From May to October 2023, samara development of A. truncatum was monitored from anthesis to physiological maturity. Samples were collected in Pingdi Township, Panzhou City, Guizhou Province, China at two defined developmental stages. After collection, samaras were dissected, and the pericarp was removed with tweezers and a dissecting needle to expose the kernel. At physiological maturity, the seed coat is brown and the kernel is bright yellow. Developmental stages were defined by days after anthesis (DAA) together with external seed traits. Kernels at 150 DAA were designated as the immature stage (IMS), and kernels at 180 DAA as the mature stage (MS). Three biological replicates were collected at each stage (n = 3), for a total of six samples: Y1-1, Y1-2, and Y1-3 (IMS) and Y2-1, Y2-2, and Y2-3 (MS). Excised kernels were snap-frozen in liquid nitrogen and stored at −80°C for fatty acid analysis and RNA extraction.

Total lipids were extracted from seed samples using a chloroform–methanol mixture (1:1, v/v), following the lipid profiling protocol established by Shui et al. [26]. Triacylglycerols (TAGs) and diacylglycerols (DAGs) were quantified using an Exion UPLC–QTRAP 6500 PLUS system (Sciex), which was equipped with an electrospray ionization (ESI) source. The operating conditions included a curtain gas flow of 20, an ion spray voltage of 5500 V, a source temperature of 400°C, and ion source gases 1 and 2 both set at 35. Neutral lipids were isolated through isocratic elution on a Phenomenex Kinetex C18 column (2.6 μm, 4.6 × 100 mm) using a chloroform–methanol–0.1 mol L⁻¹ ammonium acetate mixture (100:100:4, v/v/v) at a flow rate of 160 μL min⁻¹ for 20 min. The quantification of individual TAG and DAG molecular species was performed using neutral loss MS/MS, employing a series of deuterium-labelled internal standards, including d5-TAG (16:0)₃, d5-TAG (14:0)₃, d5-TAG (18:0)₃, d5-DAG (1,3-16:0), and d5-DAG (1,3-18:1).

Raw peak areas were normalized to the corresponding deuterium-labelled internal standards. Prior to multivariate analysis, lipid abundances were log10-transformed and Pareto-scaled. Unsupervised principal component analysis (PCA) was first used to evaluate sample clustering and detect outliers. Supervised orthogonal partial least squares-discriminant analysis (OPLS-DA) was then applied to assess group separation [27].

To reduce the risk of model overfitting, the OPLS-DA model was evaluated using cross-validation and permutation testing (n = 200 permutations). Model goodness-of-fit and predictability were reported as R2 and Q2, and the permutation-derived regression intercepts were used to judge whether the supervised model was robust. Differential lipids between developmental stages were determined by combining multivariate and univariate statistics. For each lipid, a two-sided Student’s t-test was performed, and p-values were adjusted for multiple testing using the Benjamini–Hochberg false discovery rate (FDR). Lipids with VIP > 1.0, |log2(fold change)| > 1, and FDR-adjusted p < 0.05 were defined as significantly differential [28].

Total RNA was isolated from each seed sample using the mirVana miRNA Isolation Kit (Ambion, Foster City, CA, USA), adhering to the manufacturer’s protocol. The quality of the RNA was initially assessed using 1% agarose gels, and the integrity of the samples was subsequently verified with an Agilent 2100 Bioanalyzer (Agilent Technologies, Santa Clara, CA, USA). For each library, 1.5 μg of high-quality RNA served as the starting material. Poly(A)+ mRNA was isolated from total RNA using oligo(dT)-conjugated magnetic beads and was then fragmented in the presence of divalent cations at an elevated temperature. First-strand cDNA synthesis was performed using random hexamer primers and M-MuLV reverse transcriptase (RNase H–), followed by second-strand cDNA generation utilizing DNA polymerase I in conjunction with RNase H. The resulting double-stranded cDNA fragments underwent end-repair to generate blunt ends, were adenylated at the 3′ termini, and were subsequently ligated to Illumina sequencing adapters. Adapter-ligated products of approximately 250–300 bp were selected using the AMPure XP system and amplified by PCR to yield the final libraries. The indexed libraries were pooled, loaded onto a flow cell for cluster generation, and sequenced on an Illumina HiSeq platform to obtain paired-end reads.

Raw sequencing reads in FASTQ format were initially subjected to quality control using custom Perl scripts. Sequences containing adapter remnants, poly-N stretches, and low-quality reads were filtered out to obtain a set of high-quality clean reads. For each library, standard quality metrics, including Q20, Q30, GC content, and sequence duplication rate, were calculated [29]. The clean reads were subsequently assembled de novo using Trinity. Within each Trinity component, isoforms assigned to the same locus were clustered, and the longest transcript in each cluster was retained as the representative unigene [29]. Functional annotation of unigenes was performed by comparing their sequences against several public databases, including the NCBI non-redundant protein (Nr) and nucleotide (Nt) databases, Pfam, KOG/COG, Swiss-Prot, KEGG Orthology (KO), and Gene Ontology (GO). BLASTx searches were conducted with an E-value threshold of 1.0 × 10⁻⁵, and the top hits were utilized to assign putative functions and metabolic pathway information to each unigene [29].

Transcript abundance was estimated from cleaned reads mapped back to the Trinity assembly, generating raw expected counts and TPM values for each unigene. TPM/FPKM values were used only for expression visualization. Differential expression between the immature stage (IMS) and mature stage (MS) was performed on raw counts using edgeR (negative binomial model) with TMM normalization and a quasi-likelihood framework. p-values were adjusted using the Benjamini–Hochberg false discovery rate (FDR), and genes with FDR < 0.05 and |log2(fold change)| > 1 were defined as differentially expressed [30,31]. Global expression patterns of DEGs were summarized by hierarchical clustering. Global expression patterns of DEGs were summarized by hierarchical clustering. GO enrichment was conducted using GOseq to correct for gene-length bias, and KEGG pathway enrichment was assessed using KOBAS [32].

To validate the RNA-seq results, ten differentially expressed unigenes were selected for quantitative RT-PCR analysis. Gene-specific primers were designed using the PrimerQuest Tool, and the resulting PCR products were sequenced to confirm amplicon specificity. For each sample, 1 μg of total RNA was treated with gDNA Eraser and subsequently reverse-transcribed into first-strand cDNA using the PrimeScript™ RT Reagent Kit (RR047A). Each qPCR reaction (20 μL) contained 10 μL of abm® EvaGreen qPCR MasterMix (no dye), 1 μL of cDNA template, 1 μL of forward primer, 1 μL of reverse primer, and 7 μL of double-distilled water. Three biological replicates were analyzed for each sample. The cycling protocol included an initial denaturation at 95°C for 10 min, followed by 40–45 cycles of denaturation at 94°C for 15 s and annealing/extension at 60°C for 1 min. Relative transcript levels were calculated using the 2^−ΔΔCt method based on the corresponding threshold cycle (Ct) values. The primer sequences used for qRT-PCR are listed in Table S1.

Lipid composition was profiled using liquid chromatography–mass spectrometry (LC–MS). To robustly detect shifts in lipid abundance and minimize false-positive and false-negative results, three independent pairwise comparisons between sample groups were conducted. Lipid species were annotated based on retention time and MS/MS spectra, referencing published data and the LIPID MAPS database (http://www.lipidmaps.org/). A total of 100 lipid molecular species were identified, comprising 66 triacylglycerols (TAGs) and 34 diacylglycerols (DAGs). The score plot from the orthogonal partial least-squares discriminant analysis (OPLS-DA) demonstrated tight clustering of samples within groups and clear separation between groups, indicating effective discrimination of the lipid profiles (Fig. 1a). Cross-validation was employed to assess the robustness of the OPLS-DA model, further supporting the reliability of the group separation (Fig. 1b).

Equations in display format should be separated from the surrounding text, aligned to the left of the cAs indicated by lipid-class-based clustering and heat maps, significant metabolic changes were observed among lipid species and in lipid composition (VIP > 1, p < 0.05) based on GC-MS analysis between the two stages (Fig. 2). In the mature stage (MS), the content of triacylglycerols (TAGs) increased markedly Compared to the immature stage (IMS), further confirming the pivotal role of TAGs in the accumulation of seed oil [13]. Most TAGs in the mature stage exhibited a significant increase relative to the immature stage. In contrast, the content of diacylglycerols (DAGs) showed minimal change, even demonstrating a slight downward trend. The lipid composition varies between the two stages, suggesting that there are different patterns of lipid metabolism in A. truncatum seeds during MS and IMS.

As illustrated in the heatmap, seed maturity did not significantly alter the abundance of various lipid species; however, the content of major triacylglycerol (TAG) species increased (Fig. 2). Further analysis using a volcano plot, with a threshold of p-value < 0.05 and |log2(fold change)| > 1 based on the Student’s t-test, revealed 30 lipidomic metabolites (18 TAGs and 12 diacylglycerols, DAGs) exhibiting significant differential expression (Fig. 3d). The variations in the three most significant metabolites between mature seeds (MS) and immature seeds (IMS), namely TAG54:6 (18:1) (Fig. 3a), TAG54:5 (18:3) (Fig. 3b), and TAG54:6 (18:3) (Fig. 3c), are presented in the box plots. Notably, TAGs represented the majority of lipids in MS, accounting for over 91% of the total lipid content (Fig. 3e). Enhanced lipid accumulation of TAGs compared to DAGs was observed at both developmental stages.

Lipidomics revealed complex specificity in glycerolipid remodeling during seed maturation. Compared with IMS, most triacylglycerols (TAGs) increased in MS (e.g., TAG54:3 and TAG54:5), whereas several TAG species decreased (e.g., TAG58:3). In parallel, many diacylglycerols (DAGs) showed lower relative abundance in MS, consistent with DAG functioning as a central intermediate in plant glycerolipid metabolism that can be rapidly converted to TAG (via DGAT/PDAT) and interconverted with phosphatidylcholine during acyl editing. Together, we conclude that distinct metabolic patterns exist between MS and IMS, with lipid synthesis and accumulation primarily occurring in MS.

After quality filtering, we obtained 55,724,490, 48,095,836 and 57,736,034 clean reads from the immature-seed libraries (Y1-1, Y1-2 and Y1-3), and 47,965,770, 59,159,334 and 56,282,996 clean reads from the mature-seed libraries (Y2-1, Y2-2 and Y2-3). For each library, the yield of clean bases exceeded 7.19 Gb. As summarised in Table 1, the average sequencing error rate was 0.03%, the mean GC content was 46.32%, and the Q20 and Q30 values exceeded 97.50% and 93.03%, respectively, supporting the use of these data in subsequent analyses.

Clean reads were assembled de novo using Trinity into 56,225 unigenes. Unigene lengths ranged from 301 bp to 47,817 bp, with a median of 1339 bp and an N50 of 2338 bp. Of these, 19,949 unigenes were 200–500 bp, 15,337 were 500 bp–1 kb, 10,215 were 1–2 kb, and 10,724 exceeded 2 kb.

The Nr, NT, Swiss-Prot, Pfam, KOG/COG, KEGG, GO databases were used to find all of the unigenes that were found. A total of 42,674 unigenes were annotated, representing 75.89% of the overall count. Among them, 5844 (10.39%) were annotated in all seven databases (Fig. 4a). A total of 11,161 unigenes (26.15%) demonstrated significant homology with sequences in the Nr database, with 13.3% matching Citrus clementina and 11.4% matching Citrus sinensis (see Additional file 1). Furthermore, the sequences from Fig. 4b showed that 7976 unigenes (18.9%) were substantially comparable to those in the KOG database.

According to GO annotation, 29,015 (67.99%) of the matched unigenes were classified into three functional categories: Biological Process, Cellular Component, and Molecular Function. As shown in the GO enrichment histogram, “cellular process” and “metabolic process” were the main subcategories in “Biological Process”. In the “Cellular Component” category, the subcategories “cell” and “cell part” showed the highest levels of enrichment. For “Molecular Function”, most unigenes were associated with “binding” and “catalytic activity” (Fig. 4c).

KEGG analysis, as a pathway-based approach, helps to better understand specific processes, gene functions, and gene interactions at the transcriptomic level. A total of 10,536 unigenes were mapped to 19 pathways, according to search results from the KEGG database. Unigenes from A. truncatum effectively represented pathways associated with lipid and carbohydrate metabolism (Fig. 4d).

Using the FPKM method, we calculated the expression of unigenes. Notably, we identified 3512 significant differentially expressed genes (DEGs) between the MS and IMS phases, with 2068 genes upregulated (greater than 2-fold) and 1444 genes downregulated. The remaining genes exhibited downregulation (less than 0.5-fold). This finding suggests that lipid accumulation during seed maturity is accompanied by broad transcriptional reprogramming between these two stages. The proportion of expressed genes during the MS stage appears to increase in tandem with seed maturation (Fig. 5a). Hierarchical clustering analysis segregated the DEGs from the two stages into two distinct clusters. Genes in Cluster 1 were highly expressed in both the MS and IMS phases, whereas Cluster 2, which comprised the majority of the DEGs, exhibited contrasting expression profiles (Fig. 5b).

To further investigate the functional roles of the DEGs, 3512 genes were assigned to pathways in the KEGG database and to Gene Ontology (GO) categories. In the KEGG analysis, 219 DEGs were significantly enriched in 13 pathways (Fig. 6a). Among these, 202 upregulated genes were associated with 13 pathways, including “Biosynthesis of Unsaturated Fatty Acids”, “Fatty Acid Biosynthesis” and “Fatty Acid Elongation”, indicating a pronounced activation of lipid metabolic routes. In contrast, 17 downregulated genes were enriched in the “Phenylalanine, Tyrosine and Tryptophan Biosynthesis” pathway and in “Carotenoid Biosynthesis”.

GO enrichment analysis showed that, among the upregulated DEGs, 71 terms were assigned to biological processes, 266 to molecular functions and 434 to cellular components (Fig. 6b,c). For the 569 downregulated DEGs, 32 GO terms fell under biological processes, while 223 and 314 were grouped into molecular functions and cellular components, respectively. In both upregulated and downregulated gene sets, the term “metabolic process” was the most strongly enriched. Overall, genes involved in lipid metabolism tended to show higher transcript levels in the mature seed (MS) transcriptome than in the immature seed (IMS) transcriptome, consistent with an upregulation trend at the MS stage.

The expression patterns of selected unigenes during two stages of A. truncatum seed maturation were determined by RT–qPCR for nine candidate genes. Nine genes were selected from differentially expressed genes in enriched lipid-related pathways. Expression values for these genes are provided in Table S2. The expression patterns of the nine genes were consistent with the RNA-seq (FPKM) profiles (Fig. 7). Correlation analysis showed agreement between RT–qPCR and RNA-seq measurements (r > 0.815), supporting the consistency of the transcriptomic results.

AcDGAT (diacylglycerol acyltransferase) and TER (trans-2,3-enoyl-CoA reductase) showed lower expression in MS than in IMS, matching the reduced TAG biosynthesis at the MS stage. Gene-specific patterns were observed for AcKCS (β-ketoacyl-CoA synthase), AcHACD (3-hydroxyacyl-CoA dehydratase), and AcKAR (β-ketoacyl-ACP reductase). AcHACD, KAR-2, and AcKCS-2 were higher in MS, whereas AcHACD-2, AcKAR-1, and AcKCS-1 were lower in MS. AcOLEOSIN showed markedly higher expression in MS than in IMS.

By linking transcript abundance with lipid composition through correlation analysis, we selected 12 unigenes. These unigenes encode enzymes or structural proteins implicated in fatty acid elongation and glycerolipid assembly and show pronounced, stage-associated expression shifts in the RNA-seq dataset. Annotation assigned these unigenes to AcKCS, AcKAR, AcHACD, AcTER, AcDGAT, AcLPAT, and AcOLEOSIN. Pearson correlation coefficients were calculated between unigene expression (FPKM) and the relative abundances of 30 quantified lipid metabolites (TAG and DAG molecular species) across two developmental stages. All 12 unigenes showed statistically significant correlations with the 30 lipid metabolites (Fig. 8). Two AcOLEOSIN unigenes (Cluster-16167.8397 and Cluster-16167.12374) were positively correlated with most TAG species, including multiple TAG54:x and TAG52:x species, and negatively correlated with several DAG species (e.g., DAG34:1, DAG34:2, and DAG34:4). The remaining 10 unigenes, annotated as AcKAR, AcKCS, AcHACD, AcTER, AcDGAT, and AcLPAT, showed the reverse association: positive correlations with DAG species and negative correlations with most TAG species.