Plants of Arabidopsis thaliana accession Wassilewskija-4 and a homozygote T-DNA insertion knockout of the GPT2 gene (gpt2, FLAG_326E03; INRA, Versailles, France), were used in all experiments. Plants were grown for 8 weeks under an 8 h day (20°C day/16°C night) at a light intensity of 100 μmol m^-2 s^-1 (LL) under warm white LEDs (color temperature 2800–3200 K), and then transferred to an irradiance of 400 μmol m^-2 s^-1 (HL) for 7 days. All plants were harvested on Day 7 HL, controls were maintained at LL and harvested on the same day. Plants were harvested at the end of the photoperiod, with leaves identified as fully developed prior to treatment being flash frozen in liquid nitrogen directly from growth conditions.

For all measurements of gas exchange, a CIRAS1 infrared gas analyser (PP Systems, Amesbury, MA, United States) was used. All measurements were made at a CO₂ concentration of 2000 μl l^-1 at an irradiance of 2000 μmol m^-2 s^-1. Measurements of chlorophyll fluorescence were performed using a PAM-101 chlorophyll fluorimeter (Heinz Walz, Effeltrich, Germany) with data recorded with a National Instruments PCI-6220 analog to digital convertor, using laboratory written software. Fluorescence parameters were calculated as described by Maxwell and Johnson (2000). P700 redox state was measured using a Walz PAM 101 in combination with an ED-P700DW-E emitter-detector unit (Heinz Walz; Hald et al., 2008). Actinic and saturating flash light was provided by a LED Engin LZ4 warm white LED (LED Engin, San Jose, CA, United States).

Frozen leaf samples were ground to a fine powder using a pestle and mortar under liquid nitrogen. For total protein content measurements, ground leaf was normalized on a fresh weight basis and Bradford reagent (Sigma–Aldrich, St. Louis, MO, United States) was used according to the manufacturer’s instructions for estimates of total leaf protein. For LCMS analysis, five replicates were used per condition, and 20 mg FW of ground leaf was added to 200 μl of 1% Rapigest (Waters, Milford, MA, United States) in 25 mM ammonium bicarbonate. Samples were then denatured at 80°C for 10 min. A 10 μl aliquot of this sample was reduced and alkylated by dithiothreitol and iodoacetamide. Samples were then digested overnight using proteomics grade trypsin (Sigma–Aldrich) at 37°C. Rapigest was removed by the addition of trifluoroacetic acid to 1% final concentration, and incubated for 2 h at 37°C, followed by 2 h at 4°C. Samples were centrifuged at 14,000 g for 15 min and the supernatant was desalted using POROS R3 beads.

Digested samples were analyzed by LC-MS/MS using an UltiMate^® 3000 Rapid Separation LC (RSLC, Dionex Corporation, Sunnyvale, CA, United States) coupled to an Orbitrap Elite mass spectrometer (Thermo Fisher Scientific, MA, United States). Peptide mixtures were separated using a gradient starting with a mixture of two solutions: 92% of 0.1% Formic acid (FA) in water and 8% of 0.1% FA in acetonitrile, increasing up to 33% FA in acetonitrile. Run time was a total of 180 min at a flow rate of 300 nL min^-1. A 75 mm × 250 μm i.d. 1.7 μM BEH C18 analytical column (Waters) was used for separation. Peptides were selected for fragmentation automatically by data dependent analysis.

Raw data were imported into Progenesis QI (build 2.0.5556.29015; Nonlinear Dynamics, Newcastle, United Kingdom) and runs were aligned according to the default settings. Only ions with a charge state of up to +4 were considered. MS/MS data were searched against the A. thaliana TAIR 10 database and assigned to peptides using Mascot version 2.4.0 (Matrix Science, London, United Kingdom). A maximum of one missed cleavage (Trypsin) was permitted, with a peptide mass tolerance of 10 p.p.m. and an MS/MS tolerance of 0.5 Da. For a complete list of identified peptides, see Supplementary Data S2. Data were then re-imported into Progenesis to allow for assignment of proteins from peptide data.

Raw protein intensities were exported from Progenesis and, instead of normalizing all samples to a single run, a reference run was selected for each treatment based on the sample with the median total protein content for that treatment. Total protein for each sample was calculated by summing the intensities of all the quantified proteins (for further details see Supplementary Figure S2). All replicates in each condition were then normalized to the reference run for that treatment.

When a comparison of two conditions was made, a t-test was used, using the constraints detailed in the figure legends. When a comparison of more than two conditions was made, an ANOVA was used. Proteins were considered to have significantly changed in abundance when a P-value of <0.05 was reached, with a fold change of >1.2, unless otherwise stated. All data analysis was carried out in the R software package, except ANOVAs, which were performed in SPSS (IBM, Armonk, NY, United States). Exploratory hierarchical clustering (Figure 2) was done using Euclidean distance and the complete linkages method. For comparitive hierarchical clustering (Figure 6) and heatmap analysis, fold change data were calculated relative to the WT at LL and log₂ scaled. A heatmap was then generated using the heatmap.2 package in the R software, using the default settings.

The mass spectrometry proteomics data have been deposited to the ProteomeXchange Consortium via the PRIDE (Vizcaino et al., 2016) partner repository with the dataset identifier PXD006598 and 10.6019/PXD006598.

Plants were grown at LL (LL; 100 μmol m^-2 s^-1) for 8 weeks and then acclimated to high light (HL; 400 μmol m^-2 s^-1) for 7 days. This HL intensity was selected as, previously, this was shown to induce a non-stressing acclimation response (Yin and Johnson, 2000; Athanasiou, 2007). Leaves that were fully expanded before the HL treatment were used for analysis. This approach was taken to avoid examining developmental acclimation to HL, and to assess how leaves that developed at LL are able to alter their protein composition in response to an increase in light intensity (Athanasiou et al., 2010). Examination of leaf sections by microscopy confirmed that there were no obvious changes in leaf morphology induced by the HL treatment (Supplementary Figure S1). As plants of both genotypes acclimated to HL, the leaf area per unit fresh weight declined marginally, but significantly, by about 10% (Figure 1A). There was also a marked decrease in the specific leaf area, implying the accumulation of a substantial amount of dry matter in the leaves (Figure 1B).

Exposure of wild type plants to HL resulted in an increase in their maximum capacity for photosynthesis (P_max) per unit leaf weight, from 2.5 to 3.3 μmol CO₂ g^-1 hr^-1 (Figure 1C), consistent with previous results (Athanasiou et al., 2010). In contrast, plants lacking the glucose-6-phosphate translocator GPT2 were only able to reach a P_max of 2.5 in response to HL, similar to WT at LL (Figure 1C). Both WT and gpt2 leaves did, however, contain approximately 20% higher total protein content per unit fresh weight after HL acclimation (Figure 1D).

To define changes occurring in the leaf proteome in response to HL, we adopted a label-free proteomic approach. In order to maximize the reproducibility and breadth of coverage of the proteome, various methods of protein purification and separation were explored. The best results were found with a simple ‘in-solution’ digestion protocol. Leaf proteins were solubilised in a buffer containing Rapigest (Waters, Milford, MA, United States), before being digested with trypsin, desalted and analyzed using LCMS. This avoided the additional steps involved in gel-based fractionation or isobaric labeling and allowed the entire leaf proteome to be analyzed in a single MS run. A total of 3514 proteins were reproducibly identified on the basis of 36380 unique peptide IDs (see Supplementary Data S1, S2). We applied a strict filter to the dataset, so that proteins were only quantified when three or more non-redundant constituent peptides were reproducibly identified. With this filter, 1993 proteins per run could be quantified, which were involved in diverse metabolic processes (for a complete list of quantified proteins, see Supplementary Data S1).

One of the challenges of quantifying proteomes is to determine the basis for normalizing between samples. It is a common practice to assume that all samples have the same protein content per unit of material. Using the Progenesis default normalization method, it was found that in the WT, of the 1993 quantified proteins, 251 increased and 369 decreased in abundance. Proteins were considered to have significantly changed in abundance when a p-value of <0.05 or lower was reached, with a minimum fold change of 1.2, a level widely adopted in proteomics experiments (Nissom et al., 2006; Keenan et al., 2009; Serang et al., 2013; Zhang et al., 2016). In the present case, however, using such a normalization arguably gives misleading results. For example, on this basis, we would have to conclude that investment in electron transport and carbon fixation were unchanged or declined in leaves acclimated to HL. This does not reflect the increase in P_max (Figure 1C) or observations from western blotting of RBCS, which showed an increase per unit leaf weight in WT plants (data not shown). Furthermore, this normalization method showed no increase in total protein after acclimation to HL (calculated by summing all protein quantified proteins), in contrast to results from the Bradford assay, which showed a 1.2-fold increase (Figure 1D).

An assumption of the Progenesis default normalization is that the majority of detected features (peptides) do not change in abundance between samples, and thus these features can be used for normalization. A reference sample is then selected and all samples are normalized according to the intensity of the selected normalization vectors. In essence, this approach is an automated method of selecting a number of housekeeping peptide ions that can be used for normalization. Under our experimental conditions, however, this assumption was violated, as the increase in total protein content under HL (Figure 1D) resulted in an increase in many of the detected features, leading to the erroneous results as described in the previous paragraph. Therefore, a normalization method based upon the median total protein content in each experimental group was used (for further details, see Materials and Methods and Supplementary Figure S2). After normalization, a large proportion of the proteome showed a significant increase in abundance after acclimation, consistent with the increase in total leaf protein (Figure 1D). For WT and gpt2 plants, 1284 and 1119 proteins were significantly increased in abundance, respectively, with only 14 proteins being significantly decreased for both genotypes. 831 proteins were increased in both WT and gpt2.

Amongst the quantified proteins, extremely good coverage of central carbon metabolism was achieved (for further details see Supplementary Figure S3). This included quantitation of components of every major protein complex of the photosynthetic electron transport chain and every enzymatic step of the Benson-Calvin cycle. Furthermore, we were able to quantify the majority of enzymes from sucrose and starch metabolism, glycolysis and respiration. Many proteins involved in other key metabolic pathways were also quantified, including amino acid, lipid, tetrapyrrole and antioxidant metabolism. The majority of identified proteins were easily solubilised or abundant enzymes, while there was an under representation of membrane proteins.

For initial data analysis, hierarchical clustering was performed (Figure 2). LL samples were clearly separated from HL samples, indicating that the light treatment caused a change in the proteome of both WT and gpt2 plants and that the effect of the treatment was stronger than the effect of the genotype. When the LL cluster is examined, there is only poor discrimination between WT and gpt2, suggesting the LL proteomes are rather similar between genotypes. However, following acclimation to HL, the genotypes separate into distinct clusters, indicating different responses are occurring.

The proteomic analysis provided good coverage of the protein complexes of the thylakoid membrane. Figure 3 summarizes how different gene products associated with the light reactions respond to HL in both WT and gpt2 plants. All proteins involved in core PSII function (PSBA-F) were detected, and most PSI core proteins (PSAA-F). Even highly homologous isoforms of the same polypeptide could, in some cases, be discriminated (e.g., PSBO1 and PSBO2, which share 95% amino acid homology; Dasgupta et al., 2005). The analysis also provided good coverage of the rest of photosynthetic electron transfer, including quantification of every subunit from the CF1 domain of the chloroplast ATP synthase, both plastocyanin and photosynthetic FNR isoforms, and most, but not all, subunits of the cytochrome (Cyt) b₆f complex.

Amongst the proteins involved in light capture, similar trends were seen in the two genotypes. The majority of detected PSII antenna proteins showed no significant change in response to HL, however, specific proteins did change. LHCB 4.3 (AT2G40100) showed a more than threefold increase in both genotypes, while LHCB 1.4 (AT2G34430) was decreased by 50% in WT plants but only 30% in gpt2 plants. There was also a significant increase in LHCA 5 (AT1G45474) in both genotypes. Thus, these results suggest a reorganization of the light harvesting proteins is occurring in both genotypes.

In the WT, there were also increases of around 50% in the protein NPQ4 (PSBS; AT1G44575), which is required for non-photochemical quenching. A similar increase was also seen in NPQ1 (violaxanthin deepoxidase). In gpt2 plants, NPQ4 behaved in a similar manner to the WT, however, NPQ1 was not significantly altered in response to HL.

Although a large proportion of the proteome was increased in absolute terms in response to HL, the majority of PSII core proteins showed no change in abundance (including PSBB-D). PSBA (ATCG00020) did show a small but significant (1.25-fold) increase in WT plants. In gpt2, PSBA was detected at a higher level than WT at LL but did not change significantly in response to HL. Assuming that these proteins are only functional as part of a complex, it is unlikely that an increase in a single subunit reflects a change in the abundance of the reaction center complex. It was possible to detect isoforms of all proteins of the oxygen evolving complex of PSII (OEC), with both isoforms being quantified for PSBO and PSBQ. Amongst the quantified proteins, a change in the relative abundance of the different isoforms was seen. For example, there was an increase in PSBQ1 (AT4G21280), accompanied by a similar decrease in PSBQ2 (AT4G05180), suggesting that at HL, the Q1 isoform is favored. Similarly, for PSBO, we saw no increase in the O1 (AT5G66570) isoform, yet an increase in the O2 (AT3G50820) isoform. Both PSBP isoforms were detected, however, due to the high sequence homology between these isoforms, there were too few unique peptides to allow quantification of PSBP2. PSBP1 and PSBR were detected but showed no change in abundance. Overall, these results indicate a changing composition of the OEC upon HL acclimation. There is no clear evidence that total PSII abundance changes upon acclimation to HL in either genotype.

The only apparently increased protein detected in the PSI core complex was PSAA (ATCG00350) in the WT, by 30%.

When proteins involved in the photosynthetic electron transport chain (PETC) were examined, a clearer pattern of increased abundance was apparent. In WT there were significant increases in 2 of the 3 Cyt b₆f subunits, 1 plastocyanin isoform, both FNR isoforms, and all detected ATP synthase subunits. Although showing a similar trend to the WT, gpt2 plants increased less of the Cyt b₆f complex, with only one detected subunit being considered significantly altered, and no change in plastocyanin content being observed.

In order to determine whether the above conclusions, from proteomic analysis of electron transport proteins are consistent with behavior of the photosynthetic apparatus, chlorophyll fluorescence analysis was performed (Figure 4). The parameter ΦPSII gives an estimation of the proportion of absorbed light energy that is used for photochemistry (Genty et al., 1989). Under saturating CO₂ and light conditions in WT plants, ΦPSII increased by around 50% at HL (Figure 4A). This compares with an increase of approximately 30% in P_max, measured under the same conditions (Figure 1C). In gpt2 plants, ΦPSII increased to a lesser extent than in the WT. NPQ was found to not change significantly in response to HL in the WT, however, in gpt2, a significant reduction in NPQ was observed in HL acclimated leaves (Figure 4B).

Proteomic data suggest that changes are occurring in the relative abundance of electron transport proteins, with photosystems being unaltered whilst proteins responsible for intersystem electron transport tending to increase. This is predicted to affect the flux through the electron transport chain. Measurements of the reduction kinetics of PSI were conducted by measuring the oxidation state of the primary electron donor, P700 (Harbinson and Hedley, 1989; Ott et al., 1999). At 2000 μmol m^-2 s^-1 light, the degree of oxidation of P700 did not vary significantly between plants maintained at LL and those transferred to HL in either genotype (Figure 4C). P700 was more oxidized in the mutant, however, in both treatments. P700 re-reduction following a transition to darkness was fitted with a single exponential decay curve, yielding a pseudo-first order rate constant indicative of the conductance of the electron transport chain (Golding and Johnson, 2003). In WT plants, the rate constant in HL plants was 60% higher than in LL plants (Figure 4D), indicating an increased conductance of the PETC, consistent with the increase in the relative abundance of electron transport components. In gpt2 plants, there was no significant increase in the rate constant in HL-acclimated plants. Taken together these measurements indicate that, in the WT, the overall flux per PSI reaction center was greater in HL plants, whilst in the gpt2 mutant, no acclimation of electron transport occurred. Combined with ΦPSII data (Figure 4A), these data indicate that HL acclimated WT plants have a greater relative capacity for linear electron flow than gpt2, consistent with our interpretation of the proteomic data. The Cyt b₆f complex is the earliest part of the PETC where differences are apparent between WT and gpt2 plants at HL.

All of the detected Benson-Calvin cycle enzymes were increased in the WT at HL, by on average 50% (Figure 5), consistent with the increased capacity for carbon fixation (Figure 1C). While the majority of enzymes were also increased in gpt2, this increase relative to the control conditions was lower for many enzymes, in the range of 10–30%. Many of the more abundant Benson-Calvin cycle enzymes were quantified using nine or more peptides, allowing extremely robust quantitation, allowing even slight differences in abundance to be resolved. Surprisingly, the only enzyme that did not show differing abundance between the genotypes at HL was Rubisco, with a 75% increase in the large subunit (RBCL; ATCG00490) for both genotypes. For the Rubisco small subunit (RBCS), several different isoforms could be detected. When common peptides for all RBCS isoforms were used to calculate total RBCS abundance, there was no difference between WT and gpt2. However, it was also possible to detect enough unique peptides to quantify RBCS 1A (AT1G67090) and 3B (AT5G38410) individually. At HL, there was a difference in the RBCS isoform balance, with a similar increase in RBCS 1A in both WT and gpt2, but an increase RBCS 3B to a significantly higher level in the WT. Even though gpt2 plants are affected in their ability to increase their capacity for carbon fixation upon HL, overall Rubisco abundance increases in a manner similar to the WT, indicating that the abundance of this enzyme is not sufficient to explain the difference in P_max. The abundance of other Benson-Calvin cycle enzymes, in particular sedoheptulose 1,7-bisphosphatase (SBPase; AT3G55800) and fructose 1,6-bisphosphate aldolase (FBA) have previously been suggested to be important in determining P_max (Zhu et al., 2007). We were able to quantify SBPase and all three chloroplast localized FBA isoforms (FBA 1–3; Lu et al., 2012). Both SBPase and FBA1 (AT2G21330) were increased to a significantly lower level at HL in gpt2 compared to the WT.

In the WT, nearly all enzymes of starch and sucrose metabolism were increased in abundance, suggesting an increased capacity for flux through these pathways (see Supplementary Figure S4). A similar pattern of increased abundance was observed in gpt2 plants, however, as with the Benson-Calvin cycle, a number of enzymes were detected at a lower level at HL, relative to the WT. There were 13% lower levels of the starch synthesis enzymes STARCH SYNTHASE 1 (SS1; AT5G24300) and ADP-GLUCOSE PYROPHOSPHORYLASE (AGPase; AT5G19220) at HL. gpt2 plants also contained on average 24% lower levels of enzymes associated with starch degradation, including 2 β-amylases (BAM3; AT4G17090 and BAM6; AT2G32290), and DISPROPORTIONATING ENZYME (DPE1; AT5G64860). Overall, lower levels of starch metabolism enzymes in gpt2 plants, compared to WT at HL, suggests a lower capacity for this pathway.

Sucrose metabolism exhibits a similar pattern of increased abundance to starch metabolism under HL, as all of the detected enzymes in sucrose synthesis are increased in WT (Supplementary Figure S4). The same is true of gpt2 plants, with the exception of selected enzymes, including SUCROSE PHOSPHATE SYNTHASE 2F (SPS2F; AT5G11110). The sucrose degrading enzymes SUCROSE SYNTHASE 1 and 4 (SUS1; AT5G20830 and SUS4; AT3G43190) are also increased in WT plants and not in gpt2 plants at HL. It should, however, be noted that the levels of SUS1 and 4 are higher in gpt2 plants at LL than in the WT. Two key regulatory enzymes, HEXOKINASE 1 (HXK1; AT4G29130) and cytosolic fructose bisphosphatase (cFBPase; AT1G43670) were increased in both genotypes, but to significantly higher levels in WT than in gpt2 at HL. HXK1 acts both catalytically in sugar metabolism but also as a sensor of glucose status (Moore et al., 2003; Cho et al., 2006), while cFBPase acts in sucrose synthesis, and also as a fructose sensor (Cho and Yoo, 2011).

The majority of enzymes involved in glycolysis and the TCA cycle were increased in both genotypes at HL (Supplementary Figure S3). There were very few differences between the WT and gpt2 mutant in the enzymes of glycolysis at HL, and a similar pattern was also seen in the TCA cycle. Generally, glycolysis and the TCA cycle were covered well in this analysis, with at least one protein quantified for every enzymatic step of the cycle. The mitochondrial electron transport chain was less well covered; however, we were still able to quantify many of the complexes based on at least one protein. When respiration is considered as a whole, there is an increase in both WT and gpt2 upon HL.

In order to further investigate the differences between the WT and gpt2 acclimation responses, taking a less targeted approach, cluster analysis was performed using a subset of proteins which were most differentially expressed (p < 0.05, greater than 1.5-fold difference in abundance) between the genotypes at either LL, HL, or both (Figure 6). Relative abundance was expressed relative to the WT at LL, so that each comparison is normalized to the same reference. Six clusters were defined, which display differing behavior. For a full list of proteins belonging to each cluster, see Supplementary Data S3. Proteins in Clusters 1 and 3 could be generally characterized as having elevated abundance at LL in gpt2 but a similar pattern of increase in both genotypes at HL. A number of proteins involved in stress responses were contained within this group, including PATHOGENESIS-RELATED GENE 5 (PR5; AT1G75040), GLUTATHIONE S-TRANSFERASE 6 (GST6; AT1G02930), and HEAT SHOCK PROTEIN 60-2 (HSP 60-2; AT2G33210). GST6 was particularly strongly increased – 9-fold and 11-fold in gpt2 and WT, respectively. Proteins involved in the isoprenoid pigment biosynthesis pathway were also found in this cluster. Overall, this suggests that, although HL induces stress response proteins in both WT and gpt2 plants, these tend to be expressed more highly in gpt2 at both LL and HL, relative to the WT.

Cluster 2 contained proteins with a general trend toward being expressed more highly in gpt2. N-MYRISTOYL TRANSFERASE-1 (NMT1; AT5G57020), regulates protein activity by lipid modification of the N-terminus (N-myristoylation) of a small number of proteins, including the global metabolic regulator, SUCROSE NON-FERMENTING RELATED KINASE 1 (SnRK1; AT3G29160; Pierre et al., 2007). This protein doubled in abundance in response to HL in the WT, while it remained unchanged in gpt2. Also in this group was PLASTID MOVEMENT IMPAIRED 2 (PMI2; AT1G66840), which was unchanged in abundance in the WT, but constitutively increased in gpt2. PMI2 is involved in chloroplast photo-relocation as part of the HL avoidance strategy, suggesting this response in enhanced in gpt2.

Cluster 4 contained proteins which showed perhaps the most striking differences between WT and gpt2, with no change in response to HL in the WT, but abundance in gpt2 of on average 7-fold higher at LL, and 11-fold at HL, relative to the WT. This group included PLASTID MOVEMENT IMPAIRED 15 (PMI15; AT5G38150). PMI2 was also constitutively increased in gpt2, and has been shown to physically interact with PMI15 (Kodama et al., 2010) to mediate chloroplast photorelocation, suggesting an enhanced HL avoidance response in gpt2. L-GALACTONO-1,4-LACTONE DEHYDROGENASE (GLDH; AT3G47930), which catalyzes the final step of ascorbate biosynthesis was also found within this group, as was MYO-INOSITOL-1-PHOSPHATE SYNTHASE 1 (MIPS1; AT4G39800). Finally, SnRK1.2 was also found within Cluster 4, and was increased four and eightfold in gpt2 LL and gpt2 HL, respectively, relative to LL WT plants.

Only two proteins were found in Cluster 5, and these were substantially increased at HL, with both being involved in anthocyanin biosynthesis. CHALCONE ISOMERASE LIKE (CHIL; AT5G05270), was increased 13-fold in the WT, but 36-fold in gpt2. UDP-GLUCOSE:CYANIDIN 5-O-GLUCOSYLTRANSFERASE (5GT; AT4G14090), an anthocyanin glycosylating enzyme (Tohge et al., 2005), was expressed to a 4-fold higher level at LL in gpt2 relative to WT, but increased by around 35-fold in both genotypes at HL relative to the WT at LL.

When proteins considered within the cluster analysis are compared as a whole, some trends can be seen. Across all clusters, there are many proteins involved protein synthesis regulation, such as ribosomal proteins, tRNAs and chaperones. Regulatory proteins, such as tetratricopeptide repeat proteins and translation initiation factor proteins were also abundant (for a full list, see Supplementary Data S3). This data suggests a differential regulation of protein composition between WT and gpt2, and that post-transcriptional regulation may be important in explaining differences between WT and gpt2. Many proteins involved in stress responses, antioxidant synthesis and HL avoidance were expressed more highly in gpt2 at LL compared to the WT, suggesting GPT2 is important under control conditions. Furthermore, while some of these stress-related proteins were induced in both WT and gpt2 at HL, generally, the induction was stronger in gpt2.