We present a new approach at the atomic level for the development of knowledge-based mean force potentials (MFPs) that can be used in fold recognition, ab initio structure prediction, comparative modelling and molecular recognition. Our method is based on atom-type definitions, raising the total frequency of the pairwise distributions and leading to very accurate and specific distance-dependent energy functions. Forty different heavy atom types were defined depending on their bond connectivity, chemical nature and location level (side-chain or backbone). Using this approach it has been possible to obtain average frequencies of pairwise contacts about 15 times higher than the ones obtained using the classic way of one heavy atom definition for each amino acid (i.e. alpha-carbon, beta-carbon, virtual centroid or virtual beta-carbon co-ordinates). In this paper we use this approach to develop a MFP that can be used in fold recognition and we compare it with a classic MFP at the amino acid level compiled from the alpha-carbon distances between the different amino acid pairs. Both potentials involve all the pairwise contacts extracted from a non-redundant folds database of 180 protein chains with a sequence identity threshold of 25%. The pairwise energy functions of the MFP at the atomic level have a deep and very well defined minimum for each pairwise interaction, in contrast to the same curves obtained from the MFP developed at the amino acid level, which generally have multiple minima with similar depth. Our results also show that this MFP is able to produce very similar energy profiles for couples of proteins that share a very low sequence identity but are closely related at the structural level. When these profiles are plotted considering the structure-structure alignment, they are mostly superimposed, showing a correlation with the structure-structure similarity. In the same test, the MFP at the amino acid level fails to produce similar profiles. We suggest that using this MFP at the atomic level in the last stages of fold recognition or threading, when some candidates are available, can improve the sequence-structure alignments and, therefore, the final models. We also discuss the possibility of using this approach in the development of new MFPs to be used in ab initio structure prediction, comparative modelling and molecular recognition procedures.