Adult male-specific pathogen-free Hooded Wistar (12- to 13- weeks old) and Long Evans (7–8 weeks old) rats were procured from Animal Resources Centre (Western Australia, Australia). Hooded Wistar rats were group housed (3–4 rats/cage) in large, open-top polypropylene basin cages (56.5 × 38.5 × 19.5 cm, l × w × h) while Long Evans rats were pair-housed in standard open-top plastic cages (38 × 27 × 15 cm, l × w × h). All rats were provided with standard rat chow (Barastoc, Ridley Corporation, VIC, Australia) and tap water ad libitum, wood shavings and shredded paper as bedding, and maintained under controlled temperature (23 ± 1°C) and lighting (reversed 12:12 h light: dark cycle; lights off at 1000 or 1,100 h) conditions. All procedures were conducted in accordance with the National Health and Medical Research Council of Australia Code of Practice for the Care of Experimental Animals and received by the RMIT University Animal Ethics Committee (approval number 1402) and La Trobe University Ethics Committee (approval number 18–15).

Following a 2- to 3-week acclimation period, rats of each strain were randomly allocated into two treatment groups (short-term used Long Evans n = 14/group; long-term used Hooded Wistar n = 7–8/group): control and calorie restriction (CR). Controls were allowed ad libitum access to food throughout experimentation. The CR groups received 75% (25% restriction) of the amount of food consumed by the age- and strain-matched control rats, delivered daily within the hour before lights out (1000–1,100 h). Food intake of the CR group was initially determined by calculating the food intake of all rats across the last 48-h period of acclimation, and thereafter over 48-h every month. Food consumption of the control groups of both strains was consistent throughout experimentation with daily consumption ranging from 20 to 25 g per rat, equating to a mean daily food intake ranging from 15 to 19 g per rat for the CR groups. Long Evans rats underwent short-term CR, with CR extending for a total of 2.5 months, while the Hooded Wistar rats underwent long-term CR, with CR extending for 15 months. CR was initiated at approximately 3 months of age for both short-term and long-term CR rats. See Supplementary material S1 for body weight data across the duration of the experiments.

In the long-term CR experiment, behavioral testing in the elevated plus maze (EPM) and open field (OF) test occurred at 6, 12, and 18 months of age, corresponding to 4, 8, and 14 months of CR. These data were recently reported (Govic et al., 2022) and were used for comparison with the performed short-term CR in this study. Animals from the long-term CR study also underwent testing for acoustic startle reflex at 12 months of age. In the short-term CR experiment, behavioral testing in the EPM occurred 4 weeks following the initiation of CR. All animals were tested during the active dark portion of the light: dark cycle, approximately 1 h after lights-off and 2 h after the provision of food. To reduce the impact of circadian patterns on behavior, testing did not extend beyond the first half of the dark cycle and was preceded with a 30-min acclimation period to the testing conditions. Mazes/boxes were cleaned with 70% ethanol between tests. A closed-circuit camera mounted above the EPM allowed behavior to be recorded and tracked with Ethovision XT (Noldus, SDR Clinical Tech, Middle Cove, NSW, Australia) ethological tracking software which was operated in an adjacent room. While the individuals conducting the testing were not blind to group allocation, data collection and analysis via Ethovision were performed in a blind manner. Distance travelled (cm) was additionally calculated as an index of general locomotor behavior.

Following 2.5 months of CR (6 months of age) for the short-term CR group and 15 months of CR (at 18 months of age) for the long-term CR group, rats were humanely euthanized via an overdose of pentobarbital sodium (short-term study) or carbon dioxide asphyxiation (long-term study) and rapidly decapitated with the aid of a guillotine 2–4 h after lights-out. The hypothalamus, amygdala, pituitary, and prefrontal cortex (short-term CR only) were rapidly dissected out with the aid of a rat brain atlas and 1 mm coronal brain block (Braintree Scientific, Braintree, MA). Adrenal glands of the rats undergoing long-term CR were collected instead of the prefrontal cortex. Adrenal glands were collected from the short-term CR animals but due to methodological reasons were unable to be processed. Tissues were immediately flash frozen or stored in the Allprotect® Tissue Reagent (Qiagen, Hilden, Germany) and subsequently placed into a − 80°C freezer.

Total RNA was extracted from the rat brain tissue and adrenals (n = 5/region/group/strain) using the E.Z.N.A.® DNA/RNA Isolation Kit (Omega Bio-tek Inc., Georgia, USA) according to the manufacturer’s protocol. The nanodrop 2000c spectrophotometer (Thermo Scientific Inc., Waltham, MA, USA) was used to determine total RNA concentration, while integrity was assessed using the Agilent 2,200 TapeStation instrument (Agilent Technologies, Santa Clara, CA, USA). The TruSeq Stranded mRNA Library Prep Kit (Illumina, Inc., San Diego, CA, USA) was used to generate RNA sequencing libraries. The libraries were quantified and qualified using the High Sensitivity D1000 Screen Tape on an Agilent 2,200 TapeStation instrument. The libraries were normalized, pooled, and subjected to cluster and paired-end sequencing was performed for 150 cycles on a HiSeqX10 instrument (Illumina, Inc. San Diego, CA, USA) according to the manufacturer’s instructions.

The analysis of the generated sequencing reads was performed using CLC Genomics Workbench package 22 (QIAGEN) (Liu and Di, 2020) and Galaxy Australia¹ (Jalili et al., 2020), including quality control of sequencing reads, trimming, mapping, and finding the differentially expressed genes. Rat reference genome and its annotation (Rattus norvegicus.mRatBN7.2) were downloaded from the Ensembl genome browser² and used for mapping and expression analysis. Mapping was performed based on the following parameters: mismatch cost = 2, insertion cost = 3, deletion cost = 3, minimum length fraction = 0.8, and minimum similarity fraction = 0.8. The generalized linear model (GLM) based on negative binomial distribution (Robinson et al., 2010) was employed for differential expression analysis. The p-values were also corrected with false discovery rate (FDR) for multiple testing. The use of the GLM allows for curves to be fit to expression values without assuming that the error on the values is normally distributed. Fold changes were calculated from the GLM, which corrects for differences in library size between the samples and the effects of confounding factors. The Wald test was applied to calculate the p-values and FDR p-values for comparison of all group pairs. p-values and FDR p-values (corrected) were used for the selection of genes with significant differential expression in comparison of short-term CR against the control group and long-term CR against the control group in each of the studied brain regions.

The ensemble learning approach, based on voting of attribute weighting models, was employed in this study for biomarker discovery. The seven attribute weighting (feature selection) algorithms, including weighting by Info Gain, Info Gain Ratio, Rule, chi squared, Gini Index, Uncertainty, and Relief were applied, as previously described (Govic et al., 2022). Then, the genes and the studied tissue were ranked based on the overall received weights from attribute weighting models. The ones with higher cumulative weights were selected as key responsive genes. Attribute weighting models were successful in the discovery of tissue-independent transcriptomic signatures of CR.

PCA analysis based on the correlation matrix using transcriptomic signatures of each of the short-term and long-term CR treatments in each of the studied tissues was performed. The PCA analysis and PCA plot were generated by the Minitab21 statistical package.³ The signature was defined based on the top 20 responding genes to each condition derived overall weights of attribute weighting models.

The top 20 genes selected by attribute weighting models as responding genes to short- and long-term CR were further classified based on protein classes, including protein kinase, protein phosphatase, receptor, RNA transcript, secretory protein, chromatin-associated protein, transcription factor, and transporter using gene ontology (GO) information, comparative GO, and the Pathway Studio tool (Ashburner et al., 2000; Fruzangohar et al., 2013; Consortium GO, 2019). All regulatory genes, secretory proteins, and extracellular biomarker genes that were identified in the top 20 were reported. We then employed literature mining by MedScan (Novichkova et al., 2003), a natural language processing (NLP) implemented in the Pathway Studio webtool (Elsevier) (Nikitin et al., 2003), to shed light on the relations between responding genes in the transcriptomic signature of short- and long-term CR and anxiety, depression, aging, and HPA axis from full texts of published articles, as previously described (Alanazi et al., 2018; Mohammadi-Dehcheshmeh et al., 2021; Govic et al., 2022). The confidence index was prefiltered by the software to note biological facts supported by >1 independent sentence(s) within the literature, with >3 being a high confidence, 2, medium confidence and 1, low confidence. Furthermore, the sentences detected by the software were manually checked to ensure the accuracy of the relation. In the field of literature/text mining, the robustness and confidence of a particular interaction is often assessed by examining the number of references, that correspond to the number of sentences in published resources, supporting that interaction. This approach has been applied by Farhadian et al. (2018) and Govic et al. (2022).

Literature mining was applied to find the possible link between the available drugs and the responding genes to short- and long-term CR in this study. This signature was utilized to identify the optimal drug combination (drug repurposing) that mimics the effect of CR on the transcriptome.

As shown in Figure 1, short-term CR increased the proportional frequency of entries into the open arms relative to the closed arms [E_M = 28.82, (−5.62, 60.44), D_p = 95%, ROPE_p < 0.01] despite no difference in total distance travelled [E_M = 12.13, (−223.5, 240.56), D_p = 54%, ROPE_p > 0.99], indicating reduced anxiety-like behavior by short-term CR. Null hypothesis testing obtained the same result (Supplementary material S2). Anxiety-like behavior results for animals involved in the long-term experiment were recently published by Govic et al. (2022).

Supplementary material S3 provides the results of running seven attribute weighting models in ranking of the genes in response to short and long-term CR. Transcriptomic data for long-term CR animals has been previously published (Govic et al., 2022) but were here categorized by protein class (transcription factor, transporter, etc.) and compared to short-term CR data.

The overall (sum of the weights) ranks the genes in discriminating either short-term CR or long-term-CR from controls. The employed weighting models have the capability to analyze the effect of categorical variables of the tissue along with the numerical data of gene expression. The tissue received low overall weights in both short-term and long-term experiments demonstrating the possibility of the development of a tissue-independent signature.

As it can be inferred from PCA analysis in Supplementary material S4, the developed short-term and long-term signatures, constructed from the top 20 responsive genes, were successful in distinguishing samples under short-term and long-term CR treatments from control samples. For example, in the prefrontal cortex, the short-term CR signature segregates CR samples from controls efficiently, accounting for 64.2% of data variation, respectively. In contrast, for long-term CR, PCA1 distinctly separated treated samples in both the amygdala and pituitary glands accounting for more than 68% of variation in the data. There is no similarity between the top 20 high-ranked genes between short and long-term CR (Supplementary material S3). It can therefore be concluded that there were ongoing transcriptomic changes with long-term CR beyond the short-term effects. The genes encoding key protein classes (transcription factors, ligands, receptors, kinases, secretory proteins, chromatin-associated proteins, and transporters) in studied tissues can be seen in Table 1 and Figure 2. Note that Figures 2F,G,J show reproductions from Govic et al. (2022) for purposes of comparison.

Transcription factors ZNF45, CRY2, EID1, and CARHSP1, and the receptor HLA-A were key regulatory proteins identified in short-term CR tissues (hypothalamus, amygdala, pituitary, and prefrontal cortex). As can be inferred from Figure 2, the transcriptome response to long-term CR is more uniform than the response to short-term CR. All four regulatory proteins, ZBTB2 (transcription factor), MCOLN1 (transporter), MAP4K2 (kinase), and SAP18 (component of the histone deacetylase complex) were upregulated in response to long-term CR in the studied tissues (hypothalamus, amygdala, pituitary, and adrenal glands) as compared to the short-term genes, which were either up or downregulated depending on the tissue type. For long-term CR, in addition to these regulatory genes, C1QA, a secretory protein, was identified as being downregulated in all tissues except for the hypothalamus (which demonstrated a trend in this direction), as previously described (Govic et al., 2022).

Based on literature mining, possible interactions of top responding transcription factors and receptors to short-term CR with neurodegeneration, anxiety, aging, and fertility are visualized in Figure 3. The underpinning mined sentences from literature and relationships are provided in Supplementary material S5. As mentioned above, the number of mined sentences in references that support a particular relationship is an index of confidence level where 3, 2, and 1 stand for high, medium, and low confidence, respectively.

HLA-A receptor, one of three genes encoding major histocompatibility complex class 1 (MHC1) and located on the cell membrane, is the hub gene in the regulatory network with many documented interactions with neurodegeneration, fertility, stress, and nervous system. While no interactions were observed with anxiety, HLA-A was found to have a positive regulatory role with aging, yielding a high confidence score (five mined sentences).

CRY2, a clock gene involved in circadian rhythms, is an interesting hub in the network that upregulates in response to depression, an affective disorder with high comorbidity with anxiety (Olfson et al., 2017). Interestingly, a negative regulatory role was demonstrated between CRY2 and obesity as well as an association between mutations in CRY2 and risk of developing obesity.

EID1, which represses transcription and regulates cell cycle and differentiation, plays a negative regulatory role in obesity and a positive regulatory role in neurodegeneration. Interestingly, there is no report on the top responding gene to short-term CR, ZNF45 transcription factor, that opens a new avenue for further investigation and understanding of the regulatory mechanism of short-term CR at the transcriptomic level. Similarly, there were no interactions observed for CARHSP1 in the network.

Figure 4 and Supplementary material S6 present the interaction of key classes of long-term responding genes as well as C1QA with anxiety, aging, and neurodegeneration. C1QA was identified as a hub in the regulatory network linking this gene with anxiety, depression, stress, aging, and neurodegeneration, as previously described (Govic et al., 2022). Our understanding of regulatory mechanisms of long-term CR is still in its infancy. MCOLN1 receptor, located on Golgi, is the only regulatory long-term responding gene linked to neurodevelopmental disorder and the nervous system.

The results of drug repurposing in short and long-term CR are presented in Figures 5, 6 and Supplementary materials S7, S8, respectively.

For the short-term CR drug repurposing analysis, four of the regulatory genes from the attribute weighting analysis—EID1, HLA-A, CARHSP1, and CRY2—were identified as having interactions with known drugs (Figure 5 and Supplementary material S7). Only one drug was found to interact with EID1 and CARHSP1, tamoxifen and tacrolimus, respectively, both demonstrating negative regulation. This was in contrast to HLA-A and CRY2, with multiple drugs showing interactions with these identified short-term CR genes. Some drugs, such as rapamycin, valproic acid, doxycycline, and dexamethasone, demonstrated various interactions (expression, positive/negative expression, and positive molecular transport) with HLA-A and CRY2. For example, valproic acid was negatively associated with the expression of CRY2 (6 mined sentences, a high confidence score), where the same drug was positively associated with the expression of HLA-A (1 mined sentence, a low confidence score).

Regarding the long-term drug repurposing analysis, four of the top-responding genes from the attribute weighting analysis (PABPN1, MAP4K2, PLCG, and COQ2) as well as two regulatory genes identified in the class-based analysis (MCOLN1 and C1QA) were found to have interactions with known drugs (Figure 6 and Supplementary material S8). Multiple and various interactions, such as positive/negative regulation and positive/negative expression, were demonstrated between known drugs and PLCG1 and C1QA, while fewer interactions were demonstrated with PABPN1, MAP4K2, COQ2, and MCOLN1. Drugs such as doxycycline, resveratrol, tetracycline, dexamethasone, quercetin, and forskolin were found to interact with two to three of the long-term CR signature genes. For example, dexamethasone was positively related to the expression of C1QA (4 mined sentences and a high confidence score) and a negative regulatory relationship was identified between this drug and PLCG1 (10 mined sentences and a high confidence score).

Interestingly, doxycycline and dexamethasone have interactions with genes in both short- and long-term CR (Figures 5, 6 and Supplementary materials S7, S8).