Machine learning and experiment reveals gut microbial predictors of functional fatty acids in pork

Abstract

High-quality pork is rich in functional fatty acids such as α-linolenic (ALA), γ-linolenic acid (GLA), docosahexaenoic acid (DHA), and eicosapentaenoic acid (EPA). While gut microbiota influence muscle fatty acid deposition, its complexity remains elusive. Using machine learning, we analyzed data from 291 pigs to identify microbial predictors of muscle fatty acids linked to meat quality. The best-performing models revealed that Escherichia was negatively associated with muscle fatty acid traits, whereas Bradyrhizobium and Lachnoclostridium were positively associated with the deposition of several functional fatty acids, particularly GLA, DHA, and EPA. Furthermore, Lachnoclostridium edouardi (LE) was selected for murine experiments, which revealed that LE administration markedly increased the levels of TFA, unsaturated fatty acids (UFA), and polyunsaturated fatty acids (PUFA) in mouse muscle tissue. Collectively, these findings establish a cross-study, interpretable framework for prioritizing gut microbial candidates associated with health-promoting fatty acid deposition in pork.

Highlights

• •Microbial α diversity exhibited positive predictive power for muscle PUFAs.
• •A core set of 11 genera from the pig gut microbe correlated with muscle fatty acids.
• •Escherichia was negatively associated with the levels of most fatty acids.
• •Lachnoclostridium were positive correlations of TFA and PUFA.
• •Lachnoclostridium edouardi administration markedly increased PUFA in muscle.

1. Introduction

Pork is one of the most widely consumed meats worldwide, and its nutritional and sensory quality is strongly influenced by intramuscular fat deposition and fatty acid composition (Jiang et al., 2023). Specifically, muscle fatty acid contents and types [saturated fatty acids (SFA) and unsaturated fatty acids (UFA)] are highly associated with the quality of pork and the health of consumers. For example, excess intake of SFAs is positively correlated with the fatty liver index, accompanying with a low-grade inflammatory response and lipid accumulation (Schoeler et al., 2023). Thus, low-fat diets (especially in terms of low SFAs) can be introduced to improve lipid metabolism or fatty liver. In addition, UFAs, such as C18:1n9c (oleic acid), C18:3n3 (α-linolenic, ALA), C18:3n6 (γ-linolenic acid, GLA), C22:6n3 (docosahexaenoic acid, DHA), and C20:5n3 (eicosapentaenoic acid, EPA), have been widely investigated for their anticancer effects, their ability to alleviate cardiovascular disease, and lifespan extension (Papsdorf et al., 2023a, Papsdorf et al., 2023b; Zhang et al., 2025). Therefore, identifying biological factors that favor the deposition of beneficial fatty acids in pork has both nutritional and production relevance.

The gut microbiota is increasingly recognized as a major regulator of host lipid metabolism (Ma et al., 2022a, Ma et al., 2022b). Previous studies in pigs have shown that gut microbial composition is associated with muscular fatty acid profiles and meat-quality traits(Ma et al., 2022a, Ma et al., 2022b; Yin et al., 2023). However, most of these studies have focused on single cohorts or specific genetic and nutritional backgrounds, making it difficult to distinguish cohort-specific patterns from microbial predictors that remain informative across heterogeneous pig populations (Martinez-Guryn et al., 2018). In addition, the complexity and high dimensionality of microbiome data limit straightforward interpretation. Machine learning provides a useful strategy for integrating multidimensional datasets and prioritizing informative predictors from complex microbial features. (Qi et al., 2025; Stokes et al., 2020).

In this study, we integrated data from 291 pigs across six independent studies involving different breeds, body weights, dietary treatments, and rearing environments. We developed and compared four machine-learning models – random forest (RF), extreme gradient boosting (XGB), LightGBM (LGB), and support vector machine (SVM) - to connect meat-quality traits, muscle fatty acid profiles, gut microbial composition, and microbial functional pathways. The optimal model was selected based on model performance evaluation, and the SHAP algorithm was used to assess the contribution of gut microbiota to model outputs as well as the direction of their associations. Based on these analyses, potential candidate microbial genera associated with the production of high-quality pork traits and UFA-rich pork were identified. Finally, selected candidate microbial genera were validated in mice to further investigate the regulatory effects of machine learning-prioritized microbes on muscle fatty acid deposition in animals.

2. Materials and methods

2.1. Biological sample collection and related data

The overall workflow of the present study is summarized in Fig. 1, including data collection from six independent pig studies, data preprocessing and feature construction, machine-learning model development, SHAP-based interpretation, and in vivo validation of prioritized microbial candidates. In this study, a dataset of 291 samples was collected from six pig studies containing fatty Ningxiang pigs (raised in Hunan Province), fatty Wujing pigs (raised in Yunnan Province), mini Bama pigs (raised in Hunan Province), and lean Duroc×Landrace×Yorkshire hybrid pigs (DLY, raised in Hunan and Hubei Provinces) with different body weights and dietary treatments (Supplementary Table 1). The marbling score, pH index, and meat color (a, b, and L value) (at 45 min or 24 h after slaughtering) of the longissimus dorsi muscle between the 6th and 7th ribs were used to evaluate meat quality, and the detection methods are described in our previous study(Ma et al., 2025; Yin et al., 2023). Thirty-one fatty acids were assessed using gas chromatography, and TFA, SFA, UFA, monounsaturated fatty acids (MUFA), polyunsaturated fatty acids (PUFA), the SFA/TFA ratio, the UFA/TFA ratio, the MUFA/TFA ratio, and the PUFA/TFA ratio were calculated as predictors of meat quality (Supplementary Table 2). The gut microbiota was tested by full-length 16S rDNA sequencing (DLY raised in Hunan Provinces and Ningxiang pigs) (Yin et al., 2023), metagenomics (DLY raised in Hunan Provinces and Ningxiang pigs) (Yin et al., 2023), or 16S rDNA sequencing (DLY raised in Hubei Provinces, fatty Wujing pigs and Bama pigs). Community evenness and richness were evaluated by microbial α diversity (ACE, Chao1, Simpson, and Shannon indices). Thirteen phyla, the top 50 genera, and microbial functional annotations (level 3) according to KEGG were used as predictors of muscle fatty acids (Supplementary Table 3).

Fig. 1.

Overall workflow of the study integrating multi-cohort pig microbiome data, machine-learning modelling, and experimental validation.

2.2. Machine learning

The integrated dataset was preprocessed by removing outliers, harmonizing units, imputing missing values, and normalizing input and output variables to a comparable scale. Missing values were imputed using the K-nearest neighbors (KNN) method. Because different algorithms vary substantially in how they handle high-dimensional data and assign feature importance, we employed four widely used machine-learning models in gut microbiota research—random forest (RF), extreme gradient boosting (XGB), LightGBM (LGB), and support vector machine (SVM)—to identify robust key gut microbial features and improve the reliability of the findings. For the RF, XGB, and LGB models, exhaustive grid search over all hyperparameters would have resulted in excessive computational cost and prolonged training time. Therefore, to improve tuning efficiency and reduce model development time, we adopted a stepwise hyperparameter optimization strategy. First, a relatively high learning rate was used to rapidly identify an appropriate range for the optimal tree structure. Next, tree-specific parameters were tuned, including max_depth, min_child_weight, gamma, subsample, and colsample_bytree. This was followed by optimization of regularization parameters (lambda and alpha), which reduce model complexity and can improve model performance. The models were developed based on the type and features of the training dataset, and the hyperparameters were adjusted according to the dataset's characteristics and training performance (Zoabi et al., 2021). After optimizing the hyperparameters of the machine learning models, the training performance and prediction accuracy of the models were evaluated on the test dataset. The models were gradually optimized and calibrated based on their performance under different parameters.

The preprocessed data points were randomly divided into a training subset (70%) and a testing subset (30%) for multiple training sessions. To ensure the relative independence of the training and validation sets, reduce the risk of overfitting, and improve model stability, ten-fold cross-validation (10-fold cross-validation, 10-fold CV) was used to optimize hyperparameters, thereby enhancing the predictive performance and generalization ability of the models. Meanwhile, an independent test set was retained for external validation of model performance. The performance of the models was evaluated using the coefficient of determination (R²). The permutation importance provides a model-independent method for calculating the importance of features. Randomly permuting a feature in the dataset and comparing the differences between the new and original results allowed us to determine the impact of that feature on the model when it changed. After constructing the ensemble models, permutation importance was used to measure the importance of individual features in the prediction process. The change in prediction performance was evaluated by eliminating each feature (randomly shuffling the feature 50 times). SHapley Additive exPlanations (SHAP) generates a value (referred to as the SHAP value) for each input feature that reflects the feature importance for a specific sample [eq. (1)]. Some factors had a positive impact on the prediction probability, while others had a negative impact (Zoabi et al., 2021).

$${\text{SHAP}}_{\mathit{feature}}\left(\mathrm{x}\right)=\sum _{\mathrm{s}\mathit{et}:\mathit{feature}\in \mathit{set}}{\left[\left|\mathit{set}\right|\times \left(\genfrac{}{}{0pt}{}{F}{\mathit{set}}\right)\right]}^{-1}\left[{\mathit{Predict}}_{\mathit{set}}\left(x\right)-{\mathit{Predict}}_{\mathit{set}\backslash \mathit{feature}}\left(x\right)\right]$$ (2)

2.3. Lachnoclostridium edouardi treatment

All animal experiments reported in this manuscript were approved by the Hunan Agricultural University Institutional Animal Care and Use Committee (HAU-2020-05). The experimental procedures complied with the Guidelines for the Care and Use of Laboratory Animals stipulated in the “Regulations for the Management of Laboratory Animals” and the “Measures for the Management of Laboratory Animals in Hunan Province.” Six-week-old male ICR mice were obtained from SLAC Laboratory Animal Center (Changsha, China). Mice were randomly assigned to four groups (Control, LE1, LE2, and LE3; n = 11 per group) and subjected to a 42-day intervention. The control group was fed a basal diet, while the LE1, LE2, and LE3 groups were fed the basal diet supplemented with Lachnoclostridium edouardi at doses of 1 × 10⁸ CFU/kg, 1 × 10⁹ CFU/kg and 1 × 10¹⁰ CFU/kg(Chen et al., 2016; Zhu et al., 2021). Throughout the experiment, all animals had ad libitum access to food and water. At the end of the experiment, the mice were euthanized by anesthesia followed by cervical dislocation, in accordance with the institutional guidelines for animal care and use. Subcutaneous adipose tissue (SAT), abdominal adipose tissue (AAT), and perirenal adipose tissue (PEAT) were weighed, and skeletal muscle fatty acids were collected and subsequently determined.

2.4. Analysis of medium- and long-chain fatty acids in muscle

Based on previously described methods(Tian et al., 2025; Yin et al., 2023), thirty-one fatty acids in muscle samples from the control and LE3 groups were quantified using GC–MS, and the contents of TFA, SFA, UFA, MUFA, and PUFA were calculated.

2.5. Data analysis

Model performance was evaluated using multiple regression metrics, including the coefficient of determination (R²) and root mean square error (RMSE). Final model selection was based primarily on R², while RMSE and cross-validation variability were also taken into consideration to provide a more comprehensive assessment of predictive performance. The SHAP algorithm was subsequently applied to analyze and interpret the feature effects of the model and to select and identify the most important features. In SHAP plots, features are ranked from top to bottom by importance. Red and blue indicate higher and lower feature values, respectively. When red points are concentrated on the right, the feature is positively associated with the predicted trait; when concentrated on the left, it is negatively associated. Greater concentration indicates a stronger effect. For features exceeding 10, the top 10 features were selected for analysis. In addition, due to the nonnormal distribution of the Escherichia data, the correlations between fatty acid degradation and Lachnoclostridium, Escherichia, and Bradyrhizobium were analyzed using Spearman correlation rather than the SHAP algorithm.

Mouse experimental data are expressed as mean ± standard error of the mean (SEM). Statistical analysis was performed using one-way analysis of variance (ANOVA) and subsequent Tukey's test or Student's t-test, performed with IBM SPSS Statistics 20. A p-value <0.05 was defined as statistically significant. All graphs were produced via GraphPad Prism 9.

3. Results

3.1. Muscle fatty acids, particularly UFAs, were predictive of meat quality

The predictive ability of the four machine learning models (RF, XGB, LGB, and SVM) was evaluated by training on a dataset of 291 samples from six pig studies (Supplementary Table 1). The prediction performance was ranked as RF > XGB > LGB > SVM, and the high regression coefficient (R²) values (≥0.83) indicated that the RF model was suitable for predicting the influence of muscle fatty acids on the meat quality (Fig. 2). For RMSE, only LGB showed a relatively high value. Therefore, the RF model was further used to predict the relative importance of muscle fatty acids in terms of meat quality. A set of the most informative features of muscle fatty acids, also referred to as ‘predictors’, was extracted from the RF model. The results revealed that muscle PUFAs and the MUFA/TFA ratio had strong positive predictive power for meat color, but the importance of PUFAs decreased and the MUFA/TFA ratio increased at 24 h after sacrifice (Supplementary Fig. 1 A, B). Supplementary SHAP plots further showed that the SFA/TFA ratio had a significant negative effect on the b (Supplementary Fig. 1C, D) and L values (Supplementary Fig. 1 E, F). The marbling score, which is an intramuscular fat correlation index, was positively associated with the muscle UFA/TFA ratio, especially the MUFA/TFA ratio (Supplementary Fig. 1 G). The muscle pH index reflects the oxidative reactivity of fatty acids, and the pH index at 45 min was positively related to PUFA content, while it was negatively correlated with TFA at 24 h (Supplementary Fig. 1H, I).

Fig. 2.

Machine learning models for assessing the effect of muscle fatty acids on meat quality. (A) Comparison of model performance based on R² for specific fatty acids. (B) Comparison of model performance based on RMSE for specific fatty acids. (C) Comparison of model performance based on R² for classifying fatty acid variables. (D) Comparison of model performance based on RMSE for classifying fatty acid variables.

Specifically, 31 fatty acids in the muscle were evaluated for relative importance, and 6 features were found to be predictors of meat quality. The C22:0 content had a significant correlation with the meat color a value (Fig. 3A, B) and marbling score (Fig. 3 G). For the b value, C21:0 had a significant positive correlation at 45 min (Fig. 3 C), and C18:3n3 was positively correlated at 24 h (Fig. 3 D). C12:0 and C21:0 were negatively correlated with the L value at 45 min (Fig. 3 E) and 24 h (Fig. 3 F), respectively. Muscle pH was positively related to C18:2n6c content at 45 min (Fig. 3 H), while a negative correlation with C20:1 was observed at 24 h (Fig. 3 I). In summary, higher levels of muscle fatty acids were associated with more desirable meat-quality pork with better meat color, and increasing UFA abundance was predicted to improve the marbling score of meat. However, muscle C21:0 and C20:1 were not desirable due to the negative effects on meat color and pH.

Fig. 3.

The importance of attributes and correlation of specific fatty acids on meat quality.(A) a value at 45 min. (B) a value at 24 h. (C) b value at 45 min. (D) b value at 24 h. (E) L value at 45 min. (F) L value at 24 h. (G) Marbling score. (H) pH at 45 min. (I) pH at 24 h. In each panel, features are ranked from top to bottom by importance. When red points are concentrated on the right, the feature is positively associated with the predicted trait; when concentrated on the left, it is negatively associated. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

3.2. Microbial α diversity correlates with desirable meat-quality pork rich in PUFAs

The prediction models of the gut microbiota demonstrated that, in terms of microbial α diversity and microbial abundance at the phylum level, the RF model had a better fitting performance than did the XGB, LGB, and SVM models (Fig. 4A, B. R² ≥ 0.87 for α diversity; R² ≥ 0.85 for microbial phyla). Subsequent relative importance analysis was performed using the RF model for the prediction of microbial α diversity and phyla. At the genus level and for microbial functional annotations, the best prediction performance was observed for the LGB model (Fig. 4 C, D. R² ≥ 0.92 for microbial genera; R² ≥ 0.93 for microbial functional annotations). SVM showed a higher RMSE value (Fig. 4). Thus, the roles of genera and microbial pathways in determining muscle fatty acid levels were predicted using the LGB model.

Fig. 4.

Machine learning models for assessing the effect of the gut microbiota on muscle fatty acids. (A) Model comparison of α diversity based on R² and RMSE. (B) Model comparison of phylum-level microbial abundance based on R² and RMSE. C) Model comparison of genus-level microbial abundance based on R² and RMSE. (D) Model comparison of microbial functional annotations at KEGG level 3 based on R² and RMSE.

Community evenness and richness are generally evaluated by microbial α diversity using the ACE, Chao1, Simpson, and Shannon indices. According to the relative importance calculated by the RF model, ACE exhibited a negative correlation with muscle MUFAs (Fig. 5 A), PUFAs (Fig. 5 B), UFAs (Fig. 5 C), SFAs (Fig. 5 D), and TFAs (Fig. 5 E). Data mining revealed that a core subset, which contributed to microbial α diversity, exhibited strong predictive power for muscle-specific fatty acids. Primarily, microbial α diversity (Chao1 and ACE) was positively correlated with most UFAs (i.e., C14:1, C17:1, C18:3n6, C20:3n3, C20:4n6, C20:5n3, C22:1n9, and C22:6n3) and negatively correlated with muscle SFAs (i.e., C6:0, C8:0, C10:0, C12:0, C14:0, C15:0, C16:0, C17:0, C18:0, C20:0, and C21:0) (Fig. 5 F—O and Supplementary Fig. 2 A-U). In summary, greater gut microbial evenness and richness was associated with pork rich in GLA, DHA, and EPA.

Fig. 5.

The importance of the attributes and the correlation of α diversity with significant fatty acids. (A) MUFA. (B) PUFA. (C) UFA. (D) SFA. (E) TFA. (F) C12:0. (G) C18:2n6c. (H) C18:3n3. (I) C20:1. (J) C21:0. (K) C22:0. (L) C18:1n9c. (M) C18:3n6 (N) C20:5n3. (O) C22:6n3. In each panel, features are ranked from top to bottom by importance. When red points are concentrated on the right, the feature is positively associated with the predicted trait; when concentrated on the left, it is negatively associated. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

3.3. The gut microbiota is involved in fatty acid deposition in muscle

According to the relative importance results calculated by the RF model at the phylum level, two phyla were positively correlated with muscle fatty acids. Notably, one of the subgraphs containing muscle UFAs (Fig. 6 C) and one containing TFAs (Fig. 6 E) was found in the gut, including Actinobacteria. The subgraphs correlated with Spirochaetes contained MUFAs (Fig. 6 A), PUFAs (Fig. 6 B), and SFAs (Fig. 6 D).

Fig. 6.

The importance of the attributes and correlations of microbial abundance at the phylum level with significant fatty acids. (A) MUFA. (B) PUFA. (C) UFA. (D) SFA. (E) TFA. (F) C12:0. (G) C18:2n6c. (H) C18:3n3. (I) C20:1. (J) C21:0. (K) C22:0. (L) C18:1n9c. (M) C18:3n6 (N) C20:5n3. (O) C22:6n3. In each panel, features are ranked from top to bottom by importance. When red points are concentrated on the right, the feature is positively associated with the predicted trait; when concentrated on the left, it is negatively associated. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

For each specific fatty acid, Actinobacteria was negatively correlated with muscle C6:0 (Supplementary Fig. 3 A), C8:0 (Supplementary Fig. 3 B), C10:0 (Supplementary Fig. 3C), and C12:0 (Fig. 6 F), and positively correlated with C22:2 content (Supplementary Fig. 3 R). Spirochaetes was positively correlated with C14:0 (Supplementary Fig. 3 D), C15:0 (Supplementary Fig. 3 F), C16:0 (Supplementary Fig. 3 G), C16:1 (Supplementary Fig. 3H), C17:0 (Supplementary Fig. 3 I), C17.1 (Supplementary Fig. 3 J), C18:2n6c (Fig. 6 G), C18:1n9c (Fig. 6 L), C18:3n6 (Fig. 6 M), C20:0 (Supplementary Fig. 3 L), C20:1 (Fig. 6 I), C20:2 (Supplementary Fig. 3 M), and C20:4n6 (Supplementary Fig. 3 P) depositions in the muscle. Muscle C18:0 (Supplementary Fig. 3 K), C20:3n6 (Supplementary Fig. 3 O), and C21:0 (Fig. 6 J) concentrations were negatively correlated with Proteobacteria. C18:3n3 (Fig. 6 H) and C20:3n3 (Supplementary Fig. 3 N) were principally, positively correlated with Cyanobacteria. Firmicutes (negative for C22:0, C23:0, and C24:0) (Fig. 6K; Supplementary Fig. 3 S, T) and Bacteroidetes (positive for C20:5n3 and C22:6n3) (Fig. 6 N, O) were the two most abundant phyla in most pigs, and the F/B ratio, a microbial marker for lipid metabolic disorder, was positively correlated with C14:1 (Supplementary Fig. 3 E) and C24:1 (Supplementary Fig. 3 U) but negatively related to C22:1n9 (Supplementary Fig. 3 Q) and C22:1 (Supplementary Fig. 3 Q).

According to the predictions of the LGB model, the relative importance of gut microbiota at the genus level in predicting muscle fatty acids was further evaluated. Only Escherichia, a common pathogen in the gut, was identified to be negatively correlated with muscle fatty acids, including MUFAs (Fig. 7 A), PUFAs (Fig. 7 B), UFAs (Fig. 7 C), SFA (Fig. 6 D), and TFAs (Fig. 7 E), indicating that higher Escherichia abundance was generally unfavorable for muscle fatty acid deposition in pigs.

Fig. 7.

The importance of the attributes and correlations of microbial abundance at the genus level with significant fatty acids. (A) MUFA. (B) PUFA. (C) UFA. (D) SFA. (E) TFA. (F) C12:0. (G) C18:2n6c. (H) C18:3n3. (I) C20:1. (J) C21:0. (K) C22:0. (L) C18:1n9c. (M) C18:3n6 (N) C20:5n3. (O) C22:6n3. In each panel, features are ranked from top to bottom by importance. When red points are concentrated on the right, the feature is positively associated with the predicted trait; when concentrated on the left, it is negatively associated. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

The relationships between gut genera and specific fatty acids were further analyzed. Eleven of 50 genera were associated with muscle fatty acids, including Psychrobacter (C6:0, C8:0, and C10:0) (Supplementary Fig. 4 A-C), Bacteroides (C12:0) (Fig. 7 F), Escherichia (C14:0, C14:1, C15:0, C16:0, C16:1, C18:0, C18:1n9c, C18:2n6c, C20:0, C21:0, and C22:1n9) (Fig. 7 G, J, L and Supplementary Fig. D, F, G, H, K, L), Clostridium (C17:0, C20:1, C20:3n6, and C22:2) (Fig. 7 I and Supplementary Fig. 4 I, O, R), Campylobacter (C17:1 and C20:3n3, C22:0, and C24:0) (Fig. 7 K and Supplementary Fig. 4 J, N, T), Bradyrhizobium (C18:3n6) (Fig. 7 M), Phascolarctobacterium (C18:3n3, C20:5n3, and C22:6n3) (Fig. 7 H, N, O), Terrisporobacter (C20:2) (Supplementary Fig. 4 M), Ruminococcus (C20:4n6) (Supplementary Fig. 4 P), Corynebacterium (C23:0) (Supplementary Fig. 4 S), and Dehalobacter (C24:1) (Supplementary Fig. 4 U). Overall, Bradyrhizobium, Phascolarctobacterium, Terrisporobacter, Ruminococcus, and Dehalobacter were the main species associated with the abundance of very long-chain UFAs, such as C18:3n3, C18:3n6, C20:2, C20:4n6, C20:5n3, C22:6n3, and C24:1.

3.4. Gut microbiota-targeted fatty acid degradation is negatively correlated with muscle fatty acid deposition

To further investigate the relationships among gut microbiota, muscle fatty acids, and gut KEGG pathways, with particular emphasis on the fatty acid degradation pathway, we analyzed level 3 microbial functional annotations using the LGB model (Supplementary Figs. 5 and 6). Notably, UFAs (Supplementary Fig. 5C), especially PUFAs (Supplementary Fig. 5 B), were negatively correlated with microbial metabolism, likely related to fatty acid degradation. Indeed, the fatty acid degradation pathway was negatively correlated with muscle C8:0 (Supplementary Fig. 6 B), C17:1 (Supplementary Fig. 6 J), C18:2n6c (Supplementary Fig. 5 G), C18:3n3 (Supplementary Fig. 5H), C18:3n6 (Supplementary Fig. 5 M), C20:2 (Supplementary Fig. 6 M), C20:3n3 (Supplementary Fig. 6 N), C20:4n6 (Supplementary Fig. 6 P), C20:5n3 (Supplementary Fig. 5 N), C22:0 (Supplementary Fig. 5 K), C23:0 (Supplementary Fig. 6 S), and C24:0 (Supplementary Fig. 6 T) content. Microbial metabolism in diverse environments and muscle MUFAs (Supplementary Fig. 5 A), SFAs (Supplementary Fig. 5 D), and TFAs (Supplementary Fig. 5 E) were also correlated, indicating the potential role of environment-related microbial changes in the muscle fatty acid phenotypes of pigs. Typically, muscle fatty acids can also be influenced by propanoate metabolism (C14:1) (Supplementary Fig. 6 E); glycine, serine and threonine metabolism (C6:0, C10:0, C12:0, and C22:6n3) (Supplementary Fig. 5 F, O and Supplementary Fig. 6 A, C); alanine, aspartate and glutamate metabolism (C22:2) (Supplementary Fig. 6 R); microbial metabolism in diverse environments (C14:0, C15:0, C16:0, C16:1, C17:0, C18:0, C18:1n9c, C20:0, C20:3n6, C21:0, C22:1n9, and C24:1) (Supplementary Fig. 5 L, J and Supplementary Fig. 6 D, F, G, H, K, L, O, Q, U); and aminoacyl-tRNA biosynthesis (C20:1) (Supplementary Fig. 5I).

3.5. Microbial predictors are associated with functional fatty acid-rich pork

A higher dietary level of UFAs, especially ALA, oleic acid, GLA, DHA, and EPA, is recommended for a healthy lifespan. Therefore, we further used the LGB model to identify microbial genera most strongly associated with these fatty acid traits. SHAP analysis showed that Dorea was positively associated with muscle C18:1n9c (oleic acid), whereas Escherichia, Terrisporobacter, and Actinobacillus were negatively associated with this trait (Fig. 7L). In addition, Bradyrhizobium and Lachnoclostridium were positively associated with C18:3n6 (GLA) and, together with other genera, were also identified as important predictors of C20:5n3 (EPA) and C22:6n3 (DHA) (Fig. 7 M, O, N). Enterococcus and Phascolarctobacterium and C18:3n3 (ALA), a precursor of DHA and EPA (Fig. 7H) were observed. Two microbial species from the genera (Bradyrhizobium and Lachnoclostridium) shared with the correlations, which were expected to provide an opportunity to promote GLA, DHA, and EPA deposition in the muscle. Although positively related to C22:6n3, Escherichia was also negatively correlated with C18:1n9c, C18:3n3, C18:3n6, and C20:5n3; thus, lower Escherichia abundance was associated with desirable meat-quality pork rich in functional fatty acids. These results suggest that specific genera may serve as candidate predictive biomarkers or potential targets for future validation in relation to the deposition of functional fatty acids.

The muscle PUFA content was negatively correlated with microbial metabolism related to fatty acid degradation (Supplementary Fig. 5 B). Here, we further analyzed the relationships between the key predictors of functional fatty acids (i.e., Lachnoclostridium, Bradyrhizobium, and Escherichia) and the fatty acid degradation pathway. The important microbes that influenced the degradation of fatty acids were ranked in the following order: Lachnoclostridium > Escherichia > Bradyrhizobium. Correlation analysis revealed a positive correlation between Escherichia and fatty acid degradation, while Lachnoclostridium and Bradyrhizobium exhibited a negative correlation (Supplementary Table 4). These results indicate that Lachnoclostridium, Bradyrhizobium, and Escherichia were closely associated with the fatty acid degradation pathway and muscle functional fatty acid traits, highlighting them as candidate microbial predictors for further mechanistic validation.

3.6. Lachnoclostridium edouardi was associated with altered muscle fatty acid composition

Because Lachnoclostridium was consistently identified as an important feature in the prediction models and was negatively associated with the fatty acid degradation pathway, Lachnoclostridium edouardi was selected for in vivo validation in a mouse model. LE administration significantly increased body weight but had no observable effect on SAT, AAT, or PEAT (Fig. 8A). Subsequent analysis of muscle fatty acids revealed a significant reduction in C16:0, accompanied by marked increases in C18:1n9c, C18:2n6c, TFA, UFA, PUFA (Fig. 8B, C).

Fig. 8.

Effects of Lachnoclostridium edouardi on the relative weight of adipose tissue and muscle fatty acids in mice (A) Body weight and relative weight of adipose tissue. (B) Muscle specific fatty acids. (C) Muscle classifying fatty acids. The symbol “a, b” indicated values within a row with different superscripts differ significantly (P < 0.05). The symbol “*” indicated differences among the groups were compared (P < 0.05).

4. Discussion

The relationships between the gut microbiota, muscle fatty acids, and meat quality have been analyzed in various pig studies (Jie Ma, et al., 2022; Yin et al., 2023). However, the diversity of individual differences and host background information, in certain cases, can affect the accuracy of investigations on the gut microbiota and obscure the actual relationship between the microbiota and host physiological states (Chen et al., 2022). Machine learning methods involve a collection of data analysis techniques aimed at establishing predictive models from multidimensional datasets to make predictions about possible outcomes (Camacho et al., 2018). However, previous pig studies have mainly focused on association analyses within single cohorts or under specific genetic and nutritional backgrounds. In addition, few studies have jointly modeled gut microbial diversity, taxonomic composition, microbial functional pathways, muscle fatty acid profiles, and meat-quality traits within one interpretable framework. In the present study, we integrated data from 291 pigs across six independent studies and compared four machine-learning models to identify microbial predictors associated with pork quality and muscle fatty acid composition. Generally, the low R² values of gut microbiota prediction models are insufficient for accurately calculating microbial associations with host metabolism, and further integration methods are required to obtain better and more robust predictions (Cammarota et al., 2020). The current predictive models achieved performances with R² values >0.8, supporting their utility for prioritizing informative microbial features in this dataset.

The quality evaluation of meat primarily involves comprehensive assessment based on indicators such as meat color and pH (Duan et al., 2022). Meat color is typically evaluated using L, a, and b values, which reflect the freshness of the meat. A higher a value or lower b and L indices are considered to indicate more vivid meat color and high quality (Chen et al., 2022; Wang et al., 2022). In this study, machine learning indicated that muscle PUFA levels were significantly positively correlated with the a value, while the b and L values were negatively correlated with the SFA/TFA ratio, suggesting that fatty acids (both PUFAs and SFAs) improve meat color. A higher pH of meat implies slower anaerobic glycolysis in muscle tissue, more stable protein structure, and better preservation performance (L. Wang et al., 2022). At 45 min, PUFAs were positively related to pH, while TFAs were negatively related to pH. Together, the manipulation of meat quality was strongly associated with muscle fatty acids by increasing PUFAs and SFAs. Indeed, compared with lean breeds, most fatty pig breeds have higher levels of PUFAs and SFAs, resulting in superior meat quality and flavor (Duan et al., 2022; L. Wang et al., 2022; Yin et al., 2023).

Machine learning has been applied to predict diseases related to the gut microbiota, such as cancer (Cammarota et al., 2020), inflammation (Barberio et al., 2022), cardiovascular diseases (Aryal et al., 2020), and diabetes (Gou et al., 2021), and to explore the potential causal relationships of microbes with these diseases. The relationship between the gut microbiome and lipid metabolism in individuals with obesity has been widely investigated using machine learning (Su et al., 2022). For example, Bacteroides caccae, Odoribacter splanchnicus and Roseburia hominis were identified as obesity-related biomarkers using a vector machine model (0.485 R²) in 2262 Chinese individuals, and obese individuals exhibited decreased diversity of these species (Liu et al., 2022). Here, we also observed a negative correlation between muscle fatty acid deposition and microbial α diversity; thus, increasing gut microbial evenness and richness should produce desirable meat-quality pork rich in PUFAs, especially GLA, DHA, and EPA. Gut microbial richness is mainly shaped by dietary composition. For example, herbivores were reported to have the highest genus-level richness, followed by omnivores and carnivores (Ley et al., 2008), and dietary intervention (energy-restricted high-protein diet) improved the status of patients with low microbial richness (Cotillard et al., 2013). A high bacterial richness was associated with marked improvement in overall adiposity, insulin resistance, dyslipidemia, and systemic inflammation than was a low bacterial richness (Cotillard et al., 2013), which might also be associated with the metabolic and anti-inflammatory benefits of PUFAs.

Dietary oleic acid (C18:1n9c), a key cis-MUFA in colostrum and camel milk (Mehra et al., 2021), is reported to extend lifespan by altering intestinal lipid droplets and improving lipid oxidation (Papsdorf et al., 2023a, Papsdorf et al., 2023b). In the pig industry, C18:1n9c in the muscle is the main MUFA in pig breeds with superior meat quality (Duan et al., 2022). The data mining also indicated a positive correlation between pH-based meat quality and muscle C18:1n9c content, which was mainly predicted by increases in Dorea and decreases in Escherichia, Terrisporobacter, and Actinobacillus in the gut. Indeed, a greater abundance of gut Dorea accompanied by increased serum C18:1n9c levels had an anti-inflammatory effect in an animal model of ulcerative colitis (Fernández et al., 2020). An obese-linked fatty acid profile was characterized by a decrease in functional fatty acids (Fernández-Navarro et al., 2019), and increasing the intake of PUFAs, especially GLA, DHA, and EPA, is also recommended because of the decreased incidence of cardiovascular and metabolic diseases (Gould et al., 2022; Paredes et al., 2023). Thus, foods rich in these functional fatty acids, such as meat, nuts, and edible oils, have attracted much attention due to their metabolic benefits. Here, microbial strategies for producing GLA-, DHA-, and EPA-enriched meat were shown to increase the abundances of gut Bradyrhizobium and Lachnoclostridium or reduce the abundance of Escherichia. Bradyrhizobium is commonly associated with soil and plant environments. It may have been introduced into the gut through fiber-rich plant-based feed and subsequently contributed to intestinal colonization, thereby influencing fatty acid metabolism (Ding et al., 2018). This may partly explain the high functional fatty acid deposition in Chinese local fatty pigs with high ability to digest crude fibers. Lachnoclostridium species respond to dietary n-3 PUFAs (Tabata et al., 2021) and are widely involved in host lipid biosynthesis (Nogal et al., 2021), with potential implications for ameliorating obesity and type 2 diabetes (Li et al., 2023). Thus, increasing Bradyrhizobium and Lachnoclostridium abundances by dietary intervention (i.e., fiber, fatty acids, or probiotics) may facilitate the deposition of fatty acids in the muscle. Pigs and mice share similarities in microbiota–host metabolic interactions, and the composition of the gut microbiota, rather than the host species itself, is considered a key factor regulating lipid metabolism (Ridaura et al., 2013; Wang et al., 2026) In addition, the microbial genera associated with fatty acid degradation were ranked as follows: Lachnoclostridium > Bradyrhizobium. Therefore, these genera should be interpreted as model-prioritized microbial predictors. Based on this ranking, Lachnoclostridium edouardi was selected for in vivo validation in mice. The results supported the biological relevance of Lachnoclostridium, as its administration was associated with increased muscle UFA and PUFA levels.

Another important finding from the present study is that the abundance of Escherichia, which is correlated with fatty acid degradation, serves as a key negative predictor of the formation of desirable meat-quality pork that is rich in functional fatty acids. In general, studies of Escherichia species have focused mainly on the pathogenic effects of these bacteria in diarrhea and gastrointestinal diseases (Hansen et al., 2021), and the role of the current predictive model in muscle fatty acid deposition, especially for ALA, GLA, DHA, and EPA, has not previously been reported. Similarly, compared with fatty acid-producing Jinhua pigs, lean Landrace pigs (23–28%) had more Escherichia species (< 5%), which might explain the greater fatty acid composition in the muscle of fatty pigs (Xiao et al., 2018). Indeed, Escherichia infection is characterized by intestinal fatty acid malabsorption and lipid metabolic disorders (H. Liu et al., 2019). Conversely, antibacterial effects were also confirmed for various functional fatty acids, and dietary supplementation with DHA and EPA inhibited the growth of Escherichia (Zhuang et al., 2021). Accordingly, the Escherichia abundance in the gut should be considered cautiously not only due to its potential to cause diarrhea but also because of its negative effect on meat quality, especially in terms of functional fatty acid deposition.

5. Conclusions

In summary, this study established an interpretable machine-learning framework linking gut microbiota with muscle fatty acid composition and pork quality across multiple pig cohorts. By integrating taxonomic and functional microbial features, we identified several key candidate genera, including Lachnoclostridium, Bradyrhizobium, and Escherichia, that were closely associated with the deposition of functional fatty acids in muscle. Among them, experimental validation with Lachnoclostridium edouardi supported the model prediction that modulation of the gut microbiota could increase muscle UFA and PUFA levels. These findings provide a data-driven basis for developing microbiota-targeted strategies to improve the nutritional quality of pork. Future studies should further validate these findings in pig models and expand sample size for machine-learning analysis to obtain more robust conclusions and improve predictive accuracy.

CRediT authorship contribution statement

Lingfeng Pan: Writing – original draft, Methodology. Zejin Li: Writing – original draft, Methodology. Qian Zhu: Investigation. Chenyu Wang: Investigation. Qing Ouyang: Investigation. Xingguo Huang: Visualization. Caiyuan Zhou: Investigation. Ifen Hung: Investigation. Chunxue Liu: Investigation. Kang Xu: Visualization. Jie Yin: Writing – review & editing, Supervision, Project administration, Funding acquisition, Conceptualization. Yuying Li: Writing – review & editing, Supervision, Project administration, Methodology, Funding acquisition, Conceptualization. Yulong Yin: Supervision.

Funding

This work was supported by the National Natural Science Foundation of China (32172761 and U20A2055), Hunan Provincial Scientific and Technological Innovation Team (2021RC4060), “Huxiang Young Talents Plan” Project of Hunan Province (2022RC1157), Key Research and Development Program of Hunan Province (2023NK2018) and Agricultural Science and Technology Innovation Project Special Fund of Chinese Academy of Agricultural Sciences (ASTIP-IBFC and CAAS-IBFC-2025-01).

Declaration of competing interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Appendix A. Supplementary data

Supplementary material 1

Supplementary material 2

Data availability

Data will be made available on request.