Abstract: Machine learning algorithms are widely used to interpret well logging data. To enhance the algorithms’ robustness, shuffling the well logging data is an unavoidable feature engineering before training models. However, latent information stored between different well logging types and depth is destroyed during the shuffle. To investigate the influence of latent information, this study implements graph convolution networks (GCNs), long-short temporal memory models, recurrent neural networks, temporal convolution networks, and two artificial neural networks to predict the microbial lithology in the fourth member of the Dengying Formation, Moxi gas field, central Sichuan Basin. Results indicate that the GCN model outperforms other models. The accuracy, F1-score, and area under curve of the GCN model are 0.90, 0.90, and 0.95, respectively. Experimental results indicate that the time-series data facilitates lithology prediction and helps determine lithological fluctuations in the vertical direction. All types of logs from the spectral in the GCN model and also facilitates lithology identification. Only on condition combined with latent information, the GCN model reaches excellent microbialite classification resolution at the centimeter scale. Ultimately, the two actual cases show tricks for using GCN models to predict potential microbialite in other formations and areas, proving that the GCN model can be adopted in the industry.