Accurate sub-seasonal to seasonal prediction of regional precipitation remains challenging using dynamical seasonal prediction models. Machine learning models have demonstrated enhanced skills in predicting modes of climate variability, like ENSO, over longer lead times. While these modes lend regional predictability on sub-seasonal to seasonal scales, the advances in the predictability of these modes does not appear to have benefited regional predictions similarly. A recent study applied a multi-input multi-output autoencoder-decoder (MIMO-AE) network to capture the non-linear relationship between tropical Pacific SST variability and Southern California [Passarella and Mahajan, 2023]. The study showed that the MIMO-AE network offered enhanced predictability when a time-series prediction model was used in combination with the MIMO-AE network. Here, we extend that work to apply a similar approach to assess the predictability of Amazon rainfall offered by Tropical Pacific variability. Our preliminary analysis with a MIMO-AE network trained only on monthly data from a single historical run with an Earth System Model reveals that similar to the previous work over Southern California, the MIMO-AE network significantly enhances the predictability of Amazon rainfall as compared to that from ENSO indices like Nino3.4 and the ENSO Longitudinal Index, up to a lead time of 3 months. When Tropical Atlantic data is also included in the network, the predictive skill of the network is further enhanced over the Amazon as well as regions of the US. We will present results on the predictive skills of a MIMO-AE network trained on ensembles of historical simulations as well as observational data.