyzed the influence of several other parameters on mRNA localization (codon bias, codon adaptation index, amino-acid content, presence and length of the mitochondrial presequence as predicted by mitoprot or targetP, intra-mitochondrial protein localization, 3' UTR M-FOLD structure, and so on). Two characteristics - the phylogenetic origin and the length of the genes - turned out to correlate with the mRNA localization properties.

By making batches of alignments, Karlberg et al. 22 classified the yeast mitochondrial proteome on the basis of each gene's evolutionary origin. The main categories they considered were: ABE (yeast genes having at least one homolog in Archaea, one homolog in Bacteria and one homolog in the eukaryotes); BE (at least one homolog in both a bacterium and a eukaryote); E (at least one homolog in a eukaryote); and SC (genes specific to S. cerevisiae). They concluded that the yeast mitochondrial proteome has a dual origin, around half of it having originated from the endosymbiont genome, while the other part evolved from the nuclear genome.

Our previous study had shown that there was indeed a link between the eukaryotic/prokaryotic origin of a mitochondrial gene and the localization of the corresponding transcript. We found a statistically significant relation when we applied a one-way analysis of variance (ANOVA) to explain 358 MFI values by four categorical evolutionary origin values. To further characterize this relation, a multiple comparison among the four groups was done. Figure 5 highlights the results of that more detailed analysis, with 99% confidence intervals for each category's mean MFI. This reveals significant and ordered differences, with (ABE, BE) having more mitochondrial-membrane-localized translation than (E, SC). Although the results are not amenable to experimental test, this statistical observation might help in better understanding the cooperation between the nuclear and the organellar genomes. Interestingly, like the few remaining mitochondrially encoded genes translated near the inner membrane, the nuclear genes coming from the mitochondrial genome tend to be translated in the vicinity of the outer mitochondrial membrane.

Concerning the length of the proteins, which in yeast is closely connected to the length of the mRNA, we found a surprising statistical relation between mRNA length and mRNA localization; longer mRNAs tend to be localized to the vicinity of mitochondria (correlation coefficient: 0.67; see Figure 6).

George et al. 23 recently proposed a molecular mechanism and a model for the coupling of translation and translocation of mitochondrial mRNAs in vivo. Their model is based on the experimental observation that the nascent-polypeptide-associated complex (NAC) promotes interaction of the ribosome with the mitochondrial surface in vivo. Lithgow's two-step model proposes that in an early phase most mitochondrial mRNAs begin to be translated in the cytosol, whereas in the late phase, when a ribosome eventually encounters the surface of a mitochondrion, the mRNA molecule remains 'stuck' to the mitochondrion until its translation is completed. On average, longer mRNAs have more ribosome-NAC complexes, which results in a higher affinity for mitochondrial outer membrane. Interestingly, in a quite different context, flock-house virus mRNA polymerase is translated on Drosophila mitochondria while its amino-terminal sequence is inserted into the mitochondrial membrane 24. That example shows that the existence of a mitochondrial targeting sequence (mts) and a mitochondria-localized mRNA is not contradictory. In addition to the coordinated action of NAC and mts, in vivo RNA targeting analyses have shown that the 3' UTR contains sufficient information to address the corresponding RNA to the vicinity of mitochondria 14.

This means that several mechanisms can control the mRNA targeting process. Some of these are essential; for instance, we recently obtained direct evidence for the genes ATP2 and ATM1 that the 3' UTR sequence is necessary for functional mitochondria. This suggests that, in the case of these two genes, the mRNA targeting process linked to the 3' UTR signal is required for mitochondrial biogenesis 25. In addition, the initiation of translation of many mRNAs is, at least in part, controlled by their 3' UTR structure. Thus, early translation cis (mts, 3' UTR) and trans (NAC) factors act synergistically to control the targeting of many mRNAs coding for mitochondrial products. It has recently been demonstrated that ribosomes can 'sense' features of nascent polypeptide chains and modulate translation accordingly. Our results suggest that the mts and the 3' UTR could be such features that control mRNA translation in connection with the mitochondrial import process 26.

What could be the role of the mRNA targeting process in the biogenesis of mitochondria? Our favourite hypothesis is to imagine that it might be part of the control of the time course of production of mitochondrial proteins. One can reasonably propose that assembly of large mitochondrial complexes requires a temporally controlled import of the different subunits. Our observation would imply that the largest subunits are more likely to be the basic elements of the architecture of the mitochondrial complexes. This assumption should be amenable to experimental assessment. This mitochondrial mRNA targeting process is not specific to S. cerevisiae. Several studies in various human cells have shown that some specific mRNAs can be found next to mitochondria 2728, which strongly suggests that our observation can be applied to many organisms.

Saccharomyces cerevisiae strain CW04 (isonuclear to W303-1B: MATa, leu2-3, ura3-1, trp1-1, ade1-2, his3-11, can1-100) was used. Yeast was grown in a galactose-rich medium (1% bacto-yeast extract, 1% bactopeptone, 0.1% KH₂PO₄, 0.12% (NH₄)₂SO₄, 2% galactose).

Mitochondria-bound polysomes and free cytoplasmic polysomes were purified as described by Ades and Butow 30, with few modifications. To convert cells to spheroplasts, they were treated for 60 min with Quantazyme (1,000 U per g per wet weight cells) (Quantum Appligene, Illkirch, France) at 30°C with shaking. Spheroplasts were allowed to regenerate for 3 h with gentle shaking at 28°C in 1 M sorbitol, 1% bactopeptone, 1% yeast extract, 0.1% KH₂PO₄, 0.12% (NH4)₂SO₄ and 2% galacotose. Regeneration was stopped by adding cycloheximide to a concentration of 300 μ/ml and 20 min later the culture was poured over frozen 1 M sorbitol, 200 μ/ml cycloheximide. The spheroplasts were harvested and washed with 1 M sorbitol, 200 μ/ml cycloheximide. One-twentieth of the spheroplast volume was frozen at -80°C. The rest of the spheroplasts were homogenized in 6 ml of the following breaking buffer: 0.6 M mannitol, 30 mM Tris-HCl pH 7.4, 100 mM KCl, 5 mM magnesium acetate, 5 mM 2-mercaptoethanol, 200 μ/ml cycloheximide, 2,000 U/g yeast protease inhibitor cocktail (Sigma) 10,000 units RNAsin ribonuclease inhibitor (Promega, Madison, WI) and 500 μ/ml heparin. Cells were disrupted with 15 strokes using a glass Teflon pestle. The homogenate was centrifuged at 4,500 rpm for 10 min; the pellet was once again disrupted and centrifuged. Both homogenates were combined and centrifuged twice to eliminate unbroken cells and nuclei. The resulting supernatant was further centrifuged at 11,000 rpm for 30 min. The pellet of crude mitochondria associated with polysomes was washed twice in mannitol-polysome buffer before storage at -80°C. Free cytoplasmic polysomes were purified from the post-mitochondrial supernatant by sedimentation through a step gradient of 2 M and 0.5 M sucrose, each of these solutions prepared in 30 mM Tris-HCl pH 7.4, 100 mM KCl, 5 mM magnesium acetate, 5 mM 2-mercaptoethanol, 200 μ/ml cycloheximide, 600 U/g yeast protease inhibitor cocktail (Sigma) 2,500 units RNAsin ribonuclease inhibitor (Promega) and 500 μ/ml heparin. The gradients were centrifuged at 40,000 rpm for 20 h at 4°C using the TST41 Ti rotor. The pellet, containing free cytoplasmic polysomes, was used for RNA extraction.

RNA was extracted from spheroplasts, mitochondria-bound polysomes and free cytoplasmic polysomes using the RNeasy Protect MIDI kit (Qiagen, Valencia, CA).

Detailed information on this aspect of the experiments can be found in Additional data files and at 29 All the raw-values and pictures (.cel and .dat files) are available for download. For each of the six arrays, mRNA quality was controlled (Agilent Bioanalyzer; Agilent, Palo Alto, CA) before hybridization. Data analysis was carried out using the Affy Package (L. Gautier and R. Izarry, unpublished), which is part of Bioconductor R bundle. Classical normalization procedures, built on the assumption that most genes expression values do not vary across arrays, could not be applied for all six arrays. Therefore, arrays were normalized on a pairwise basis using the Bolstad method: MP₁ with MP₂; FP₁ with FP₂ and TOT₁with TOT₂. The Bolstad method averages independently perfect match and mismatch signal intensity values among chips at each quantile level, which reshapes the entire intensity distributions to make them comparable. Quantile normalization tends to align intensity boxplots and flatten MvA plots. We then checked whether spike mRNAs gave similar results for the six chips; as this was almost the case, no further correction was applied. Expression values were then computed using Affymetrix's playerout algorithm. Genes (3% of total) that had at least one of the six intensity values under an arbitrary threshold were discarded from further analysis. Interestingly, when conducted with MAS5.0, the analysis gave the same kinds of results, but with larger variability.

We used standard one-way analysis of variance. Multiple comparison analysis was then applied using the Tukey-Kramer test with α = 1%. This test is known to be somewhat conservative for unequally sized subpopulations. The Bonferroni test gave the same results (ABE, BE) > (E, SC).