Development of a machine learning potential for nucleotides in water

Recent experimental research showed that nucleotides, under favorable conditions of temperature and concentration, can self-assemble into liquid crystals. The mechanism involves the stacking of nucleotides into columnar aggregates. It has been proposed that this ordered structure can favor the polymerization of long nucleotide chains, which is a fundamental step toward the so called “RNA world”. In this thesis, starting from ab initio molecular dynamics simulations, at the density func- tional theory level, an all-atom potential for nucleotides in water, based on an implicit neural network representation, has been developed. Its stability and accuracy have been tested and its predictions on simple model systems have been compared with data generated both ab initio and using currently available empirical force field for nucleic acids.