ACS Medicinal Chemistry Letters
Letter
synthesized a number of compounds based on the scaffold for
the primary study of the structure−activity relationship. We
found that only compounds with small substituents, such as a
hydrogen atom or a methyl group at position 2 of the scaffold,
exhibit favorable in vitro activity, while compounds with bulky
groups at the same position were less active. Moreover, the
volume and length of the substituents at position 4 also
influence the antifungal activity. So, compound 94 with 2-
methyl group and 4-cyclopropyl-NH substituent appears to
have the most promising MIC80 of 0.6 μM for C. neoformans
and MIC 0.39−0.78 μM for C. gattii. The MIC80 was 0.79−
1.56 μM against a fluconazole (FLC) resistant strain with a
CC50 of 9.6 μM in human hepatoma cells resulting in a
specificity index, SI = 16, for C. neoformans. Compound 94 was
also tested for synergy in a checkerboard assay with AmB and
FLC, where the FICIs were calculated to be 1.25 and 1.5,
respectively (data not shown). These data suggest an
indifference but, importantly, no antagonism between com-
pound 94 and either AmB or FLC.
Machine Learning. Assay Central was used to generate
machine learning models that used a Bayesian algorithm and
ECFP6 fingerprints alone, for the 121 compounds with C.
neoformans data. The threshold for actives was MIC80 12.5−
25.0 μM. The 5-fold cross validation ROC was 0.70, and other
statistics (Figure 3A, Supporting Information (SI), Table S1)
were acceptable for such a relatively small data set. This data
was then used to score a set of 51 additional analogues
(external set test ROC 0.77, Figure 3B). From past experience,
as we add more data to the machine learning model the cross-
validation statistics generally improve (172 compounds, ROC
0.90, Figure 3C) and generate better predictive models for
optimization of the molecules. We are thus able to lower the
threshold for an active to MIC80 < 3.12 μM in order to predict
more potent compounds using this iterative approach.
We have also generated a Bayesian model with over 3000
compounds with data from the NIAID ChemDB HIV,
Opportunistic Infection and Tuberculosis Therapeutics Data-
base, with excellent statistics overall at a threshold of 10 μM
(ROC 0.87, Figure 3D), which is potentially useful for
exploring more chemical diversity because of the broader
makeup of the training set. We used these data from the 172
compounds tested as an external test set for this training
model, which yielded reasonable statistics (ROC 0.81; Figure
4).
The molecules of most interest highlighted in this study
(Table 1) were used with this literature C. neoformans model
with a cutoff of 10 μM (ROC 0.62, SI, Figure S1A) or 3.2 μM
(ROC 0.82, SI, Figure S1B). The predictions and model
domains (a measure of applicability) can be seen in more detail
in SI, Table S2. As we also generated in vitro data for C. gattii
in this study, we used this set from Table 1 to construct a
preliminary model with a cutoff at MIC 1.52−3.12 μM (ROC
0.62, SI, Figure S2). These machine learning models will be
put to further use for virtual screening of compound libraries of
commercial molecules in order to find additional molecules for
testing against these fungi. These Bayesian models can also be
used to help us further optimize the current chemical series
alongside models for ADME/Tox properties.
Figure 4. Statistics for the Cryptococcus data from this study when
used as an external test set using the literature cryptococcus model
derived from the NIAID ChemDB database as a training set (10 μM
activity threshold).
virus.35 Most recently, we have used this software to assist in
drug discovery for chordoma,36 Neisseria gonorrheae,37 HIV,38
and Staphylococcus aureus.39 This study represents our first
attempt at applying this approach to antifungal drug discovery.
We have now identified a series of 5-nitro-6-thiocyanatopyr-
imidine antifungal drug candidates using in vitro and
computational machine learning approaches that have been
shown to inhibit C. neoformans and C. gattii growth at
submicromolar levels (Figure 1). These may represent a
starting point for further hit-to-lead optimization and target
identification.
Synthesis of 5-Nitro-6-thiocyanatopyrimidines. 2,4-
Disubstituted 5-nitro-6-thiocyanatopyrimidines studied here
were synthesized in four steps according to Figure 2. The
starting 2-substituted 4,6-dihydroxypyrimidines were nitrated
at the 5-position using fuming nitric acid and sulfuric acid in a
catalytic amount. Further treatment of these intermediates
with phosphorus oxychloride resulted in the formation of 2-
substituted 4,6-dichloro-5-nitropyrimidines 7−12. At the next
stage, 4,6-dichloropyrimidines were reacted with the corre-
sponding amines in the acetate form in dioxane (Figure 2c) or
with the corresponding sodium alkoxides in alcohol (Figure
2d), which led to the replacement of only one chlorine atom in
the 4-position to amines or alcohols. The final 5-nitro-6-
thiocyanatopyrimidines 63−112 (Table 1) were obtained by
nucleophilic substitution of the second chlorine atom with
potassium thiocyanate in good yields.
In Vitro Assays and Preliminary Structure−Activity
Relationships Study. Using a whole-cell phenotypic screen
approach, we tested a set of 121 chemically diverse compounds
from our laboratory library and measured the MIC80 of these
compounds against the pathogenic C. neoformans laboratory
strain KN99.28 We identified that 17 out of 121 hits show
activity with the MIC80 values less than or equal to 50 μM.
Among them, the 5-nitro-6-thiocyanatopyrimidines represent
the most interesting series for further research. Next, we
During recent patent preparation, we became aware of a
Russian patent describing 2-nitroheterylthiocyanates with
activity against Candida, Aspergillus, and Fusarium strains
only.41 Interestingly, they did not test against Cryptococcus
778
ACS Med. Chem. Lett. 2021, 12, 774−781
Donlin, Maureen J.
