In recent years, different machine learning methods have been proposed for total energy prediction and molecular design. Having a robust architecture and a good descriptor is crucial in neural networks as they can significantly impact the final results. Here we investigate the use of multilayer perceptron neural networks (ReMLP-NET) and transformer encoders (Transformer-Potential) in molecular energy prediction. The methods are trained on a dataset called Retrievium, which mainly contains organic structures with elements H, C, N, O, S, and Cl, originating from the GDB13 dataset. The atomic environment vector is generated based on the symmetry functions used in ANI-2x, where it spans both configurational and conformational space, with the addition of genetic algorithm optimized parameters in the case of ReMLP-NET. Unlike ANI-2x, which includes a separate network for each element, both ReMLP-NET and Transformer-Potential use a single network. Different preprocessing steps are performed on the employed dataset to accelerate the learning process and improve the results. The ReMLP-NET and Transformer-Potential methods are compared with the well-known ANI-2x in terms of total energy, where they obtained results of 1.29 and 1.33 kcal/mol for MAE on average, respectively, compared to 1.53 kcal/mol for ANI-2x.