Computational prediction of a 3D structure of a molecular complex, also known as docking, is essential in modern biological research. It can complement MD, provide working directions to experimentalists, etc. We are invested in fragment-based docking, specifically for the single-stranded RNA-protein complexes. Why ssRNA specifically? Generally speaking, molecular flexibility is a scourge of docking: it increases its complexity and decreases results' reliability. High flexibility leads to a near-infinite number of docking models, the processing of which is too expensive computationally. Hence, highly flexible ssRNA is a challenging target to work on. A fragment-based docking approach was developed to tackle this high flexibility issue [1]. Its core idea is to split the ligand into overlapping fragments, and dock them onto the rigid receptor separately, assembling the fragments back into the whole ligand afterwards. For the full procedure to succeed, each fragment must return at least one correct pose (so-called near-native): this is the sampling problem. The poses are obtained by minimisation using a differentiable energy function. Then, before assembling, docked fragments must be filtered, keeping a high percentage of near-natives. Otherwise, the assembly task once again becomes too expensive computationally: this is the scoring problem. The filtration is done using a scoring function. We are working with the ATTRACT docking engine, where the same function is used both for sampling and scoring. It has the shape of a Lennard-Jones potential, and 2 parameters per atom type pair (1054 in total). The current parameters were obtained in 2010 by extraction of the statistical potentials from RNA-protein crystal structures and were optimized by a random Monte Carlo-like strategy [2].