Single-cell mapping of homocysteine-induced perturbations in avian embryogenesis using LMO and BD rhapsody transcriptomics Authors: Stéphanie Maupetit-Méhouas, Felipe Maurelia, Pouria Hosseinnia, Nicolas Allègre, Yoan Renaud, Jonas Cruzel, Claire Chazaud, Charlène Guillot Published: 2026 ABSTRACT SummarySingle-cell RNA sequencing enables the study of cellular heterogeneity and the effects of environmental perturbations across complex tissues. However, subtle biological responses can be obscured by experimental noise, sample-specific processing, and computational integration. Here, we combined lipid-modified oligonucleotide cell tagging with the BD Rhapsody platform to multiplex early embryonic chick cells exposed to different homocysteine conditions. This species-agnostic approach enabled robust sample barcoding, simultaneous processing of treatment conditions, and recovery of high-quality single-cell transcriptomes from heterogeneous embryonic tail tissues. By comparing non-integrated and integrated analyses, we found that integration preserved broad cell identities but could mask cluster-specific transcriptional signatures in sensitive populations. Using the non-integrated multiplexed dataset, we identified homocysteine-associated transcriptional changes in neuromesodermal progenitor-derived populations and linked them to posterior mesodermal and axis elongation defects. This workflow provides an adaptable strategy for detecting subtle treatment effects in non-traditional developmental models.Keywords: BD Rhapsody, lipid modified oligonucleotides, chick embryonic cells, multiplexing, single-cell RNA sequencing, environmental perturbations Introduction Single-cell RNA sequencing (scRNA-seq) has emerged as a powerful tool for exploring biological heterogeneity in human and model organisms at the single-cell level.1,2,3 Recent advances in scRNA-seq have enabled the incorporation of cells from multiple samples into single libraries through multiplexing strategies, thereby improving sample throughput while minimizing technical batch effects, library preparation time, and overall costs.4,5 Such strategies allow for direct comparison of different samples under genetic or environmental perturbations while minimizing bioinformatic processing.6,7 Sample multiplexing is generally achieved by exploiting either preexisting genetic diversity or by introducing sample-specific barcodes before pooling, a process commonly referred to as “hashing”. Several approaches have been developed for this purpose, including oligo-labeled antibodies, chemical oligo-tagging, and lipid-based oligonucleotide hashing, implemented either via single-component cholesterol-modified oligos (CMOs) or via lipid-modified oligonucleotides (LMOs) assembled as a two-component anchor/co-anchor system (MULTI-seq LMO).8,9,10 Nevertheless, most of these strategies have been optimized and validated primarily in human and mouse systems, limiting their direct applicability to non-traditional or embryonic model organisms. One of the disadvantages of CMO and LMO hashing is that cell labeling introduces additional processing steps prior to single-cell encapsulation, which can directly affect cell viability and, consequently, the quality of the analysis.11 To mitigate this effect, the MULTI-seq LMO design increases overall hydrophobicity and improves membrane retention of the labeling complex compared to CMOs, supporting more stable barcode association during washing and handling.11 This issue is particularly relevant for fragile samples with low cell numbers or high complexity, such as in vivo embryonic animal model tissues. To date, no study has validated the efficiency of LMO-based labeling in the chick model, particularly in the context of fragile, highly heterogeneous early embryonic cells. Here, we generated a fluorescent LMO probe to assess the ability of this labeling strategy to target dissociated embryonic chicken cells, and to validate both cell viability after treatment and the homogeneity of labeling across different cell and tissue types. In this study, sample multiplexing was subsequently performed exclusively using MULTI-seq LMO lignoceric amide-modified Anchor and palmitic amide-modified Co-Anchor. The BD Rhapsody single-cell RNA sequencing platform offers a notable advantage for processing samples with high cell heterogeneity and RNA content, enabling the generation of robust, high-quality data from in vivo animal tissues.12 Compared with other platforms such as 10× Genomics Chromium, BD Rhapsody demonstrated superior performance with small-sized cells, making it particularly suitable for developmental context where cell dimensions are reduced.13 Currently, the most widely used multiplexing strategy for this platform relies on antibody-based cell tagging with species-specific pan-cell antibodies, allowing the analysis of up to 40,000 cells per experiment.12 Nevertheless, this approach faces important limitations: antibody-based tagging depends on epitope recognition, which may fail to capture the full cellular diversity within the samples.14 Such constrains increase the risk of missed cell populations, especially considering that available reagents are mainly restricted to human and mouse samples. To date, no multiplexing strategy is available for chick samples using the BD Rhapsody system, highlighting an important methodological gap for single-cell studies in this model. By leveraging LMO-based multiplexing, which can target all animal species and cell types, we can overcome the limitations associated with antibody-dependent strategies and perform multiplexing of embryonic samples under different environmental conditions. In addition to enabling the analysis of up to 40,000 cells per experiment, the BD Rhapsody microwell-based scRNA-seq method produces high-quality data from cell suspensions with a minimal starting point of 50% viable cells, partly due to a post-loading washing step that removes dead cells and environmental RNA from the media.12 Here, we first validated LMO labeling in heterogeneous chick embryonic tissues and then applied this strategy to multiplex these samples using the BD Rhapsody platform. Because LMO labeling introduces additional sample-handling steps prior to capture, this workflow may also introduce technical variability across treatment conditions, raising the question of whether data integration is required to appropriately compare multiplexed samples generated with this approach. Although a similar LMO-based strategy has recently been used for single-cell capture of tumor cells in mice,13 its application to early chick embryonic cells under different environmental conditions has not been reported. Moreover, as direct comparisons between integrated and non-integrated analyses remain limited in this experimental context, it is still unclear to what extent integration is usable when coupling LMO multiplexing with BD Rhapsody in embryonic chick samples. The chick model is an exquisite model for developmental and biomedical research.15,16 Used in clinics for teratoma potency of small molecules, it is considered a suitable model for the analysis of cancer cells to diverse treatments/chemicals.17,18 Thanks to its easy accessibility, it has been a great model to study the role of environmental perturbations during development19,20,21,22 and developmental programs such as gastrulation and body axis formation.23,24 Maternal hyperhomocysteinemia is increasingly recognized as an environmental and metabolic perturbation associated with adverse pregnancy outcomes and altered fetal development in humans, including neural tube defects, placental dysfunction, and fetal growth restriction.25,26 However, modeling this type of developmental perturbation in vivo can be challenging in mammalian systems, where maternal physiology, placental transfer, and embryonic exposure are tightly intertwined.27,28 In this work, we used the chick embryo as an experimentally accessible model that allows direct and sustained elevation of homocysteine levels without maternal intervention, thereby providing a useful framework to assess its effects on posterior body axis elongation. Indeed, previous work showed that adding homocysteine to the avian embryo induces neural tube-like defects, indicating an impairment of the neural tube morphogenesis process.29,30 However, how it impacts the associated developmental program and cell states during embryogenesis is not fully understood. Here, we decided to combine targeted homocysteine treatment with multiplexed single-cell RNA sequencing to identify the impact of homocysteine treatment on the transcriptomic cell state of the cells involved in posterior body axis formation. In our study, we combine the benefits of pan-cell tagging from LMOs and the analysis of only viable cells to achieve high-quality single-cell RNA sequencing of vertebrates’ embryonic samples under varying treatment conditions simultaneously, with a capacity of up to 40,000 living cells per experiment. Results Non-toxic doses of homocysteine treatment induce posterior axis defects in Gallus gallus embryos In this study, we selected homocysteine (HCys) doses to approximate physiologically relevant and elevated exposure ranges described in humans. In human plasma, total homocysteine is generally considered within the normal range below approximately 15 μM, whereas higher concentrations are associated with hyperhomocysteinemia and disease states.31 Accordingly, we performed subembryonic injections into the yolk of 20 μL of HCys 0.5 mM and 19 mM to achieve final treatment doses of 0.5 μM (low dose, LD) and 19 μM (high dose, HD) HCys (Figure 1A). Notably, the HD used here remains far away from teratogenic doses previously reported in avian embryos, where treatment with 200 mM D,L-homocysteine increased serum homocysteine to >150 μM.30 After HCys treatment, we saw a visual decrease in the caudal region of the treated embryos, compared to control ones. Occasional defects in the heart and anterior neural tube were also observed (data not shown). These results are consistent with previous knowledge in human embryos showing a correlation between NTD phenotypes and high HCys levels.32 In chick and mouse embryos, it has been shown that treatment with HCys can induce neural tube defect (NTD)-like phenotypes.33,34 Labeling of embryonic chick cells with lipid-modified oligonucleotides (LMO) (A) Schematic representation of the experimental procedure shows chicken embryos treated with homocysteine at different concentrations- and incubated for 72 h (left). Schematic representation of dissociated cells barcoded with LMO (anchor and co-anchor), labeled with a fluorescent probe (Atto550 dye) for fluorescence detection (right). (B) Histogram of single-cell population labeled with LMOs at 20 nM and 200 nM, analyzed on BD FACS Melody. (C) Density plot shows the distribution of LMO-labeled cells within live and dead populations, based on FVS 660 and Atto 550 intensities in single cells measured on a BD FACS Melody. Approximately 15% are dead (upper right), 79% are alive and LMO-labeled (lower right), and 4.8% are alive but non-labeled (lower left). (D) Representative confocal image of a dissociated live cell (Hoechst positive) labeled with LMOs (Atto550 positive), acquired at 20× magnification on a Zeiss LSM 980. Scale bars, 20 μm. Lipid-modified oligonucleotides uniformly label viable chick embryonic cells Lipid Modified Oligonucleotides (LMOs) offer a versatile sample barcoding solution compatible with various cell types or nuclei.4,9 Their hydrophobic extremity can insert into accessible plasma membranes, facilitating pooled single-cell sample processing from varying treatment conditions. Comprising a Lignoceric amide-modified anchor DNA oligo and a palmitic amide-modified co-anchor DNA oligo, these lipid-modified oligos seamlessly embed into cell or nucleus membranes, providing a landing pad for DNA barcodes with complementary 5′ sequences in their “handle” region.9 To assess LMO insertion into membranes of chick cells, we designed a fluorescent probe capable of hybridizing to the "handle" region at the 3′ terminal end of the anchor, specifically labeled with Atto550 (Figure 1A, right). Using the viability stain (FVS 660 BD Biosciences), we first distinguished live and dead cells by flow cytometry. We then included a negative control for LMO labeling (probe only) to define the specificity of the fluorescent signal and establish the gate for LMO-positive cells based on the background/non-specific Atto550 signal (Figure S1A). This strategy allowed us to distinguish LMO-labeled from unlabeled cells while excluding dead cells and background fluorescence. Insertion efficiency was then analyzed by flow cytometry using a BD FACS Melody. We tested three final LMO concentrations (2 nM, 20 nM, and 200 nM) to label dissociated cells (Table S1). At the manufacturer-recommended LMO concentration of 200 nM, we observed more uniform labeling, as shown by the histogram distribution (Figure 1B). Lower concentrations revealed a biphasic behavior, with highly labeled and weakly labeled cells (Figure 1B). Using the manufacturer-recommended concentration of 200 nM, we achieved LMO labeling in 79% of live single cells, with 15% mortality (Figures 1C and 1D). Overall, LMO is an effective method to label heterogeneous chick embryonic cells. LMO-based multiplexing enables sample barcoding in the BD rhapsody workflow For labeling each cell from a specific sample, a unique Sample Barcode (SBC) was synthesized with unique 8-nucleotide barcodes. These sequences feature a 5′ region complementary to the “handle” of the anchor, followed by validated barcode sequences9 and a 3′ polyA tail to attach to the Rhapsody bead (Table S2). To perform LMO labeling of chick embryonic cells, cells are dissociated and labeled per sample by a specific mixture of anchor-co-anchor and SBC (See STAR Methods). After incubation on ice, excess LMOs were removed by centrifugation steps (2 times). After assessing final cell viability after LMO labeling, 13,000 viable cells of each sample were loaded onto the cartridge. The steps of cell capture were performed following those outlined in the BD Rhapsody protocol and the library preparation steps (Figure 2). Steps of mRNA and LMO capture from single-cell-derived embryonic chick tissue using the Rhapsody Cartridge (A) Diagram illustrates the main steps of cell capture in the BD Rhapsody system. Unlike standard capture methods, the sample barcode construct (SBC) contains a polyA region, enabling the identifier to hybridize to the beads after cell lysis. This feature ensures its retention during purification and amplification during the reverse transcription step, thereby allowing efficient capture and downstream analysis. (B) LMO library amplification by PCR on the supernatant after lysis, enabling the amplification and sequencing of barcode regions for cell identity assignment. Custom primers are listed in Table S3. (C) mRNA library amplification by PCR on the purified beads, required for whole-transcriptome library preparation as indicated in the BD Rhapsody protocol. A modified BD AbSeq protocol recovers LMO-derived sample barcodes mRNAs and LMOs from individual cells were captured on Rhapsody beads through polyA/polyT-based interactions (Figure 2A). The cellular barcodes carried by the beads enable assignment of mRNA transcripts to individual cells, whereas LMO-derived sample barcodes allow each cell to be associated with its biological sample of origin. After capturing mRNAs and LMOs on beads, reverse transcription and exonuclease treatments are performed following the established BD protocol. The supernatant from the beads is preserved and processed separately to generate a library specific to the LMOs, allowing for the barcoding of cells (Figure 2B and STAR Methods). Here, we used a specific primer named Multi Seq Primer that is complementary to the "Handle" region to perform the PCR1 step, which also contains the complementary region to allow the binding of Primer PCR2 (Table S3). Simultaneously, a transcriptome library is generated using beads loaded with cDNA (Figure 2C). This process involves amplification by random priming and extension, followed by two PCR steps. To sequence the two libraries, we performed paired-end 150 bp sequencing on a high-throughput Illumina platform. Depth of sequencing for the LMO library was 2,000 reads/cells and 50,000 reads/cells for the Whole Transcriptome Analysis library. BD rhapsody analysis successfully retrieves both transcriptome and LMO barcode libraries BD Rhapsody provides an online platform, i.e., “Seven Bridges”, with an in-house analysis pipeline, allowing users to check the sequencing quality, align and annotate the reads, and generate the necessary output files required for downstream analysis. We adapted the VDJ function alignment to retrieve the LMO sequences within the BD Rhapsody WTA Analysis Pipeline. To do so, we provided the FASTQ files from the WTA and LMO sequencing, the reference genome, the transcriptome annotation, and the Barcode sequences as a supplemental reference file. Output files include expression matrices filtered or unfiltered, BAM files, and a pipeline report indicating the quality of the sequencing and alignment (Data S1). Quality controls of the alignment indicate successful amplification of both mRNA and LMOs. Specifically, 96.41% of reads passed the quality filters, with 75.6% of cellular reads aligning uniquely for mRNA, while 97.16% of reads passed quality filters, and 79.87% of cellular reads aligned uniquely for LMOs (Data S1). The t-SNE plots (Figure S1B, upper) depict the sum of gene sets from the WTA library for all three samples combined, with 28,146 total genes detected. Good expression levels of each sample’s barcodes were also retrieved for each sample, with a broadly unimodal, approximately symmetric distribution, with most cells showing around 100 barcode counts (log-transformed value ≈2; Figure S1B, lower). This pattern supports a relatively homogeneous pooling of samples and indicates correct retrieval of LMO-derived barcodes together with the expected recovery of mRNA transcripts. LMO demultiplexing identifies high-quality single-barcoded chick embryonic cells To analyze the data, we utilized the published Seurat pipelines to demultiplex the data, assign the correct LMO identity, and perform quality control.13 Applying stringent filters for LMO numbers per cell (>100 barcodes per cell), we retrieved 16,385 single cells barcoded for analysis (Figure S1E). The effect of this filtering step is shown in Figure S1C, which presents violin plots of nFeature_RNA, nCount_RNA, and percent.mt before and after QC filtering. To further remove potential artifactual profiles, doublets were identified computationally, and their distribution in the dataset is shown in Figure S1D, where doublets are highlighted in red and singlets in gray on the UMAP projection. High-quality cells were defined as those with less than 25% mitochondrial gene content and expression of 1,000 to 10,000 genes. After filtering, we retained 9,231 high-quality single-barcoded cells across the three samples for downstream analysis (Figure S1E). Integration strategies preserve broad cell identities but alter sensitive tail-forming populations Technical noise can arise during single-cell RNA sequencing experiments at multiple stages, including cell dissociation, handling, and library preparation, where differences in cell stress, RNA capture efficiency, or reverse transcription can introduce variability unrelated to biology. Additional processing steps such as washing, labeling, or multiplexing may further affect cell recovery and transcript detection, leading to batch-specific biases. These sources of variation can influence clustering and gene expression patterns, making careful normalization and integration important when comparing samples. While here we used simultaneous encapsulation and sequencing steps, technical noise could still remain and affect the interpretation of the biological results. To assess if/how computational integration affects the identification of tail formation cell populations, we first established a non-integrated baseline dataset (Figure 3A left) and assigned cluster identities based on positive differentially expressed genes (DEGs) with high adjusted significance, validated using the GEISHA database (Figure 3A right; Figure S2) and our previous knowledge of existing cell states in this dissected area.23 This baseline served as a reference to compare the effects of different integration strategies on cluster structure and biological identity. Baseline clustering and evaluation of integration strategies for tail formation cell populations (A) UMAP of the non-integrated baseline dataset, showing (left) distribution of cells per sample (SBC1: Control; SBC3: Low dose HCys; SBC7: High dose HCys) and clusters assigned to tail formation cell populations (right). Cell identities were annotated using validated marker genes from the GEISHA database. (B) UMAPs of RPCA and Harmony-integrated datasets, including the default setting (rpca_main) and two additional parameterizations (s1 and s2). Cluster colors were matched to the corresponding baseline populations. For each integrated cluster, similarity to the baseline was assessed as the percentage of shared positive differentially expressed genes (DEGs), calculated using the top 200 positive DEGs ranked by adjusted p value. (C) Number of positive DEGs identified per cluster in the baseline dataset and in the three RPCA integration conditions (rpca_main, rpca_s1, and rpca_s2). (D) Reference genes for the late neural tube population corresponding to baseline cluster 5, together with GEISHA in situ hybridization images for NOTCH2 and ROBO1, supporting the annotation of this cluster as late neural tube. We then generated integrated datasets using two commonly used approaches, RPCA35 and Harmony,36 each tested under a default configuration and two additional parameterizations (Figure 3B). For RPCA, we varied the number of integration features (nfeatures) and the stringency of anchor filtering (k.filter), while keeping anchor detection constant (k.anchor = 5). Specifically, nfeatures controls how many variable genes are used to define the integration space, whereas k.filter determines how stringently low-confidence anchors are removed during integration; higher values make anchor filtering more permissive, while lower values apply a stricter filter. We also adjusted k.weight, which controls how many anchors contribute to the weighting step when generating the integrated embeddings. Using this framework, we tested three RPCA settings: rpca_main (nfeatures = 2000, k.anchor = 5, k.filter = 50, k.weight = 25), rpca_s1 (nfeatures = 1800, k.anchor = 5, k.filter = 40, k.weight = 20), and rpca_s2 (nfeatures = 2500, k.anchor = 5, k.filter = 100, k.weight = 50). For Harmony, we varied the parameters controlling the strength and structure of batch correction. In particular, theta is the diversity penalty, which determines how strongly Harmony encourages mixing of cells from different samples within clusters; larger values promote stronger batch mixing. lambda is a ridge penalty that controls the aggressiveness of the correction, with smaller values leading to stronger correction. Sigma defines the width of the soft k-means clustering, and nclust sets the number of clusters used internally by the Harmony model. In our analysis, we kept lambda, sigma, and nclust constant and varied theta to test different correction strengths: harmony_main (theta = 2, lambda = 1, sigma = 0.1, nclust = 50), harmony_s1 (theta = 1, lambda = 1, sigma = 0.1, nclust = 50), and harmony_s2 (theta = 3, lambda = 1, sigma = 0.1, nclust = 50). Cluster colors in the integrated UMAPs were matched to the corresponding baseline populations, and for each integrated cluster, we quantified similarity to the baseline as the percentage of shared positive DEGs, using the top 200 positive DEGs ranked by adjusted p value. Overall, both integration methods preserved the general structure of the dataset, but important differences emerged for specific populations. In particular, Harmony did not resolve the neuromesodermal progenitor (NMP) population at this clustering resolution, whereas RPCA detected NMPs in the main and s1 settings only. This indicates that the ability to preserve biologically relevant populations depends on both the integration method and parameter choice. A more detailed comparison based on shared DEGs showed high overlap with the baseline for most clusters across the RPCA-integrated datasets, supporting broad conservation of cluster identity (Figure 3B). However, one notable exception was the late neural tube population, corresponding to baseline cluster 5 (dark red), which showed 0 shared DEGs with its corresponding integrated cluster in rpca_main and rpca_s1 (pale red). Consistent with this observation, the number of positive DEGs identified per cluster revealed that cluster 2 in the integrated RPCA datasets contained fewer than 10 positive DEGs, whereas most other clusters retained numbers of DEGs comparable to the non-integrated baseline (Figure 3C). Thus, integration had a disproportionately strong effect on this specific population. Importantly, the apparent loss of DEGs in the integrated late neural tube cluster did not reflect a complete loss of biological identity. When we examined the expression of late neural tube reference genes validated in GEISHA, including NOTCH2 and ROBO1, the corresponding integrated cluster still retained the expected biological signature of this population (Figure 3D). This suggests that integration, particularly in the RPCA settings tested here, can overcorrect certain clusters, making them transcriptionally more similar to neighboring populations and thereby masking cluster-specific DEGs without fully erasing their underlying identity. Together, these results show that integration affects some cell populations more strongly than others. In our dataset, the non-integrated BD Rhapsody baseline was essential to preserve the transcriptomic distinctiveness of sensitive populations such as the late neural tube cluster. This is particularly relevant for downstream comparisons across treatments, because overcorrection of specific clusters could obscure biologically meaningful changes in gene expression that would otherwise remain detectable in the non-integrated analysis. Because multiplex datasets contain limited technical variation, any correction introduced during integration more directly reflects parameter-driven effects, making them an ideal framework for selecting settings that minimize transcriptional distortion prior to integrating additional datasets. Homocysteine treatment induces subcluster-specific transcriptional changes in NMP-derived populations To further evaluate whether homocysteine treatment altered the representation of major tail cell populations, we first compared the relative contribution of cells per cluster across conditions using the non-integrated, RPCA- and Harmony-integrated datasets (Figure S3). Overall, several populations showed similar distributions across control and treated samples, indicating that treatment did not globally reshape all embryonic cell identities. For example, the lateral plate mesoderm showed no evident change in proportional representation across conditions, whereas blood cells were consistently increased in both treatment groups. In addition, populations associated with neural tissues, surprisingly, did not show major shifts in their relative proportions after treatment, while differences in the NMPs and the paraxial mesoderm cluster were visible. These results suggest that, at the level of broad cluster annotation, homocysteine treatment does not uniformly alter cell composition across all tail populations. However, because cluster assignment at this level captures the global transcriptional profile of each population, subtle treatment-associated shifts within developmentally dynamic progenitor compartments may remain unresolved. We therefore focused on the clusters associated with tail tissues affected in NTD-like phenotypes, which are the neuromesodermal progenitor (NMP), the paraxial mesoderm, and early neural tube clusters. Since NMPs are bipotent cells that can differentiate into mesoderm or neural tissue, changes in their proportion and transcriptional signature can impact the proper development of these tissues. To determine whether treatment altered transcriptional states within this developmental continuum, we reclustered these three populations (Figure 4A, upper). Subclustering analysis and morphological effects of homocysteine treatment during tail formation (A) Subclustering of the NMP, paraxial mesoderm, and early neural tube populations identified five clusters after segmentation of cells according to the UMAPs of SBC1 (control), SBC3 (low-dose HCys), and SBC7 (high-dose HCys) (upper). Expression of the marker genes TBXT and SOX2 for NMPs, MEOX1 for paraxial mesoderm, and PAX6 for early neural tube allowed the putative identity of these subclusters to be assigned. (B) Upper: normalized number of treatment-induced differentially expressed genes (DEGs) per subcluster compared to the control treatment. Lower: normalized number of treatment-induced DEGs per subcluster normalized by cluster size. Low dose vs. baseline (gold); High dose vs. baseline (red). (C) Representative images of embryos under control, low-dose HCys, and high-dose HCys treatments, shown in dorsal and ventral views. Dashed boxes indicate the dissected region. Scale bars, 200 mm. (D) Global measurements of embryo size by condition, including total embryo size, anterior size (from the blood vessel branching to the nose), and posterior size (from the blood vessel branching to the end of the tail). n = 27. Statistical significance was assessed using a two-tailed t test. Posterior size: control vs LD p = 0.0428; control vs HD p = 0.0160. (E) Representative images of the tail end under each condition, with immunostaining against TBXT (green) and SOX2 (purple). SOX2 marks early neural tissue (NT), double-positive cells correspond to NMPs, and TBXT marks mesodermal tissue. Using the DAPI channel, the last formed somite (LSOM) was delineated as a landmark of mesodermal fate acquisition. Scale bars, 200 μm. (F) Quantification shows the effect of treatment on mean LSOM width and mean LSOM length. n = 27. Statistical significance was assessed using a two-tailed t test. Width: control vs LD p = 0.0059; control vs HD p = 0.0157. Length: control vs LD p = 0.0239. (G) Quantification of the distance between the LSOM and the tail end, including mean PSM width, mean width of the posterior NT, mean width of the anterior NT, and the distance from the NT to the tail end. n = 27. Statistical significance was assessed using a two-tailed t test. LSOM-END: control vs LD p < 0.0001; control vs HD p = 0.0020. This subclustering analysis identified five subclusters, whose putative identities were assigned using canonical markers of NMP and lineage commitment.37,38 Expression of TBXT and SOX2 was used to identify NMP-like states, MEOX1 to identify late paraxial mesoderm, and PAX6 to identify early neural cells (Figure 4A). Based on these markers, cluster 2 corresponded to a population of pure NMPs, whereas part of cluster 1 represented neural cells, cluster 4 paraxial mesoderm (MSGN1 positive), and cluster 0 somite identities. In contrast, cluster 3, which was overrepresented in treated embryos, had markers of neural, NMPs and paraxial mesodermal identities, indicating specific transcriptional cell states of these cellular identities arising in treated embryos. We next examined treatment-induced transcriptional responses within each subcluster by quantifying the number of significant differentially expressed genes (DEGs) in low-dose vs. control and high-dose vs. control comparisons (Figure 4B, upper). This analysis revealed that low-dose treatment reduced the number of DEGs in cluster 1, composed of neural cells and NMP-like cells, whereas high-dose treatment was associated with an increased number of DEGs in the more mesodermal subcluster. Because larger clusters may inherently yield higher DEG numbers, we then normalized DEG counts by cluster size (Figure 4B, lower). After normalization, both treatments showed broadly similar effects across most subclusters. Importantly, cluster 3 (HCys-NMPs), Cluster 1 (neural), and cluster 4 (paraxial mesoderm) show strong DEG differences in the treated versus control cases, indicating that even though these cells maintain their marker genes and cell identity, HCys treatment can induce some transcriptional changes. Together, these results indicate that homocysteine treatment affects the transcriptional output of NMP-related populations in a subcluster-specific manner. Homocysteine preferentially disrupts posterior mesodermal output and axis elongation To validate whether these transcriptional changes were associated with morphological defects, we then examined the impact of treatment on embryo morphology (Figures 4C–4G). Representative dorsal and ventral views of control, low-dose HCys, and high-dose HCys embryos are shown in Figure 4C, with dashed boxes indicating the end-tail region. Global morphometric measurements confirmed that the treatment effect was region-specific: although overall embryo size and anterior size were not significantly altered, posterior size was significantly reduced under both low- and high-dose treatment (Figure 4D). These data support the idea that homocysteine preferentially affects posterior axis development. To investigate this phenotype at higher resolution, we analyzed the caudal region by immunostaining against TBXT and SOX2 (Figure 4E). In this context, SOX2 marks early neural tissue, TBXT labels mesodermal tissue, and double-positive cells correspond to NMPs. Using the DAPI channel, we delineated the last formed somite (LSOM) as a landmark of mesodermal segmentation. Quantification of these images showed that LSOM width was significantly reduced under both treatment conditions, whereas LSOM length was significantly reduced only under low-dose HCys (Figure 4F). These observations show significant variation in somite shapes in HCys-treated embryos, which is consistent with the single-cell RNA sequencing showing mesodermal lineage variation. This indicates that homocysteine may affect the transition of progenitors toward mesodermal derivatives, since the LSOM represents one of the intermediate outputs of this lineage. Consistent with this interpretation, measurements of the caudal domain further showed a significant reduction in the distance between the LSOM and the tail end in treated embryos (Figure 4G), suggesting a reduction in the pool of cells contributing to posterior body axis elongation. By contrast, no significant changes were observed in the mean width of the PSM, of the posterior forming neural tube, of the anterior neural tube facing the last somite formed, or the distance from the neural tube to the tail end. This relative preservation of neural dimensions is consistent with both the proportional analysis shown in Figure S3 and the lack of major treatment-associated changes in broad neural populations, supporting the conclusion that homocysteine primarily affects posterior mesodermal output and axis elongation, rather than causing a generalized disruption of the neural tissue. Overall, this study establishes, to our knowledge, the first application of BD Rhapsody coupled to LMO-based multiplexing in embryonic chick cells for single-cell RNA sequencing. We show that this strategy enables robust sample barcoding, efficient recovery of high-quality singlet cells, and reliable identification of embryonic tail-forming populations in a highly heterogeneous and dissociation-sensitive tissue. By systematically comparing non-integrated and integrated analyses, we further demonstrate that computational integration can disproportionately affect specific cell populations, in some cases masking cluster-specific transcriptional signatures while preserving broader cell identity. Using this framework, we uncovered that homocysteine treatment preferentially perturbs posterior mesodermal output and body axis elongation, while producing more limited effects on broad neural population proportions. Together, these findings not only provide a practical multiplexing workflow for the chick model, but also highlight the value of non-integrated multiplexed single-cell analysis for detecting biologically meaningful transcriptional changes that may otherwise be obscured in standard integrated pipelines. Discussion Single-cell RNA sequencing (scRNA-seq) has revolutionized the study of cellular heterogeneity, providing unparalleled insights into complex biological systems.2 However, different multiplexing strategies have a profound impact on data quality, scalability, and applicability across species. To date, most validated protocols have relied on antibody-based tagging, such as TotalSeq HashTag antibodies, which are restricted to human and mouse systems.4,8,10 In contrast, our study demonstrates that Lipid-Modified Oligonucleotides (LMOs) constitute a versatile and species-agnostic labeling method that can be easily combined with the BD Rhapsody microwell-based platform. This combination enables robust multiplexing of chick embryonic cells during early development, thereby addressing critical challenges in assessing the impact of environmental perturbations on dissociation-sensitive and heterogeneous embryonic tissues. Compared to antibody-based approaches, LMO labeling offers several advantages. First, its independence from epitope recognition avoids potential biases coming from limited antibody availability or uneven cell surface marker exposure, thereby offering coverage for the highly heterogeneous early embryonic tissue. Second, the combination of LMO labeling with BD Rhapsody’s washing steps ensures that only viable cells can contribute to downstream analysis, effectively minimizing background signal from ambient RNA. At the recommended concentration of 200 nM of LMO label, we achieved uniform labeling of ∼80% of viable cells, while maintaining scalability to up to 40,000 cells per cartridge. By retaining only cells with a high barcode count (>100), we obtained 16,000 cells for downstream analysis. Our sequencing depths—2,000 reads per cell for LMO libraries and 50,000 reads per cell for transcriptome libraries—meet or exceed industry benchmarks for multiplexed workflows.5 These results are consistent with their prior use on PBMCs and cell lines, demonstrating that LMO labeling achieves a high signal-to-noise ratio when optimized.9,11 Importantly, all the cell states retrieved in the LMO multiplexed combined with BD Rhapsody were previously found when using no multiplexing combined with indrop single-cell sequencing, albeit with deeper sequencing depth, and more genes were recovered with this novel technology.23,39 LMOs have also been successfully applied in other vertebrate models, including zebrafish brain development,11 supporting their broad applicability across species. In the avian model, MULTI-seq has been recently applied to study intestinal development at later embryonic stages using a 10× Genomics platform.40 Our work extends these findings by providing a demonstration of LMO insertion in chick embryonic cells, as well as an adaptor to directly test the efficient insertion of LMO in living cells. This not only establishes a validated workflow for multiplexed scRNA-seq in early chick embryos—a critical window for body axis formation with highly heterogeneous populations that had not been addressed previously—but also provides, for the first time, a reproducible pipeline to perform LMO barcoding of embryonic cells with the BD Rhapsody platform. This novel combination provides a robust and scalable approach for dissociation-sensitive embryonic tissues, expanding the applicability of both LMOs and BD Rhapsody beyond human and mouse systems. A central challenge of scRNA-seq is distinguishing true biological variation from technical noise. When comparing different data types, batch correction algorithms are often used to mitigate technical artifacts.41,42 However, these algorithms can sometimes overcorrect, removing genuine biological information. Overcorrection can make downstream analyses problematic and often result in incorrect conclusions. In our study, LMO barcoding allowed us to compartmentalize cells from the outset, thereby reducing technical variability and eliminating the need for computational batch correction. This approach ensures precise sample identification and comprehensive transcriptomic profiling. Indeed, we observe in our analysis that all cells occupy similar spaces in the UMAP plot without any batch effect corrections. More importantly, when using batch effect correction algorithms, some cell types were merged, and subtle variations after treatment were hidden. This suggests that our multiplexing strategy provides a significant advantage, preserving subtle transcriptomic changes, such as those induced by environmental treatments that might otherwise be corrected by computational integration, thereby enabling more accurate biological interpretations. From a biological perspective, our results provide a new model system to study the role of elevated homocysteine in perturbing early embryonic development in chicken embryos. Previous studies have associated elevated homocysteine with neural tube defects,43,44,45 impaired neural crest delamination and migration,46,47 and altered vascular function.48,49,50,51 In line with this, we identified transcriptional variance in several cell populations linked to these processes, underscoring the broader impact of homocysteine treatment on early embryonic lineages. Importantly, both integrated and non-integrated analyses consistently revealed a population of cells associated with blood progenitors -one of the best-documented lineages affected by homocysteine treatment-further supporting the biological relevance of our system. Focusing on NMPs, paraxial mesoderm, and neural tube cells, which represent the major tissues involved in neural tube morphogenesis, we detected minimal transcriptional and proportional changes in neural identities, while mesodermal and NMP populations were more substantially affected. This novel finding suggests that caudal NTD-like phenotypes may emerge from mesodermal lineage impairment, rather than from primary disruption of neural identities. This interpretation is supported by the critical contribution of mesoderm to neural tube morphogenesis and posterior axis elongation.24,52,53,54 Furthermore, these results are consistent with the multitailed phenotype observed in HCys-treated embryos and with mesodermal outgrowth described in both diabetic and non-diabetic mouse models of neural tube defects.33,55 Nevertheless, the precise molecular mechanism by which homocysteine exerts these effects remains speculative, and further work is needed to dissect the direct effects of this molecule and its indirect consequences in disrupting one-carbon metabolism during early development. Together, our findings demonstrate that integrating LMO tagging with the BD Rhapsody platform provides a powerful approach for high-quality scRNA-seq of complex samples including early heterogeneous embryonic tissues. This methodology offers significant potential for expanding single-cell studies to include non-mouse or human organisms and broadens the scope of multiplexing workflows to challenging dissociation-sensitive systems and environmentally challenged conditions where changes may be more subtle than genetic perturbations. Limitations of the study One limitation of our study lies in the reproducibility of the injection procedure, as each egg presents intrinsic differences that may introduce variability in the results after treatment. To mitigate this, we pooled multiple samples to minimize the impact of inter-embryo variation. Another limitation concerns the number of cells analyzed: although the dataset is robust, it may be biased by the capacity of the BD Rhapsody cartridge to capture different cell populations equally. Nonetheless, the identification of expected cell populations is consistent with our previous dataset generated without multiplexing and with a drop-based single-cell platform,23 supporting the accuracy of the current data. Finally, regarding biological interpretation, this work represents an initial step. While our analyses reveal transcriptomic differences between treated and untreated embryos, further validation—through hybridization chain reaction (HCR) or by functional assays—will be necessary to confirm the observed changes and to better understand the mechanisms underlying homocysteine’s effects during early embryonic development. Resource availability Lead contact Requests for further information and resources should be directed to and will be fulfilled by the lead contact, Charlene Guillot (charlene.guillot@uca.fr). Materials availability This study did not generate new unique reagents. Custom oligonucleotides used for LMO-based sample barcoding and library preparation are listed in the supplemental information. Any additional information regarding materials used in this study is available from the lead contact upon request. Data and code availability •Single-cell RNA-seq data generated and analyzed in this study have been deposited in the NCBI Gene Expression Omnibus under accession number GEO: GSE303866 and are publicly available as of the date of publication. Single-cell RNA-seq data generated and analyzed in this study have been deposited in the NCBI Gene Expression Omnibus under accession number GEO: GSE303866 and are publicly available as of the date of publication. •All original code has been deposited at Zenodo at https://doi.org/10.5281/zenodo.20593689 and is publicly available as of the date of publication. The code is also available at GitHub: https://github.com/GuillotLabUCA/Maupetit-Maurelia-et-al-2026. All original code has been deposited at Zenodo at https://doi.org/10.5281/zenodo.20593689 and is publicly available as of the date of publication. The code is also available at GitHub: https://github.com/GuillotLabUCA/Maupetit-Maurelia-et-al-2026. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Acknowledgments The authors thank Dr. Krzysztof Jagla from the Institute of Genetics and Reproduction for their helpful discussions, suggestions, and comments. We are also grateful to the BD team and the Seven Bridges support members for their assistance and technical help. The authors would also like to acknowledge the outstanding Romane Bertrand and the members of the Guillot lab, and the technical services provided by the Institute of Genetics, Reproduction and Development platforms: SC3 – Single Cell & Cell Culture Facilities, the Bioinformatics Platform (BIM), and the Clermont Confocal Imaging Platform (CLIC), for providing feedback and help with the organization of the experiments. We also thank the administrative and technical shared support staff of the institute for their constant assistance. This work was supported by the CIR3 i-SIte Young Group Leader Fellowship (I04ABMECA/IR21GUILLOTCHAIR) to C.G., and by the INSERM/UCA Junior Professor Chair and the French National Research Agency (ANR-22-CPJ2-0084-01) to C.G. Author contributions S.M.-M.: writing – review and editing, writing – original draft, visualization, methodology, and investigation. F.M.G.: writing – review and editing, visualization, software, investigation, formal analysis, and data curation. P.H.: writing – review and editing, software, formal analysis, and data curation, N.A.: methodology and investigation. Y.R.: software, formal analysis, and data curation. J.C.: methodology and investigation. C.C.: review and editing, supervision, and resources. C.G.: writing – review and editing, writing – original draft, project administration, supervision, resources, methodology, investigation, formal analysis, data curation, funding acquisition, and conceptualization. Declaration of interests The authors declare no competing interests. Declaration of generative AI and AI-assisted technologies in the writing process During the preparation of this work, the authors used ChatGPT (OpenAI) to assist with English-language editing, refinement of sentence structure, improvement of clarity, and troubleshooting or drafting code for data analysis workflows. The tool was not used to generate scientific conclusions, analyze data independently, or create original experimental results. After using this tool, the authors reviewed, edited, and validated the content and code as needed and take full responsibility for the content of the publication. STAR★Methods Key resources table REAGENT or RESOURCESOURCEIDENTIFIERChemicals, peptides, and recombinant proteinsFixable Viability Stain 660BD BiosciencesCat# 564405; RRID: AB_2869571MULTI-Seq LMO Lig-AnchorGene LinkCat# 26-2202-02MULTI-Seq LMO Palm Co-AnchorGene LinkCat# 26-2204-02AccutaseGibcoCat# A1110501Hank’s Balanced Salt SolutionSigma-AldrichCat# H8264Bovine serum albuminSigma-AldrichCat# 05470L-HomocysteineSigma-AldrichCat# 69453Anti-SOX2, rateBioscienceCat# 14-9811-82; RRID: AB_11219471Anti-TBXT, goatR&D SystemsCat# AF2085Phosphate-buffered salineEurobio ScientificCat# CS1PBS02-01Critical commercial assaysBD Rhapsody Cartridge Reagent KitBD BiosciencesCat# 633731BD Rhapsody Cartridge KitBD BiosciencesCat# 633733BD Rhapsody cDNA KitBD BiosciencesCat# 633773BD Rhapsody WTA Amplification KitBD BiosciencesCat# 633801Agencourt AMPure XP magnetic beadsBeckman Coulter Life SciencesCat# A63880Deposited dataSingle-cell RNA-seq dataThis paperGEO: GSE303866Analysis codeThis paperhttps://doi.org/10.5281/zenodo.20593689Experimental models: Organisms/strainsGallus gallus, JA57 strainEARL Les Bruyères, Dangers, FranceN/AOligonucleotidesLMO sample barcodes (SBCs)This paperTable S2Custom primers and Atto550-labelled LMO detection probeThis paperTable S3Software and algorithmsBD Rhapsody WTA Analysis Pipeline, Revision 8Seven Bridgeshttps://www.sevenbridges.comRR Project for Statistical Computingv4.4.0; https://www.r-project.orgRStudioPosithttps://posit.co/download/rstudio-desktopSeuratSatija Labv5.1.0; https://satijalab.org/seurat/MULTI-seqMcGinnis et al.9https://github.com/chris-mcginnis-ucsf/MULTI-seqscDblFinderBioconductor[v1.24.0; https://bioconductor.org/packages/release/bioc/html/scDblFinder.htmlNo integration scRNAseq pipelineThis paperhttps://github.com/GuillotLabUCA/Maupetit-Maurelia-et-al-2026Other40 μm cell strainerCorningCat# CLS431750BD FACSMelodyTM cell sorterBD BiosciencesFACSMelodyTM cell sorter4200 TapeStation systemAgilent TechnologiesG2991BANovaSeq sequencing platformIllumina/NovogeneNovaSeq Experimental model and study participant details Fertilized Gallus gallus eggs, JA57 strain, were obtained from EARL Les Bruyères, Dangers, France, maintained at 12°C and further incubated under standard conditions. To induce experimental treatments, embryos received subembryonic injections into the yolk with 20 μL of L-homocysteine solution (Hcys, Sigma-Aldrich, Cat#69453) or control solution. Low-dose homocysteine-treated embryos received 20 μL of a 0.5 mM homocysteine solution, high-dose homocysteine-treated embryos received 20 μL of a 19 mM homocysteine solution, and control embryos received 20 μL of Hank’s Balanced Salt Solution (HBSS; Sigma-Aldrich, Cat# H8264). Assuming dilution within an approximate yolk volume of 20 mL, these injection conditions were estimated to result in final exposure doses of approximately 0.5 μM and 19 μM homocysteine for the low-dose and high-dose conditions, respectively. Embryos were collected at 72 h post-incubation, corresponding to Hamburger–Hamilton stages 18–19. Treated and control embryos were processed in parallel under identical conditions to ensure comparability. For each experimental condition, two posterior embryonic regions were dissected from four embryos: the tail bud from two embryos, representing the posterior-most structures, and the posterior end from two embryos, including the last formed somite, presomitic mesoderm, posterior neural tube, and tail bud. Dissected tissues were briefly rinsed in cold HBSS and enzymatically dissociated in Accutase (Gibco, Cat# A1110501) for 15 min at 37°C, followed by gentle mechanical trituration. The resulting cell suspension was filtered through a 40 μm cell strainer, and cell density and viability were evaluated using a hemocytometer. Cells were then centrifuged at 350 × g for 5 min at 4°C and resuspended in 0.25% BSA in PBS before downstream processing. Embryonic sex was not determined because sex cannot be reliably assessed at the developmental stages analyzed in this study; therefore, potential sex-specific effects could not be evaluated. Ethical statement This study did not require ethical approval, as it involved only early-stage Gallus gallus embryos collected before hatching, in compliance with European Directive 2010/63/EU. Method details Validation of LMO insertion into the membrane of dissociated chick cell by flow cytometry Lipid-modified oligonucleotides (LMOs) were used exclusively for sample hashing. Specifically, MULTI-Seq LMO Lig-Anchor (Gene Link; Cat# 26-2202-02) and MULTI-Seq LMO Palm Co-Anchor (Gene Link; Cat# 26-2204-02) reagents were used to label dissociated chick embryonic cells. For LMO labeling, 5 × 105 dissociated cells were incubated on ice with Anchor LMOs for 5 min, followed by the addition of an equal concentration of co-Anchor LMOs for an additional 5 min. Three final Anchor/co-Anchor concentrations were tested in parallel: 200 nM, 20 nM, and 2 nM. To simultaneously monitor cell viability and LMO incorporation, cells were co-incubated with Fixable Viability Stain 660 (FVS660; BD Biosciences, Cat# 564405; 1:200 from stock) and a complementary Atto550-labelled probe targeting the LMO handle sequence at a final concentration of 0.1 μM. After incubation, cells were centrifuged at 500×g for 10 min at 4°C, resuspended in HBSS, and analyzed on a BD FACS Melody cytometer. Controls included unstained cells, FVS660-only cells, probe-only cells, and LMO + probe-labelled cells, which were used to define compensation, background fluorescence, and gating strategy. Dead cells were excluded using FVS660 staining, and LMO-positive cells were identified based on Atto550 fluorescence above the probe-only background. LMO-based sample multiplexing and BD rhapsody single-cell capture Single-cell dissociation protocols were optimized to achieve >90% viability and minimize doublet formation before capture. For LMO-based sample barcoding, 2.5 × 105 cells from each tissue sample were incubated in PBS supplemented with 200 nM Anchor LMO and 200 nM tissue-specific sample barcode oligonucleotide. Each sample barcode was assigned to a specific tissue or treatment condition. After gentle agitation and 5 min incubation on ice, 200 nM co-Anchor LMO was added for an additional 5 min. Cells were washed twice by centrifugation at 500×g for 10 min at 4°C to remove excess unbound LMOs. Cell viability and concentration were then reassessed. Based on viable cell counts, 16,000 live cells per sample, up to 40,000 total cells per capture, were loaded into a primed BD Rhapsody cartridge according to the manufacturer’s instructions. Cell capture beads were dispensed into each microwell, and cells were lysed to release mRNA and LMO-derived sample barcodes. Both mRNA molecules and polyadenylated sample barcode oligonucleotides were subsequently captured by the beads through polyA/polyT interactions. After bead washing, reverse transcription was performed, following the BD Rhapsody workflow. LMO barcode library To recover LMO-derived sample barcode information, reverse transcription products not bound to the beads were collected from the supernatant and processed separately from the whole-transcriptome library. A first PCR amplification was carried out using the MULTI-seq additive primer, which specifically recognizes the sample barcode handle sequence. The expected amplicon size of approximately 138 bp was verified after AMPure XP bead purification (Beckman Coulter Life Sciences, Cat# A63880) using a 4200 TapeStation system (Agilent). A second PCR was then performed to add Illumina P5 and P7 sequencing adapters using custom reverse primers listed in Table S3. The resulting final LMO barcode library had an expected size of approximately 158 bp. Whole-transcriptome library Beads containing captured mRNA were processed using the BD Rhapsody WTA Amplification Kit according to the manufacturer’s instructions, without modifications. This workflow generated whole-transcriptome amplification libraries representative of the captured single-cell transcriptomes. Sequencing LMO barcode and whole-transcriptome libraries were pooled at equimolar concentrations and sequenced by Novogene on an Illumina NovaSeq platform using paired-end 150 bp sequencing. Sequencing depth was configured to obtain approximately 2,000 reads per cell for the LMO barcode library and approximately 50,000 reads per cell for the whole-transcriptome library, ensuring sufficient coverage for robust demultiplexing and transcript quantification. Bioinformatic processing and downstream single-cell analysis FASTQ files from whole-transcriptome and LMO barcode libraries were processed using the BD Rhapsody WTA Analysis Pipeline on the Seven Bridges platform, Revision 8. To retrieve LMO barcode sequences, the VDJ alignment module was adapted using the sample barcode sequences as a supplemental reference. This step generated filtered and unfiltered gene expression matrices, BAM files, LMO barcode count matrices, and quality-control reports. For sample demultiplexing, LMO-derived unique molecular identifier (UMI) counts were analyzed in R v4.4.0 using Seurat v5.1.0 and MULTI-seq packages. Each cell was assigned to a sample only if it contained at least 100 total LMO counts and if at least two-thirds of total LMO counts corresponded to a single sample-specific barcode. Cells that did not meet these criteria were excluded from downstream analyses, thereby reducing ambiguous assignments and minimizing misclassification due to low barcode signal or mixed labeling. For transcriptome-based quality control, cells were retained only if they showed between 1,000 and 10,000 detected genes and less than 25% mitochondrial gene content. These filtering criteria were applied after demultiplexing to remove low-quality cells prior to downstream clustering and doublet detection. To further control for artifactual cell profiles, doublets were identified computationally using scDblFinder on the quality-filtered dataset, with sample identity explicitly provided to the algorithm so that doublet classification was performed in a sample-aware manner. Cells classified as doublets were excluded, and only cells annotated as singlets were retained for downstream analyses. Normalized datasets were subjected to dimensionality reduction by principal component analysis (PCA), and principal components were selected based on the inflection point of the elbow plot. Unsupervised Louvain clustering and UMAP visualization were performed using 20 principal components and 50 nearest neighbors. For integrated analyses, samples were combined using either Seurat reciprocal PCA integration or Harmony. Integration parameters are described in the results section and corresponding figure legends. Marker genes for each cluster were identified using receiver operating characteristic (ROC) tests, and classification power was assessed by the area under the ROC curve (AUC). Positive marker genes were ranked according to adjusted p value and used to support cluster identity assignment by cross-referencing known developmental regulators and previously validated embryonic expression data, including GEISHA annotations and published chick embryonic datasets. For treatment-associated transcriptional analyses, differentially expressed genes were identified between treated and control conditions within annotated clusters or subclusters. For integration benchmarking, integrated clusters were compared with the non-integrated baseline annotation by calculating the percentage of shared positive differentially expressed genes. For each cluster, the top 200 positive marker genes ranked by adjusted p value were used for overlap comparisons. The number of positive differentially expressed genes per cluster was also quantified to evaluate whether integration reduced cluster-specific transcriptional signatures. Morphological and immunostaining analysis of posterior axis development To evaluate the effect of homocysteine treatment on posterior axis development, control and homocysteine-treated embryos were imaged and analyzed morphometrically. Global embryo size, anterior region size, and posterior region size were measured from representative dorsal and ventral images. The caudal region was further analyzed by immunostaining against TBXT (R&D Systems, Cat#AF2085) and SOX2 (eBioscience, Cat#14-9811-82) to identify mesodermal, neural, and neuromesodermal progenitor-associated domains. DAPI staining was used to visualize nuclei and delineate anatomical landmarks, including the last formed somite. Measurements included total embryo size, anterior and posterior region size, last formed somite width and length, distance from the last formed somite to the tail end, presomitic mesoderm width, posterior neural tube width, anterior neural tube width adjacent to the last formed somite, and distance from the neural tube to the tail end. Quantification was performed using predefined anatomical landmarks and compared across control, low-dose homocysteine, and high-dose homocysteine conditions. Quantification and statistical analysis Single-cell RNA-seq data processing, demultiplexing, quality control, doublet removal, clustering, integration, and differential expression analyses were performed using R v4.4.0, Seurat v5.1.0, MULTI-seq, scDblFinder, Seurat reciprocal PCA integration, and Harmony, as detailed in the method details section. Cells were included in downstream analyses only if they showed unambiguous LMO barcode assignment, 1,000–10,000 detected genes, less than 25% mitochondrial gene content, and were classified as singlets. Cells failing these criteria were excluded. Cluster marker genes and treatment-associated differentially expressed genes were identified using receiver operating characteristic tests implemented in Seurat. Classification power was assessed using the area under the receiver operating characteristic curve. For integration benchmarking, clusters were compared using the percentage of shared positive differentially expressed genes, calculated from the top 200 positive marker genes ranked by adjusted p value. Morphometric quantifications were performed using predefined anatomical landmarks. Statistical tests used for each comparison, exact values of n, and the definition of n are indicated in the corresponding figure legends. Unless otherwise stated, n represents the number of embryos for morphological analyses and the number of cells for single-cell RNA-seq analyses. Data are presented as mean ± SEM unless otherwise indicated. Statistical significance was defined as p < 0.05. Exact p values and multiple-testing correction methods are reported in the figures, figure legends, or supplemental tables. No formal sample size estimation was performed. Sample sizes were determined based on embryo availability, technical feasibility, and standard practice for exploratory embryological and single-cell RNA-seq experiments. Embryos and samples from different treatment conditions were processed in parallel under identical experimental conditions to minimize technical variation. No randomization or blinding was applied during sample preparation. Image quantification and computational analyses were performed using predefined inclusion and exclusion criteria. Additional resources •BD Rhapsody single-cell analysis resources and documentation: https://www.bdbiosciences.com/en-us/products/single-cell-multiomics. BD Rhapsody single-cell analysis resources and documentation: https://www.bdbiosciences.com/en-us/products/single-cell-multiomics. •Seurat single-cell analysis documentation: https://satijalab.org/seurat/. Seurat single-cell analysis documentation: https://satijalab.org/seurat/. •GEISHA database used to support chick embryonic gene expression annotation: https://geisha.arizona.edu/geisha/. GEISHA database used to support chick embryonic gene expression annotation: https://geisha.arizona.edu/geisha/. Footnotes Supplemental information can be found online at https://doi.org/10.1016/j.isci.2026.117392. Supplemental information Document S1. Figures S1–S3 and Tables S1–S3 Data S1. BD Rhapsody/Seven Bridges HTML pipeline report used to generate curated Seurat objects prior to downstream analysis, related to Figure 2 ∗In the report, SBC2 = SBC1. References Associated Data Supplementary Materials Document S1. Figures S1–S3 and Tables S1–S3 Data S1. BD Rhapsody/Seven Bridges HTML pipeline report used to generate curated Seurat objects prior to downstream analysis, related to Figure 2 ∗In the report, SBC2 = SBC1. Data Availability Statement •Single-cell RNA-seq data generated and analyzed in this study have been deposited in the NCBI Gene Expression Omnibus under accession number GEO: GSE303866 and are publicly available as of the date of publication. Single-cell RNA-seq data generated and analyzed in this study have been deposited in the NCBI Gene Expression Omnibus under accession number GEO: GSE303866 and are publicly available as of the date of publication. •All original code has been deposited at Zenodo at https://doi.org/10.5281/zenodo.20593689 and is publicly available as of the date of publication. The code is also available at GitHub: https://github.com/GuillotLabUCA/Maupetit-Maurelia-et-al-2026. All original code has been deposited at Zenodo at https://doi.org/10.5281/zenodo.20593689 and is publicly available as of the date of publication. The code is also available at GitHub: https://github.com/GuillotLabUCA/Maupetit-Maurelia-et-al-2026. •Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request. Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.