The total organic carbon (TOC) content of shale formations is one of the important parameters to characterize the organic richness of unconventional shale reservoirs and identify shale production sweet spots. We have used traditional approaches like the Schmoker density log method and shallow neural networks to estimate TOC from seismic data and well logs. The Schmoker method is easy to use but does not consider any rock constituent changes affecting the bulk density. However, many neural network case studies have blindly fed all of the attributes to networks without understanding physical or statistical relationships between TOC and seismic properties. Therefore, we develop a combined exploratory data analysis for rock and seismic properties and deep learning approach to prevent the overfitting problems and obtain a more reliable TOC estimation model. First, we conduct statistical analysis of seismic properties from well logs and define the influential properties highly correlated with the TOC logs, such as bulk density, P-wave velocity, and Poisson’s ratio. Second, we derive a multivariate linear regression model with the three properties and analyze the statistical contribution of each property based on sensitivity tests. Third, we apply a deep neural network to find the nonlinear prediction model with a better fit to the TOC data than traditional approaches. We test our methods on real core and well log data from the Wolfcamp shale formation in the Permian Basin, Texas. In conclusion, our method can provide a more accurate and explainable TOC estimation from seismic properties for improved shale characterization.