*3.6. Polymorph Assessment*

The objective of this analysis revolved around stability assessment of the hydrogen bond network present in both forms, given the complexity of the chemical environment of ganciclovir. The reliability and predictive ability of the fitted logistic regression model were mirrored by the value of the area under the ROC curve, which was equal to 0.878, and by the reduction from the null to the residual deviance [44,45]. The same procedure was performed for each form individually, and since they are polymorphs of the anhydrous form of ganciclovir, very similar coefficients were obtained (Please refer to Supplementary Materials). Minor differences can be attributed to the non-deterministic nature of the process involving the fitting of the model. This outcome confirms the robustness of this method, which is capable of assessing the stability of various polymorphs having different hydrogen bonds simultaneously.

The final model was used to calculate the propensities of all possible intermolecular hydrogen bonds, with those in form II having the highest probabilities (more conventional contacts) (Please refer to Supplementary Materials). The overall low likelihood of the intermolecular bonds in form I was illustrated in the putative structure landscape, which categorized this polymorph as having the least stable hypothetical forms (Please refer to Supplementary Materials). The utilization of every functional group in form I, for intermolecular bonding, might have been prioritized over the formation of the fewer and more probable contacts.

It is understandable that due to low frequencies, the contacts in form I were ranked as having lower probability, and hence low stability was predicted. However, this outcome was not in agreemen<sup>t</sup> with the fact that form I is the thermodynamically stable polymorph, under ambient conditions. At this point, it is essential to recall the statistical mechanism behind the construction of the Hydrogen Bond Propensity (HBP) model, which is based on the occurrence of hydrogen bonds in similar chemical environments. Therefore, in some examples, the resultant model might not be sufficient to explain the complexity of hydrogen bonding and to capture the collective effects of multiple factors that determine the polymorph stability [6]. In his research paper, Abramov commented how in general, a HBP model cannot account for enantiotropic relationships between polymorphs, such as the one between form I and II [6]. Moreover, one has to take into account that the directional features and geometric parameters of the contacts are not being considered in the model, and these characteristics have a significant effect on stability.

In an attempt to overcome these limitations, the HBP model was constructed using a much larger training set, but similar results were obtained. Despite such limitations, there was still valuable information that could be extracted from these results. The putative structure landscape in Figure 11 portrays the presence of data points located very close to form II, representing reasonable hypothetical structures having very strong hydrogen bond interactions. The viability of the formation of such forms is highly encouraging in view of further research dedicated to the exploration of other possible ganciclovir polymorphs.

**Figure 11.** The propensity participation chart output showing form II (violet dot), which is found at the location with the optimal conditions (large values for both axes) that are usually associated with thermodynamically stable crystal forms.
