Abstract Background & Aims Liver regeneration is essential for recovery following injury, but this process can be impaired by factors such as sex, age, metabolic disorders, fibrosis, and immunosuppressive therapies. We aimed to identify key transcriptomic, proteomic, and serum biomarkers of regeneration in mouse models under these diverse conditions using systems biology and machine learning approaches. Methods Six mouse models, each undergoing 75% hepatectomy, were used to study regeneration across distinct clinical contexts: young males and females, aged mice, stage 2 fibrosis, steatosis, and tacrolimus exposure. A novel contrastive deep learning framework with triplet loss was developed to map regenerative trajectories and identify genes associated with regenerative efficiency. Results Despite achieving ≥75% liver mass restoration by day 7, regeneration was significantly delayed in aged, steatotic, and fibrotic models, as indicated by reduced Ki-67 staining on day 2 (p <0.0001 for all). Interestingly, fibrotic livers exhibited reduced collagen deposition and partial regression to stage 1 fibrosis post-hepatectomy. Transcriptomic and proteomic analyses revealed consistent downregulation of cell cycle genes in impaired regeneration. The deep learning model integrating clinical and transcriptomic data predicted regenerative outcomes with 87.9% accuracy. SHAP (SHapley Additive exPlanations) highlighted six key predictive genes: Wee1, Rbl1, Gnl3, Mdm2, Cdk2, and Ccne2. Proteomic validation and human SPLiT-seq (split-pool ligation-based transcriptome sequencing) data further supported their relevance across species. Conclusions This study identifies conserved cell cycle regulators underlying efficient liver regeneration and provides a predictive framework for evaluating regenerative capacity. The integration of deep learning and multi-omics profiling provides a promising approach to better understand liver regeneration and may help guide therapeutic strategies, especially in complex clinical settings. Impact and implications The aim of this study was to identify key transcriptomic, proteomic, and serum biomarkers of regeneration in mouse models under diverse conditions, using systems biology and machine learning approaches. Key molecular drivers of liver regeneration across diverse clinical conditions were identified using innovative deep learning and multi-omics approaches. By identifying conserved cell cycle genes predictive of regenerative outcomes, this study offers a powerful framework to assess and potentially enhance liver recovery in older patients, those with fibrosis or steatosis, and/or those under immunosuppression. Keywords: Liver regeneration, partial hepatectomy, deep learning, transcriptome analysis, proteome analysis Graphical abstract Image 1 [43]Open in a new tab Highlights * • Liver regeneration was delayed in aged, fibrotic, and steatotic mouse models. * • Deep learning predicted efficient vs. suboptimal regeneration with 87.9% accuracy. * • Contrastive learning captured regeneration patterns from transcriptome and serum data. * • SHAP highlighted key cell cycle genes, confirmed by proteomics and SPLiT-seq. * • Downregulation of cell cycle genes was associated with impaired liver regeneration. Introduction The liver is the only visceral organ with a remarkable capacity to regenerate, often restoring up to 70% of its original mass. Nearly 90% of the original liver mass can reconstitute from the liver remnant within 3 months post-hepatectomy (PH),[44]^1 with the highest rates of regeneration occurring in the first month.[45]^2 This unique ability is an essential component of various surgical interventions such as resection for neoplasms,[46]^3 portal vein embolization,[47]^4 and living donor liver transplantation.[48]^5 When the liver’s ability to regenerate is compromised due to underlying steatosis, fibrosis, or immunosuppression, post-hepatectomy liver failure (PHLF) – a devastating complication – can occur. PHLF accounts for over half of all PH mortality,[49]^6^,[50]^7 with affected patients facing a 6.7-fold increase in 90-day mortality and a median survival of 10 months compared to 42 months in those without PHLF.[51]^8 Additionally, the liver’s regenerative capacity may also lead to the regression of liver fibrosis once an insult is removed.[52]^9^,[53]^10 Preoperative planning to optimize outcomes critically centers around the liver’s underlying ability to regenerate and the likelihood of PHLF.[54]^11 Liver regeneration is a complicated process involving well-orchestrated activation of non-parenchymal cells in the injured area and proliferation of undamaged hepatocytes. Involvement of the Hippo-YAP,[55]^12 IL-6,[56]^13 TNF-a,[57]^14 MAPK, PI3K/STAT,[58]^15 mTORC2,[59]^16 and ALR[60]^17 pathways has been identified in regeneration of the normal mouse liver. More specifically, cell cycle regulators and checkpoint proteins such as Wee1,[61]^18 Rbl1,[62]^19 CDK2,[63]^20 and cyclin E2[64]^21 have been shown to play a role in hepatocellular proliferation. In addition, other protein modulators (e.g. GNL3, MDM2) that regulate tumor suppressors or cellular repair may be active in the regenerating liver.[65]^22^,[66]^23 However, it remains unclear why the liver does not regenerate as effectively under certain clinical conditions (e.g. steatosis, fibrosis) compared to normal livers, or how this process can be improved. Studies have shown that normal livers can regenerate twice as quickly as livers affected by hepatitis or cirrhosis, even after similar extents of resection.[67]^24^,[68]^25 Identifying the molecular pathways that play a role in regenerating the liver under various clinical conditions may yield valuable insights for the development of future therapies. While previous efforts have focused on individual signaling pathways, systems-level interpretation of regeneration dynamics across different biological contexts remains a challenge, particularly due to the complexity and high dimensionality of omics data. Deep learning has enabled the discovery of complex, nonlinear correlations between variables in high-dimensional datasets by learning hierarchical representations that capture underlying data structure beyond the capacity of traditional methods. Deep learning is a powerful tool to model biological processes by extracting meaningful patterns from noisy, high-dimensional datasets.[69]^26^,[70]^27 Contrastive learning, a particular class of deep learning used here, enables the discovery of hidden structures within high-dimensional omics data in a label efficient manner. In this study, we utilized transcriptomic and proteomic profiling to identify genes, proteins, signaling pathways, and biological processes involved in liver regeneration following 75% hepatectomy under various clinical contexts, including sex, age, fibrosis, steatosis, and immunosuppression. In parallel, we developed a novel, noise resilient approach based on contrastive machine learning to uncover the underlying manifold (simplified representation) in the space of transcriptomics profiles and serum markers that describes liver regeneration levels. To evaluate the quality of the learned representations, we developed two additional classifiers: (1) a k-nearest neighbor (k-NN) model and (2) a fine-tuning–based approach. As demonstrated by our results, such a manifold can be used to predict or classify the degree of liver regeneration. Materials and methods Animals and induction of liver diseases All C57BL/6J mice were obtained from Jackson Laboratories (Bar Harbor, ME, USA) and housed in a specific pathogen-free environment at the Animal Resources Center, Toronto, ON. All animal procedures were conducted in accordance with protocols approved by the Canadian Council on Animal Care (Animal Use Protocol 6045). Mice were housed in standard cages maintained at 22 °C with a 12:12-hour light-dark cycle and had ad libitum access to chow and water. To examine the effects of sex and age on liver regeneration, we used three groups: young male (n = 19) and young female (n = 19) mice aged 8–10 weeks, and aged male (n = 20) mice aged 33–35 weeks. To model liver regeneration under diseased conditions, we employed the following disease-specific mouse models, all of which underwent 75% partial hepatectomy: Liver fibrosis mouse model (Stage 2-3; Metavir score): Male C57BL/6J mice (8-10 weeks old) were treated with carbon tetrachloride (CCl[4]) (2 μl/g, diluted 1:4 in corn oil) (Sigma-Aldrich, #289116, Saint-Louis, MO, USA) by intraperitoneal injection twice per week over a 6- (n = 23), 9- (n = 15), or 12-week period (n = 15)[71]^28 and were fed with chow food and normal water. These mice are referred to as fibrotic males throughout the manuscript. Metabolomic-associated steatosis liver disease (MASLD) mouse model: Male C57BL/6J mice (8-10 weeks old; n = 23) were fed with an enriched fat and sucrose diet (Surwit Diet) composed of 58% fat, 17% protein, and 25% carbohydrate provided by Research Diets (#[72]D12331, New Brunswick, NJ, USA) for a 11-week period, with free access to normal water. These mice are referred to as MASLD males throughout the manuscript. Tacrolimus-treated mouse model: Tacrolimus was administered daily to both sham (n = 3) and 75% hepatectomy male mice (n = 17) until the day of sacrifice. Assessing the successful induction of liver disease conditions Resected liver tissues from mice treated with CCl[4] and fed a high-fat diet were processed for histological analysis. H&E staining was used to assess general liver morphology and pathology. Additionally, Picrosirius red staining was performed to evaluate collagen deposition, a marker of fibrosis. The presence of lipid droplets was assessed using Oil red O staining. These stains provided a comprehensive assessment of liver injury, fibrosis, and steatosis in the induced disease model. Animal anesthesia and monitoring the animal’s recovery post-surgery Mice were injected subcutaneously with slow-release buprenorphine (0.5 mg/kg) 2 h prior to surgery. For anesthesia induction, mice were placed in a Plexiglas chamber and exposed to 5% isoflurane in 100% oxygen (1.5-2 L/min). Once anesthetized, mice were transferred to a heated pad and maintained on 2% isoflurane through a suitable mouthpiece throughout the surgery. Intraoperative monitoring included assessing the pedal withdrawal reflex, ear pinch reflex, corneal reflex, respiratory rate, and heart rate. Following surgery, mice were kept in 100% oxygen until fully awake, after which they were transferred to new cages with soft bedding and heat for 1 h. To prevent dehydration, mice were injected with 500 μl of saline subcutaneously immediately after the procedure. Monitoring during the 24–72 h following surgery or tacrolimus treatment was focused on activity and behavior. We ensured that the mice were mobile and able to move around, and observed for signs of pain or discomfort, such as excessive grooming, hiding, or aggression. If any of these signs were present, veterinary advice was sought, and buprenorphine (0.5 mg/kg) was re-administered. We monitored food and water intake for up to 7 days, which marked the endpoint of the experiment. Additionally, we inspected the incision site for signs of infection, such as redness and swelling, and ensured the wound remained clean and dry. Sutures were checked for integrity. If any signs of improper recovery were observed, the mice were humanely euthanized after consultation with the veterinarian. Animal surgery The 2/3 hepatectomy was performed according to the procedure described in the manuscript of Mitchell, C and Willenbring, H (Nature Protocols, 2008). Briefly, mice were anesthetized, as described above, before the surgery. A midline laparotomy was performed via a 3 cm long abdominal skin and muscle incision to expose the xiphoid process, the abdominal organs and liver. The median and left lateral lobes were resected, and the right and caudal lobes remained for liver regeneration. The intraperitoneal cavity and organs were washed with sterile warm saline solution and the abdomen was closed with sutures. Control mice underwent laparotomy under identical analgesic and anesthetic protocols, involving a 3 cm abdominal incision to expose the liver without resection. Controls and mice that underwent hepatectomy were maintained in the same post-operative conditions, with full access to chow food and normal water, and without antibiotic treatment. Tissue and blood collection Blood was collected from live mice via the saphenous vein for serum biomarker profiling. For euthanasia, mice were re-anesthetized with isoflurane before cervical dislocation. The abdomen was carefully opened using sterile techniques. The liver was then exposed, and the surrounding tissues were gently separated. The liver was excised by cutting the connecting ligaments and blood vessels, ensuring minimal damage. The harvested liver was immediately placed in cold PBS. The resected livers were cut into small pieces for further analysis. These pieces were either snap-frozen and stored at -80 °C or preserved in RNAlater solution (Thermo Scientific, #AM7020) and stored at -20 °C, or fixed in neutral-buffered 10% formalin for paraffin embedding preparation. Serum biomarker profiling Blood samples were centrifuged at 10,000 x g for 5 min at room temperature. Mouse serum samples were transferred into new tubes and stored at -20 °C. As part of a comprehensive diagnostic profile, the concentration of albumin, alkaline phosphatase (ALP), alanine aminotransferase, amylase, total bilirubin, blood urea nitrogen, calcium, phosphorus, glucose, sodium, potassium, total protein, and globulin were measured, in 120 ml of serum, by the Animal Resources Centre (UHN Toronto) with the VETSCAN® VS2 (ABAXIS, Union City, CA, USA). For more details, see the supplementary materials and methods. Transcriptome analysis Total RNA extracted from liver tissues was profiled utilizing the Affymetrix Mouse Gene 2.0 ST array provided by ThermoFisher Scientific. The data obtained were pre-processed, annotated, and statistically analyzed using the Transcriptome Analysis Console software (ThermoFisher Scientific). For pathway analysis, differentially expressed genes identified from this process were analyzed using Ingenuity Pathway Analysis (IPA) software (Qiagen). For more details, see the supplementary materials and methods. Proteome analysis Proteomic profiling of snap-frozen liver tissues was conducted using the tandem mass tag multiplex method. The raw proteomic data obtained were processed using the PAW pipeline,[73]^29 and statistical analyses were carried out using the edgeR package[74]^30 in the R programming environment. For more details, see the supplementary materials and methods. Development and validation of a machine learning model for predicting liver regeneration efficiency A machine learning model was developed to learn the underlying manifold (a lower dimensional representation) of liver regeneration levels using transcriptomic profiles and serum markers, with Ki-67 positivity[75]^31^,[76]^32 as the primary outcome. The model architecture includes an encoder that integrates these inputs to generate embeddings reflecting the liver’s regenerative state, which typically progresses from baseline, reaches peak regeneration, and then returns to a resting level. To assess the quality of the learned representations, two additional classifiers were used: (1) a k-NN classifier to evaluate how well the embeddings distinguish between regeneration states, and (2) a single-layer neural network built on top of the encoder to enable feature interpretation through an end-to-end model. The transcriptomic input to the model consisted of 57 genes from four cell cycle-related pathways – chromosomal replication, G1/S checkpoint regulation, regulation by BGT family proteins, and cyclin control – identified through differential gene expression and pathway enrichment analyses and supported by prior studies[77][33], [78][34], [79][35], [80][36] (see Discussion). These pathways were found to be consistently downregulated in older, steatotic, and fibrotic mice compared to healthy controls. The final model was trained on expression levels of these 57 genes and 14 serum markers. The model used contrastive labels to learn a feature space that separates samples based on their transcriptome profiles and/or serum markers into two groups: efficient liver regeneration and suboptimal liver regeneration. It was trained using a triplet loss function, where anchor, positive, and negative triplets were generated by sampling from the efficient liver regeneration and suboptimal liver regeneration classes, enabling the model to focus on relative relationships. This approach learned a nonlinear, 16-dimensional projection of the input space. A 60-40 train-test split of the samples (n = 95) was used, along with 101-stratified shuffled k-fold splits to estimate performance scores. Using a contrastive learning approach with triplet loss, 101 stratified shuffled k-splits were trained over a single epoch to minimize chances of overfitting. A k-NN classifier with the default k = 5 was used to evaluate the quality of the embeddings learned by the triplet loss model through classification performance. The model was trained in two configurations: (1) transcriptomic profiles alone and (2) a combination of transcriptomic profiles and clinical variables. The model was implemented in two configurations: (1) transcriptomic profiles alone and (2) transcriptomic profiles combined with clinical variables. To define the threshold separating efficient from suboptimal liver regeneration, we analyzed Ki-67 distributions in the training data ([81]Fig. S20). A threshold of 2.4% was selected based on the observed separation between the sham or 7-day PH groups and the 2-day PH group and is further supported by prior studies (see Discussion). To ensure robustness, we conducted a sensitivity analysis by testing how slight variations in the threshold (ranging from 2.2% to 2.6%) affected the classification of samples into efficient and suboptimal regeneration groups. To interpret model predictions, we applied SHapley Additive exPlanations (SHAP) analysis,[82]^37 combined with multilayer perceptron classifiers trained on the latent embeddings. We extracted the top 10 features from the highest-performing models and identified those that most consistently discriminated between the two regeneration groups. These SHAP-prioritized features were validated using an independent proteomics dataset and human split-pool ligation-based transcriptome sequencing (SPLiT-seq) data. Details of feature importance and biological interpretation are presented in the Results. For SHAP analysis, we replaced the k-NN classifier with a neural network comprising a pre-trained encoder (triplet-loss-based), two fully connected layers separated by a dropout layer, and a softmax output. The encoder and hidden layers used ReLU (rectified linear unit)[83]^38 activation, while class weights were applied during training to address class imbalance. The encoder weights were frozen during the second phase of training, and the model was optimized using the Adagrad optimizer to minimize binary crossentropy loss. Human samples for SPLiT-seq Snap-frozen liver biopsies were obtained from two living donor liver transplantation recipients and two deceased donor liver transplantation recipients within 3 months of surgery, corresponding to the regenerative and non-regenerative phases, respectively. Samples had been stored in the Multi-Organ Transplant Program Biobank (University Health Network, REB CAPCR # 17-5311) and processed for single-nucleus RNA sequencing using SPLiT-seq, as described by Rosenberg et al. (2018).[84]^39 Frozen liver tissue was pulverized on dry ice and homogenized in nuclear isolation buffer (NIB) containing 250 mM sucrose, 25 mM KCl, 5 mM MgCl[2], 10 mM Tris-HCl (pH 7.4), 1 mM DTT, 0.1% NP-40, and RNase inhibitors. The homogenate was filtered through a 40 μm cell strainer and centrifuged at 500 x g for 5 min at 4 °C to pellet nuclei. The nuclear pellet was resuspended in NIB without NP-40, washed twice, and stained with Trypan Blue to assess nuclear integrity. Nuclei were processed using the SPLiT-seq protocol with four rounds of split-pool barcoding to uniquely label transcripts within each nucleus. In the first round, nuclei were distributed into 96-well plates containing barcoded oligonucleotide primers and reverse transcription reagents. Following reverse transcription, nuclei were pooled, washed, and redistributed for three additional rounds of barcoding using distinct sets of indexed primers. After the final barcoding step, nuclei were lysed, and cDNA was purified using SPRI beads (Beckman Coulter). cDNA libraries were amplified using Tn5 tagmentation, followed by PCR amplification with indexed primers. The final libraries were size-selected (300–800 bp) and quantified using a Bioanalyzer (Agilent) before sequencing. Libraries were sequenced on an Illumina NovaSeq 6000 platform (paired-end, 150 bp) with a targeted depth of ∼50,000 reads per nucleus. Raw sequencing reads were demuultiplexed and aligned to the human reference genome (GRCh38/hg38) using STAR, followed by UMI-based quantification. Quality control in Seurat[85]^40 (v4.0) included filtering nuclei with ≥200 detected genes, ≤5% mitochondrial RNA content, and removal of outliers to exclude doublets or degraded nuclei. Gene expression was normalized using SCTransform,[86]^41 selecting the top 3,000 highly variable genes. Batch effects were corrected using Harmony[87]^42 inside each group (regenerating/control), integrating the first 30 principal components while adjusting for different samples and processing batch differences. UMAP (uniform manifold approximation and projection) was applied to the Harmony-corrected embeddings for visualization. Graph-based clustering was performed using Seurat’s FindNeighbors() and FindClusters() functions, employing the Louvain algorithm with an optimized resolution. Statistical analysis All data are presented as mean ± SD or SEM. For comparing two groups, we used unpaired, two-tailed t-tests. When comparing three or more groups, analysis was conducted using ANOVA with Tukey’s multiple comparisons test, applicable in cases of multiple pairwise comparisons between different groups. A false discovery rate (FDR)-adjusted p value of less than 0.05 was deemed significant for all analyses, unless specified otherwise. These analyses were carried out using GraphPad Prism (version 9.3.1, Boston, MA). Results Liver-to-body weight ratio of mice after laparotomy in different models To assess whether different experimental conditions in our study ([88]Fig. 1A-D) influence the liver-to-body weight ratio, we measured this ratio across various mouse models. Liver tissues were collected from mice 2 days after laparotomy, and the liver weight was compared to the body weight of the same animals. There were no significant differences in the liver-to-body weight ratio between young male mice and the other experimental groups. However, aged mice exhibited a significantly reduced liver-to-body weight ratio ([89]Fig. 2A). Fig. 1. [90]Fig. 1 [91]Open in a new tab Mouse models for liver regeneration study. Overview of different designed mouse models for liver regeneration following partial hepatectomy. (A) young female and male mice and old male mice, (B) male mice with liver fibrosis, (C) male mice with MASLD, and (D) male mice treated with TAC. CCl[4], carbon tetrachloride; MASLD, metabolic-associated steatotic liver disease; TAC, tacrolimus. Fig. 2. [92]Fig. 2 [93]Open in a new tab Liver weight analysis across different mouse models: Sham (laparotomy) and regeneration time points. (A) Ratio of fully resected liver weight (g) to body weight (g) in sham (laparotomy) mice across different conditions. Data are presented as mean ± SEM with adjusted p values. Statistical significance was assessed using a one-way ANOVA. (B) Liver weight as a percentage of initial weight in sham (laparotomy) mice at 2 and 7 days of regeneration across various mouse models. Data are presented as mean ± SEM with adjusted p values. Statistical analyses were conducted using ANOVA followed by Tukey’s multiple comparisons test within each mouse model. MASLD, metabolic-associated steatotic liver disease. Regenerating liver weight in different mouse models To evaluate the regenerative capacity of different mouse models following 75% partial hepatectomy, we compared the weight of regenerating livers at 2 and 7 days PH with the weight of whole livers from corresponding laparotomy controls. The results demonstrated that livers from all experimental groups successfully regenerated to at least 75% of their original volume within 7 days PH ([94]Fig. 2B). Comprehensive diagnostic serum profile of sham-laparotomy and regenerating mice Serum concentrations of 14 biomarkers were profiled across all mouse groups, including laparotomy controls (n = 2-3/group), and mice at 2 and 7 days of regeneration (n = 4-8/group), to assess liver regenerative capacity. The number of mice in each group varied due to insufficient serum volume for analysis. The tacrolimus group was excluded from this profile due to inadequate serum for testing. The serum concentrations of these markers varied across laparotomy control groups. Alanine aminotransferase levels were elevated in all 2-day regeneration groups compared to their corresponding laparotomy controls, indicating hepatocellular injury from hepatic ischemia and increased bilirubin levels. ALP levels were also elevated in all groups. Additionally, reduced levels of phosphorous, calcium, total protein, and globulin were observed in the 2-day regeneration groups ([95]Table 1). Table 1. Serum comprehensive diagnostic profile of sham-laparotomy and regenerated mice. 8-10-week-old male mice 8-10-week-old female mice 33-35-week-old male mice (aged mice) Mice with fibrotic liver Mice with MASLD Sham-laparotomy __________________________________________________________________ ALB (g/L) 34 ± 0 39.5 ± 1.06∗ 33.67 ± 1.91 36.37 ± 1.66 36.33 ± 0.54∗ ALP (U/L) 36.33 ± 20.98 36.5 ± 25.81 25 ± 2.36∗ 61.33 ± 6.95∗ 28.33 ± 2.42 ALT (U/L) 57 ± 12.08 78 ± 33.94 51 ± 7.71 62.67 ± 17.55 65.33 ± 4.84 AMY (U/L) 554 ± 7.41 566.5 ± 36.416 631.5 ± 8.13∗ 658 ± 34.70 718.33 ± 34.99∗ TBIL (mmol/L) 4.33 ± 0.27 4 ± 0 4.33 ± 0.27 3.67 ± 0.27 4 ± 0 BUN (mmol/L) 7.23 ± 0.21 4.2 ± 0.21∗∗ 6.6 ± 0.60 7.53 ± 0.21 5.7 ± 0.37∗ Ca (mmol/L) 2.34 ± 0.08 2.16 ± 0.15 2.72 ± 0.07∗ 2.21 ± 0.11 2.55 ± 0.05 PHOS (mmol/L) 2.28 ± 0.16 2.56 ± 0.35 3.60 ± 0.11∗∗ 2.86 ± 0.38 2.72 ± 0.8 CRE (mmol/L) 20.33 ± 1.52 38 ± 0.71∗∗ 18 ± 0 23 ± 4.08 18 ± 0 GLU (mmol/L) 11.13 ± 0.28 9.9 ± 0.49 11.37 ± 0.12 11.57 ± 0.50 11.33 ± 0.28 Na^+ (mmol) 148.33 ± 1.19 147.5 ± 0.35 165.33 ± 3.81∗ 152 ± 2.49 153.33 ± 1.66 K^+ (mmol) 8.23 ± 0.12 7.6 ± 0 8.5 ± 0 7.53 ± 0.56 8,43 ± 0.03 TP (g/L) 46.33 ± 0.54 48 ± 2.12 52.33 ± 0.72∗∗ 44.67 ± 1.36 53 ± 0.94∗∗ GLOB (g/L) 12.67 ± 0.27 8.5 ± 1.06∗ 16.33 ± 0.72 8.33 ± 0.72∗ 16.33 ± 0.72∗ __________________________________________________________________ 2 days PHT __________________________________________________________________ ALB (g/L) 31.2 ± 0.82 37 ± 1.02 39.6 ± 2.69 35.125 ± 1.02 28.4 ± 1.99∗ ALP (U/L) 57 ± 4.37∗ 100.8 ± 7.41∗ 68.5 ± 8.52∗ 70.86 ± 5.34 73 ± 7.15∗∗ ALT (U/L) 84.8 ± 12.42 134.6 ± 42.90 113.25 ± 13.71∗ 148.71 ± 22.07 80.4 ± 6.02 AMY (U/L) 593.5 ± 24.73 567.8 ± 57.06 756.8 ± 81.04 581.38 ± 18.68 590 ± 34.31 TBIL (mmol/L) 5.2 ± 0.18 5.4 ± 0.67 5.25 ± 0.22 5.14 ± 0.24∗ 6.8 ± 1.53 BUN (mmol/L) 10.23 ± 0.99 8.84 ± 1.17 8.72 ± 1.18 6.16 ± 0.35 5.62 ± 0.47 Ca (mmol/L) 1.67 ± 0.15 1.94 ± 0.12 2.45 ± 0.21 1.96 ± 0.07 2.06 ± 0.15 PHOS (mmol/L) 2.67 ± 0.08 1.94 ± 0.12 2.45 ± 0.21∗ 2.74 ± 0.1 2.53 ± 0.19 CRE (mmol/L) 26 ± 4.01 19.6 ± 1.43∗∗ 25.2 ± 4.03 19.88 ± 1.75 19.8 ± 1.61 GLU (mmol/L) 7.78 ± 1.50 5.72 ± 0.44∗∗ 9.63 ± 0.49 7.31 ± 0.29∗∗∗∗ 6.8 ± 0.43∗∗∗ Na^+ (mmol) 150 ± 1.06 151.8 ± 1.43 173.4 ± 4.06 152 ± 1.62 147.2 ± 5.02 K^+ (mmol) 8.5 ± 0∗ 8.4 ± 0.09∗ 7.66 ± 0.47 8.13 ± 0.23 8.5 ± 0∗ TP (g/L) 36.6 ± 0.61∗∗∗∗ 41.6 ± 0.88∗ 50 ± 3.61 40.38 ± 1.12 40.8 ± 1.78∗∗ GLOB (g/L) 5.2 ± 0.72∗∗∗ 4.2 ± 0.33∗∗ 10.6 ± 1.28∗ 5.13 ± 0.33∗∗ 12.4 ± 0.83∗ __________________________________________________________________ 7 days PHT __________________________________________________________________ ALB (g/L) 40.6 ± 1.37∗ 40 ± 1.09 44.6 ± 2.74 41.83 ± 0.55∗ 34.6 ± 1.4 ALP (U/L) 53.4 ± 4.87 86.4 ± 12.87 80.4 ± 19.13 91 ± 25.15 45.4 ± 5.55 ALT (U/L) 34 ± 5.1 37.4 ± 6.43 64 ± 10.33 43.67 ± 7.17 65.8 ± 14.7 AMY (U/L) 719 ± 32.88∗ 779.8 ± 78.41 855.2 ± 61.38 712.13 ± 16.64 649.8 ± 15.04 TBIL (mmol/L) 5.8 ± 0.59 5.2 ± 0.44 5.6 ± 0,61 5.17 ± 0.55 5.4 ± 0.36∗ BUN (mmol/L) 8.62 ± 0.72 8.14 ± 0.56∗ 11.36 ± 1∗ 8.6 ± 0.36 6.92 ± 0.57 Ca (mmol/L) 2.18 ±0.05∗ 1.85 ± 0.04 2.34 ± 0.25 2.18 ± 0.05 2.38 ± 0.11 PHOS (mmol/L) 2.7 ± 0.1 3.14 ± 0.13 3.29 ± 0.11 2.98 ± 0.13 3.27 ± 0.17 CRE (mmol/L) 18 ± 0 24.8 ± 3.60 19.2 ± 1.07 18.17 ± 0.15 18 ± 0 GLU (mmol/L) 11.16 ± 0.70 10.68 ± 0.80 9.85 ± 0.19∗∗ 9.5 ± 0.77 11.24 ± 0.62 Na^+ (mmol) 149.4 ± 1.00 149.6 ± 0.88 168 ± 4.40 149.5 ± 0.90 155 ± 2.24 K^+ (mmol) 6.76 ± 0.18∗∗ 7.28 ± 0.36 8.5 ± 0 7.51 ± 0.28 8.28 ± 0.11 TP (g/L) 44 ± 0.63 44 ± 0.8 52.8 ± 3.70 43.5 ± 1.58 48.4 ± 1.85 GLOB (g/L) 3.6 ± 0.78∗∗∗ 4.2 ± 0.52∗ 8.2 ± 1.15∗∗ 2.67 ± 0.45∗∗∗ 13.6 ± 0.73 [96]Open in a new tab ALB, albumin; ALP, alkaline phosphatase; ALT, alanine aminotransferase; AMY, amylase; BUN, blood urea nitrogen; Ca, calcium; CRE, creatinine; GLOB, globulin; GLU, glucose; K+, potassium; MASLD, metabolic-associated steatotic liver disease; Na+, sodium; PHOS, phosphorus; TBIL, total bilirubin; TP, total protein. Serum extracted from mouse blood was profiled by Vetscan for the comprehensive profile. Fourteen serum markers were measured for 8-10-week male and female mice, 33-35-week male mice (aged mice), mice with fibrotic livers and mice with MASLD. Unpaired, two-tailed t-tests was performed for the statistical analysis. All data are presented as mean ± SEM. Adjusted p values: ∗<0.05, ∗∗<0.01, ∗∗∗<0.001, and ∗∗∗∗<0.0001. Regenerating liver tissues in young males, females, aged males and males treated with tacrolimus Liver tissues from young male mice and tacrolimus-treated male mice at 2 and 7 days of regeneration closely resembled those of the laparotomy controls, exhibiting minimal signs of alteration. In contrast, liver tissues from young female mice and aged male mice showed the presence of lipid droplets at 2 days of regeneration, which had resolved by 7 days, suggesting an initial phase of lipid accumulation during the early stages of liver regeneration, followed by resolution as regeneration progressed ([97]Fig. S1). Regenerating liver tissues in males with fibrotic livers Mice injected weekly with CCl[4] for 6 weeks developed liver fibrosis corresponding to Metavir stage 2-3 ([98]Fig. S2). This was accompanied by extensive collagen deposition and the activation of hepatic stellate cells, as indicated by the expression of α-SMA. However, regression of hepatic fibrosis was observed in mice at 2 and 7 days of regeneration (n = 16). This regression was marked by a significant reduction in collagen deposition and a decrease in the number of α-SMA-positive cells. Specifically, collagen deposition decreased from 4.09% in fibrotic livers to 0.88% and 0.68% in the 2-day and 7-day regenerating livers, respectively ([99]Fig. 3). Despite this regression, extended exposure to CCl[4] for 9 and 12 weeks inhibited the reduction of fibrosis during liver regeneration ([100]Figs. S3 and S4). These results indicate that even at advanced stages of fibrosis (stage 2-3), fibrotic liver tissues retain the potential to regenerate and transform into healthy liver tissue, although prolonged CCl[4] exposure may impair this regenerative capacity. Fig. 3. [101]Fig. 3 [102]Open in a new tab Regression of hepatic fibrosis during liver regeneration. (A) Collagen deposition was evaluated by Picrosirius red staining. Representative images of liver sections from control (oil injected, n = 10), CCl[4]-injected sham/resected liver tissues (n = 10), and regenerating tissues at 2 and 7 days PHT (n = 8/group). Scale bars represent 100 μm. (B) Liver sections were stained with an antibody directed against mouse α-SMA (n = 3-8 mice/group). Scale bars represent 100 μm. (C) Quantification (%) of Picrosirius red-positive areas of different liver tissues. All data are presented as mean ± SD with adjusted p values. Statistical analyses were performed using ANOVA with a Tukey’s multiple comparisons test. CCl[4], carbon tetrachloride; PHT, partial hepatectomy. Regenerating liver tissues in MASLD males MASLD mice (n = 23) displayed significant lipid accumulation in both macro- and micro-vacuoles within the liver compared to mice on a chow diet (n = 19), with lipid accumulation percentages of 44.99% and 5.57%, respectively. Serum albumin levels were significantly higher in MASLD male mice (44.52 g/L ± 0.61 g/L SEM) compared to those on the chow diet (36.5 g/L ± 0.69 g/L SEM), as were cholesterol levels (5.68 mmol/L ± 0.09 mmol/L SEM vs. 2.55 mmol/L ± 0.11 mmol/L SEM) ([103]Fig. S5). Following partial hepatectomy, liver tissues from MASLD mice displayed lipid droplets at both 2 and 7 days of regeneration. MASLD did not regress during liver regeneration ([104]Fig. S1). Genes, proteins, and signaling pathways involved in liver regeneration are influenced by various clinical liver conditions To identify the differentially expressed genes (DEGs) associated with liver regeneration under these distinct contexts, we performed a comparative analysis of gene profiles from liver samples collected at 2 and 7 days of regeneration (n = 8 per group), and compared them to those from corresponding laparotomy controls (n = 3). The number of DEGs significantly altered in each condition is provided in [105]Table S1 (FDR-adjusted p ≤0.05, p ≤0.05, and fold change ≥1.5 or ≤0.66). The highest number of DEGs was observed at 2 days of regeneration in most groups, except in the CCl[4]-injured group, where the peak occurred at 7 days of regeneration ([106]Table S1). At 2 days, only nine DEGs were common across all groups: Mybl1, Rad18, Tubb5, Tmem43, Lrrc59, Scamp5, Ccdc120, Tmem220, and Lrit2 ([107]Fig. S6). At 7 days, no DEGs were common across all groups, highlighting the differential regulation of liver regeneration under varying conditions. This finding emphasizes the involvement of condition-specific DEGs in liver regeneration. Further, we identified the DEGs shared and unique to each clinical group in this study and the group of 8–10-week-old male mice at both 2 and 7 days of regeneration, noting several common DEGs shared between groups, along with distinct DEGs for each condition ([108]Fig. S7). To investigate molecular mechanisms driving liver regeneration, we analyzed significant DEGs from day 2 of liver regeneration. The most significant pathways (p-value ≤0.05, Z-score ≥1.5 or ≤-1.5) and biological processes for each condition are shown in [109]Figs 4 and [110]S8. Our analysis revealed that no single signaling pathway was consistently significant across all mouse models, with distinct signaling pathways identified for each model ([111]Figs 4 and [112]S9; [113]Table S2). Fig. 4. [114]Fig. 4 [115]Open in a new tab Significant signaling pathways associated with each mouse model at 2 days of regeneration. Distinct significant signaling pathways were identified for each mouse model in this study. The values adjacent to each bar represent the adjusted p values for the associated pathways. Panels include: (A) young males, (B) young females, (C) aged males, (D) males with fibrotic liver, (E) males with metabolic-associated steatotic liver disease, and (F) tacrolimus-treated males. We then examined pathways related to sex, aging, fibrosis, MASLD, and tacrolimus treatment by comparing each condition to the young male group. The findings showed a complex pattern of common and distinct pathways for each clinical factor. Sex-related signaling pathways, for example, revealed 11 common pathways, with 42 distinct pathways in females and 21 in males. Aging-related pathways shared 2 common pathways, with aged male mice exhibiting 10 distinct pathways compared to 30 in young male mice. Fibrosis-related pathways shared 7 common pathways, with fibrotic mice showing 40 distinct pathways compared to 25 in young males, while MASLD-related pathways shared 4 common pathways, with 38 distinct pathways in MASLD mice and 28 in young males. Tacrolimus-related pathways shared 5 common pathways, with tacrolimus-treated mice exhibiting 6 distinct pathways compared to 27 in young males ([116]Fig. S10). To further investigate molecular dynamics, we analyzed liver protein expression in mice from different conditions subjected to laparotomy and 2 days PH (n = 3/group) using tandem mass tag and mass spectrometry. Proteins with significant differential expression (p ≤0.05, FDR-adjusted p ≤0.05, and fold change ≥1.5 or ≤-1.5) are listed in [117]Table S3. Through a focused analysis of these differentially expressed proteins (DEPs) and intersection analysis, we identified seven proteins consistently present across all models: NGAL, COQ8B, HMOX1, SPRC, CT250, CP254, and CO2 ([118]Fig. S11). We also identified the number of DEPs unique to each condition: 213 DEPs for tacrolimus-treated mice, 203 for aged mice, 162 for young female mice, 154 for fibrotic mice, 67 for MASLD mice, and 60 for young male mice ([119]Fig. S11). By comparing significant DEPs from each condition to those from the young male group, we further explored the molecular differences between conditions ([120]Fig. S12). IPA of the DEPs revealed only four common signaling pathways across all groups: MODY (maturity onset diabetes of young) signaling, LXR/RXR activation, FXR/RXR activation, and acute phase response signaling, along with several distinct signaling pathways for each condition ([121]Fig. S13). Additionally, we identified several distinct signaling pathways and biological processes identified for each model ([122]Figs. S14 and S15). We then analyzed pathways associated with sex, aging, fibrosis, MASLD, and tacrolimus treatment by comparing each condition to the young male group ([123]Table S4). The results revealed a complex interplay of shared and distinct pathways corresponding to each clinical condition. To assess the consistency of significant signaling pathways between transcriptomic and proteomic analyses at 2 days of regeneration, we conducted an intersection analysis for each group. This revealed conserved signaling pathways across groups, including 8 in young males, 45 in young females, 35 in aged males, 18 in fibrotic males, 26 in MASLD males, and 14 in tacrolimus-treated males ([124]Fig. S16). Among these, pathways were unique to specific groups, emphasizing the nuanced regulation of liver regeneration across conditions ([125]Fig. 5). In conclusion, our comprehensive analysis identified seven proteins consistently present across all experimental models, with a greater number of unique proteins specific to each condition. We also identified four common signaling pathways shared across all groups and numerous distinct pathways associated with different clinical conditions, reflecting the complexity of protein involvement in liver regeneration. The intersection analysis of transcriptomic and proteomic data revealed conserved signaling pathways unique to each group, suggesting the potential for tailored therapeutic strategies to enhance liver regeneration in specific clinical contexts. Fig. 5. [126]Fig. 5 [127]Open in a new tab Distinct signaling pathways involved in liver regeneration at 2 days post-hepatectomy in different mouse models. (A) Conserved transcriptomic and proteomic signaling pathways were analyzed using an UpSet plot to identify both shared and unique pathways across the study groups. (B) List of specific signaling pathways (transcriptomic and proteomic) involved in liver regeneration associated with each mouse model. MASLD, metabolic-associated steatotic liver disease. Comparative analysis of cell proliferation and cell cycle gene expression during liver regeneration Our data demonstrated that liver regeneration reached at least 75% of the original liver volume across all experimental conditions within 7 days, although regeneration rates at 2 and 7 days post-partial hepatectomy varied between conditions. To further explore these variations, we investigated genes and proteins associated with cell cycle regulation. Cell proliferation was assessed by counting Ki-67-positive cells at 2 and 7 days of regeneration. At 2 days, the young male mice group exhibited the highest percentage of Ki-67-positive cells (28.87%), while the lowest proliferation was observed in aged mice (1.52%) and tacrolimus-treated mice (3.15%). However, by 7 days, Ki-67-positive cell percentages were similar across all groups, closely resembling laparotomy or resected liver sections ([128]Figs 6 and [129]S17). Fig. 6. [130]Fig. 6 [131]Open in a new tab Young male mice showed the highest proliferative cell rate at 2 days of regeneration. (A) Liver sections stained with an antibody directed against mouse Ki-67 (n = 5-9 mice/group). Scale bars represent 200 mm. (B) Quantification (%) of nuclear mouse Ki-67-positive cells (n = 5-9 mice per group). All data are presented as mean ± SD with FDR-adjusted p values. Statistical analyses were performed using unpaired t test (control: young male mice group; tested groups: the other groups). MASLD, metabolic-associated steatotic liver disease; TAC, tacrolimus. Before partial hepatectomy, fibrotic mice showed the highest baseline Ki-67 expression (4.44%), reflecting the balance between cell damage and renewal in chronic inflammation. IPA revealed 102 genes involved in the cell cycle ([132]Table S5), including those regulating chromosomal replication, G1/S checkpoint, cyclin regulation, and cell cycle signaling. When comparing these genes’ expression at 2 days of regeneration across different mouse models, female mice showed the highest regulation of cell cycle-related genes, while aged males and tacrolimus-treated males had the fewest modulated genes ([133]Fig. S18, [134]Table S6). Additionally, we assessed the expression of these genes across conditions at 2 days of regeneration. Protein analysis of these 102 genes identified a higher expression of several minichromosome maintenance (MCM) family proteins in young males, females, aged males, and males with MASLD, while mice with liver fibrosis and tacrolimus-treated mice exhibited fewer cell cycle-associated proteins. These findings suggest that variations in cell cycle regulation and protein expression contribute to the observed differences in liver regeneration dynamics across the experimental conditions. An explainable contrastive deep learning model to predict efficient liver regeneration The 95 samples used in the machine learning model include transcriptomic profiles and serum markers derived from different time points (e.g. baseline, 2 days PH, and 7 days PH) across multiple experimental groups. All the samples at baseline sham-laparotomy (18), 2 days PH (48) and 7 days PH (48) were included to build the condition-agnostic model. A small number of samples (19) were excluded due to the unavailability of tissue for Ki-67 staining and analysis as well as the absence of serum markers. A sensitivity analysis to assess the robustness of the classification threshold ([135]Table S7) demonstrated stable classification performance across a range of Ki-67 thresholds (2.2% to 2.6%), supporting the selection of 2.4% as a conservative cut-off ([136]Fig. S20). The resulting feature space showed clear separation between samples with efficient liver regeneration at 2 days PH and those with reduced regeneration ([137]Fig. 7A). The triplet loss-based encoder, consisting of approximately 3,000 parameters, was intentionally trained for a single epoch to minimize overfitting. To address class imbalance, positive examples were paired with multiple negative examples during triplet generation, ensuring robust comparisons. Fig. 7. [138]Fig. 7 [139]Open in a new tab Machine learning analysis to distinguish efficient liver regeneration from suboptimal liver regeneration. (A) shows the transcriptome feature space of cell cycle genes learned by a model using a triplet loss function after one epoch. The feature space clearly separates samples with efficient liver regeneration at 2 days post-hepatectomy from those with reduced regeneration; (B) shows the confusion matrix for the average classification performance on the test data; (C) shows comparable performance between transcriptome-only and transcriptome+clinical models; (D) SHAP analysis of transcriptome-only and transcriptome+clinical models. The cell cycle gene, Wee1, is one. BUN, blood urea nitrogen; GLOB, globulin; Ca, calcium; Na+, sodium; PHT, partial hepatectomy; SHAP, SHapley Additive exPlanations. A k-NN classifier was applied to these embedded spaces, achieving 88.58% accuracy when using transcriptomic profiles alone, with mean specificity and sensitivity of 92.10% and 88.58%, respectively. Classification performance, visualized through confusion matrices ([140]Fig. 7B), showed that k-NN classifiers correctly identified efficient regeneration 92% of the time and suboptimal regeneration 80% of the time. Integrating clinical variables with transcriptomic profiles resulted in comparable performance, with an average accuracy of 87.89%, specificity of 92.44%, and sensitivity of 87.89% ([141]Fig. 7C). For SHAP analysis, only models with F1-scores above 89% were included, ensuring robust feature identification while maintaining a reasonable level of stringency. This threshold was selected based on the performance distribution, where 90% represented the median F1-score, allowing inclusion of the median model and all higher-performing models in an imbalanced dataset context. The transcriptome-only configuration ([142]Fig. 7D) identified Wee1 as the most influential feature, reflecting its critical role in mediating early liver regeneration and as a crucial cell cycle regulator for maintaining genomic stability. Other key predictors included Mcm6 and Top2a, both validated through proteomic analysis, which showed upregulation in younger and MASLD mice compared to fibrotic counterparts. Serum biomarkers such as sodium, globulin, blood urea nitrogen, and calcium were also frequently selected, indicating their potential relevance to regeneration outcomes. Single-cell profiling of human livers from transplant recipients: supporting evidence for mouse-derived liver regeneration pathways Visualization of the high-dimensional data using UMAP revealed distinct cellular compositions between the two groups. Regenerating livers displayed contributions from 9 cell clusters, while control livers exhibited 12 cell clusters, suggesting differences in cellular heterogeneity between regenerating and non-regenerating livers. Both groups contained abundant albumin-expressing cells, indicative of healthy, differentiated hepatocyte populations, confirming the presence of functional liver tissue. To further investigate the molecular mechanisms underlying liver regeneration, we examined the expression of GNL3, WEE1, RBL1, and MDM2, genes previously identified as key regulators of liver regeneration in mouse models through deep learning analysis. Our results revealed that these genes were expressed at significantly higher levels in regenerating human livers compared to control livers ([143]Fig. S19). This finding provides strong evidence that the molecular signature of liver regeneration identified in mice is conserved in humans, reinforcing the translational relevance of mouse models for studying human liver biology. Moreover, the increased expression of GNL3, associated with cell proliferation and nucleolar function, and WEE1, a key regulator of the G2/M cell cycle checkpoint, suggests active hepatocyte proliferation during human liver regeneration, mirroring findings in mouse models. Similarly, elevated levels of RBL1 and MDM2 in regenerating livers indicate enhanced cell cycle regulation and p53 pathway modulation, processes also implicated in mouse liver regeneration. Discussion In this study, we demonstrate that liver regeneration is significantly influenced by factors such as sex, aging, underlying fibrosis, MASLD, and pharmacological treatments like tacrolimus. Utilizing comprehensive transcriptomic and proteomic profiling, we identified critical cell cycle genes, signaling pathways, and biomarkers driving differential regenerative responses. Our findings reveal both common and distinct molecular signatures across different clinical conditions, highlighting the complexity and variability of the liver’s regenerative responses PH. Cell proliferation is essential for liver regeneration, particularly after partial hepatectomy. Clinically, estimating regenerative capacity informs the safe extent of liver resection. In our study, elevated ALP levels across groups suggested favorable regeneration, while decreased phosphate levels, commonly linked to consumption during hepatocyte proliferation, further supported active regenerative processes. Our analysis underscores Ki-67 as an effective marker for assessing cell proliferation and, consequently, regenerative capacity. Particularly, young male mice exhibited the highest percentage of Ki-67-positive cells at 2 days of regeneration, reflecting a robust cell proliferation response and rapid regeneration capability. This is consistent with the high regenerative capacity typically observed in younger rats[144]^43^,[145]^44 and mice.[146]^45^,[147]^46 Conversely, aged mice and tacrolimus-treated mice showed significantly lower Ki-67 positivity, indicating a compromised ability for cell proliferation and, consequently, a reduced regenerative capacity. This aligns well with existing literature suggesting that aging[148][47], [149][48], [150][49], [151][50] and immunosuppressive treatments adversely impact liver regeneration.[152]^51^,[153]^52 Clinically, the impaired regenerative response observed in aged individuals may partly explain the increased incidence of PHLF, with elderly patients (>75-years-old) being at approximately 9-fold greater risk.[154]^53 Fibrotic mice, on the other hand, showed increased Ki-67 positivity prior to hepatectomy, which may represent an ongoing balance between cell damage and renewal amid chronic inflammation.[155]^54 We further explored cell cycle regulation, identifying 102 significantly modulated genes during regeneration. Female mice exhibited greater modulation during regeneration, suggesting more active and effective proliferative regulation. In contrast, aged males and tacrolimus-treated males showed fewer modulated cell cycle genes, consistent with their lower proliferation rates and impaired regenerative responses. At the molecular level, MCM family proteins,[156]^55 crucial for DNA replication and cell cycle progression,[157]^56^,[158]^57 were significantly modulated. Elevated levels of MCM proteins were observed in young males, females, aged males, and MASLD mice, indicating active cell cycle progression and high hepatocyte proliferation rates in these groups. Conversely, fibrotic and tacrolimus-treated mice exhibited reduced levels of these proteins, which likely contributes to their impaired regenerative responses. MCM proteins form the pre-replicative complex necessary for initiating DNA replication during the G1 phase and are activated during the S phase to facilitate DNA synthesis and genomic stability.[159]^58^,[160]^59 Elevated MCM expression is typically consistent with increased hepatocyte proliferation and effective liver regeneration. Understanding how these proteins are regulated could inform therapeutic strategies to enhance regenerative responses. Based on the clinical characteristics observed in the mouse models – such as age, fibrosis, steatosis, and tacrolimus exposure – it is evident that these factors influence liver regeneration. However, since biomarkers were measured in the regenerating liver PH, they cannot directly inform preoperative prediction. Nonetheless, the observed clinical differences (e.g. reduced liver-to-body weight ratio in aged mice, altered serum profiles in fibrotic and steatotic models) suggest that a tool based solely on clinical features could potentially be developed to stratify patients before surgery. While the current molecular biomarkers PH do not support direct translation, the clinical intersections in our mouse models provide a rationale for developing and prospectively validating a tool based on clinical characteristics in humans. Future studies could explore the utility of preoperative blood-based biomarkers in combination with clinical characteristics to better predict regenerative capacity. We demonstrated the feasibility of using a contrastive learning framework to learn a robust representation of gene expression data associated with liver regeneration. Although the learned manifold may evolve with additional data, this approach provides a valuable framework for understanding the complex molecular relationships during liver regeneration. We selected a Ki-67 threshold of 2.4% based on data distribution and corroborated by prior studies. Sensitivity analysis by training five sets of 101 bootstrapped triplet loss models across thresholds (2.2–2.6%) confirmed the robustness of this cut-off. In prior studies, Gerlach et al.[161]^60 reported Ki-67 positivity of approximately 1.3% in sham-operated rats, 4%–8% during the first 16 h PH, and peak labeling indices exceeding 60% at 36–48 h post-PHT, reflecting active regeneration. Similarly, in human liver tissue, Kaita et al.[162]^61 found mean Ki-67 labeling indices of 2.5% in normal livers and 1.4% in inactive cirrhosis. In contrast, significantly higher values were observed in mild chronic hepatitis (28.8%), severe chronic hepatitis (41.7%), and hepatocellular carcinoma (71.3%). Although some normal samples exhibited Ki-67 indices up to 10%, the average values remained below 2.5%, supporting the use of 2.4% as a conservative threshold indicative of elevated regenerative activity. However, determining a precise threshold for Ki-67 positivity to distinguish between efficient and suboptimal regeneration is challenging due to variability across species, the underlying liver condition, extent of resection, and timing of measurement. We demonstrate that the proposed approach remains stable across subtle variations in the classification threshold and its robustness is likely to improve further with additional training data. Our contrastive triplet-loss-based model effectively distinguished efficient from suboptimal regeneration, achieving high predictive accuracy. SHAP analysis identified Wee1, Rbl1, Gnl3, Mdm2, Cdk2, and Ccne2 as the most influential predictors of regenerative capacity. These genes regulate critical checkpoints and proliferation pathways essential for hepatocyte cell cycle progression, genomic stability, and liver tissue repair. Their relevance was confirmed in an independent proteomics dataset and human SPLiT-seq data, supporting model robustness and biomarker validity. Importantly, contrastive learning mitigates label noise by focusing on relative similarity between samples, making it suitable for datasets with empirically defined thresholds like Ki-67. Even with borderline values, the model maintains stability by preserving semantic relationships in the embedding space. We used k-NN classification to directly utilize the learned distances between embeddings for classification. This approach provides an unbiased assessment of the quality of the learned representation, independent of any additional model assumptions. Despite the small dataset and class imbalance, we employed several methodological safeguards to minimize overfitting and enhance generalizability. These included dimensionality reduction through differential expression and pathway enrichment analysis, adding class weights, limiting the encoder to three hidden layers, and using contrastive triplet loss to learn robust, generalizable representations. The contrastive framework, by comparing positive and negative pairs, encourages the model to focus on relative similarities rather than memorizing specific data points. This implicitly serves as a form of data augmentation,[163]^62 helping the model learn fine-grained biological differences. We further applied 101-stratified shuffled k-fold splits and validated our findings using independent proteomic and human SPLiT-seq datasets. Nonetheless, larger and more diverse datasets will be essential to confirm and expand these insights. The preprocessing step that identified the 57 genes used to train the triplet loss network was essential to reduce model complexity, given the limited training data. While a larger neural network (wider with more inputs and with more hidden layers) likely can learn the association between key genes and liver regeneration without the need for initial filtering, such a model will have at least an order of magnitude more parameters than the current model and would require a much larger training dataset. SHAP analysis revealed consistently significant predictors across model configurations, with Wee1, Rbl1, Gnl3, Mdm2, Cdk2, and Ccne2 emerging as consistently significant across model configurations. These genes are integral to the regulation of cell cycle progression, DNA damage response, and cellular proliferation, all of which are crucial for hepatocyte regeneration and liver tissue repair. Specifically, Wee1 and Cdk2 modulate the G1/S and G2/M transitions,[164]^20^,[165]^63^,[166]^64 ensuring proper cell cycle checkpoints and mitotic entry, while Ccne2 and Rbl1 (p107) facilitate E2F-mediated transcriptional control of proliferation.[167]^65^,[168]^66 Mdm2, a key negative regulator of p53, modulates cell survival and apoptosis,[169]^67^,[170]^68 whereas Gnl3 (nucleostemin) is involved in ribosomal biogenesis and stem cell maintenance, processes critical for tissue homeostasis and regenerative potential.[171]^69^,[172]^70 The expression dynamics of Gnl3 and Cdk2 suggest sex- and age-dependent regulatory mechanisms influencing liver regeneration.[173]^20^,[174]^70^,[175]^71 Gnl3 exhibited heightened expression in young females, aged males, and tacrolimus-treated males, implying a role in regenerative competency under distinct physiological and pharmacological conditions. This pattern may reflect interactions between hormonal signaling, immune modulation, and stem cell activity in liver repair. Additionally, Cdk2’s upregulation in young females suggests a heightened proliferative response,[176]^20 potentially driven by estrogen-mediated enhancement of cyclin-dependent kinase activity, contributing to increased hepatocyte turnover and tissue restoration.[177]^72^,[178]^73 In summary, our model effectively distinguished regeneration outcomes using cell cycle-related transcriptomic profiles, with added value from clinical biomarkers. Proper regulation of these genes is essential for efficient regeneration, while dysregulation may contribute to poor outcomes. Integrating transcriptomic with clinical data poses challenges due to variability and noise. Larger, annotated datasets are needed to fully understand the impact of differential gene expression and to improve generalizability. Independent dataset validation remains essential to confirm biological relevance. A challenge with using SHAP for interpretability is that SHAP may underrepresent correlated features, potentially obscuring predictive variables. Although our initial feature selection step helps reduce redundancy, future work will explore decorrelation strategies and alternative interpretability methods. To partially address this issue, instead of averaging feature importance across models, we identify the features that most frequently appear among the top 10 important features. While this does not entirely eliminate the risk of missing correlated features, it mitigates the issue to some extent, as each model is trained independently. While mouse models offer insights, species differences limit direct translatability. Single-cell profiling of human livers is challenging due to limited biopsy tissue and nuclear-only preservation in frozen samples. Nevertheless, our single-cell profiling of human liver transplant recipients confirmed conservation of key genes – GNL3, WEE1, RBL1, and MDM2 – highlighting their relevance to human liver regeneration. This supports the validity of our mouse-derived signatures and provides a foundation for further translational studies. In conclusion, our study enhances our understanding of liver regeneration by delineating condition-specific molecular signatures and identifying predictive biomarkers that can inform personalized therapeutic strategies. These findings underscore the utility of integrating systems biology with machine learning approaches to advance precision medicine in liver regeneration, ultimately aiming to improve clinical outcomes for patients undergoing hepatic surgery or transplantation. Abbreviations ALP, alkaline phosphatase; CCl[4], carbon tetrachloride; DEGs, differentially expressed genes; DEPs, differentially expressed proteins; IPA, Ingenuity Pathway Analysis; k-NN, k-nearest neighbors; MASLD, metabolic dysfunction-associated steatotic liver disease; MCM, minichromosome maintenance; PH, post-hepatectomy; PHLF, post-hepatectomy liver failure; SHAP, SHapley Additive exPlanations; SPLiT-seq, split-pool ligation-based transcriptome sequencing. Financial support This research received funding from the Natural Sciences and Engineering Research Council of Canada (NSERC) and was also supported by the University Health Network (UHN) Foundation. Authors’ contributions A.T. Nguyen-Lefebvre and G. Oldani carried out the experiments. S. Ghosh developed the machine learning algorithm. S. Ghosh and C. Baciu were responsible for analyzing the results and drafting the manuscript. BJ. Hasjim, N. Selzner, J. Wrana, S. Naimimohasses, M. Brudno, and E. Pasini contributed to the manuscript through their review. A.T. Nguyen-Lefebvre, S. Ghosh, and M. Bhat played a pivotal role in designing and conducting experiments, analyzing the outcomes, and co-authoring the manuscript. Conflict of interest The authors declare no competing interests. Please refer to the accompanying ICMJE disclosure forms for further details. Acknowledgements