PMC

We illustrate how SMARTY scores working atom types (red on the left) compared to reference atom types (blue on the right) using a bipartite graph. (a) When there is only one generic SMARTS string (“[#6]”), this pattern is assigned to both CT and CA atom types through the PATTY-adapted scheme, but it will be matched only to the more populated reference atom type (CT highlighted in yellow); (b) when SMARTY creates a new set of working types, containing both “[#6]” and “[#6X4]”, the generic atom type will not match CT anymore, since it chemically matches the new working atom type (“[#6X4]”) instead.

The graph shows the total score (fraction of atoms matching the parm99/parm@Frosst reference atom types) as a function of iteration varies with the temperature for SMARTY_orig sampling AlkEthOH. The lines in the graph represent the temperatures 0 (pink), 0.0001 (orange), 0.01 (green) and 0.1 (gray). We can see a high variation throughout the simulation and a predominance of lower scores for extreme temperatures such as 0 and 1.0. At intermediate temperatures, such as 0.01 and 0.0001, SMARTY scored higher. The Supporting Information includes plots for simulations at all temperatures.

The numbers attached to certain arrows correspond to the probability of following that arrow; all moves are made with equal probability in SMARTY. The diamonds represent decisions and the rectangles are processes. We start the SMARTY move proposals with the working atom types and decide randomly with equal probability if we want to remove or add an atom type to the set. If we decide to remove, we randomly choose a working type to be removed and re-score the new set. If we decide to add, we pick a working atom type from the set and check if it has an alpha substituent. If the answer is no, we either add an alpha substituent or a decorator. If the answer is yes, we first check if the working atom type describes a hydrogen, and if yes, we add a beta substituent, if not, we either add an alpha or a beta substituent. The new atom type SMARTS created is added to the end of the list of its corresponding parent. The end result (black circle on the bottom) is a move proposal which is then evaluated via scoring function ().

Workflows for the SMARTY (left) and SMIRKY (right) tools are shown. There are three input data categories for SMARTY and SMIRKY (shown at top). In SMARTY (on the left), there are input files (such as base types, initial atom types, and decorators), molecules (input via a molecule set file), and reference data consisting of typed molecules (parm99/parm@Frosst atom typing was used in this work). SMIRKY has similar inputs, shown in the purple area on the right, and additionally allows the user to set the decorator odds and its reference data is fragment types (here, from smirnoff99Frosst). The algorithms are represented in the green area and are available on GitHub (https://github.com/openforcefield/smarty). Both tools begin by reading and processing the input files (top part of the green area). SMARTY (on the left) then conducts a series of moves (coral area) consisting of choosing one working atom type per iteration (icon with connected atoms), and deleting or modifying (pencil icon) this atom type, making choices made with equal probabilities (). SMIRKY (on the right) is similar, but samples over working fragment types (such as nonbonded types, bonds, angles, and proper and improper torsions; in the figure these are represented by two different icons with connected atoms) and uses weighted probabilities on their choices (). The acceptance criteria (diamond icon) for both tools are the same, using (). Both tools run a user-specified number of iterations and iterate over the steps in the coral area (repeat icon), then write out a 1nal results file after all iterations are completed.