The ANTARES neutrino telescope stopped gathering data in February 2022, after nearly 16 years of operation. The detector consisted of 12 vertical lines forming a 3D array of photo-sensors, which instrumented about 10 megatons of Mediterranean seawater. We present a method using deep learning that improves the direction reconstruction of low-energy single-line events, for which the reconstruction of the azimuth angle of the incoming neutrino is particularly difficult. We also present a combination of machine learning techniques to reconstruct the energy of the same kind of events. Our results enhance the resolution of former reconstruction techniques, at least doubling our sensitivity in the range of energy of tens of GeV, which is highly relevant for dark matter searches and other physics studies. Here, we propose a binned Dark Matter (DM) search towards the Sun for ANTARES single-line events using the new reconstruction methods. We compute the neutrino flux sensitivity for different DM annihilation channels and particle candidate masses. In this first trial, the methodology is applied to a subset of ANTARES data: these results anticipate better sensitivities for low masses of DM candidates (below $\sim150$ GeV) and/or soft spectrum channels compared to those obtained based on standard reconstruction techniques once the method is applied to the whole ANTARES dataset.