In a previous virtual screening campaign searching for DNMT1 inhibitors, eleven molecules that activate the catalytic activity were found and reported in a thesis dissertation [7] . Such results were not disclosed before in a peer-reviewed publication because they were initially regarded as negative data of a virtual screening effort. However, there is an increasing awareness of the significance of disclosing initially non-interesting compounds, aka negative data [7] .

We implemented a ligand-based approach to identify additional DNMT1 activators. We did not conduct a structure-based approach because the mechanism of enzymatic activation of the small molecules at the structural level is not known. The ligand-based approach was a fingerprint-based similarity search of ChemDiv, a commercial screening library taking as a reference eight known DNMT1 activators, namely, RG108-GGGKQVY, RG108-KQVY, L112-0668, L112-0728, L485-2669, L483-0323, L485-2761, and M056-0726 (Figure 1) most of them previously uncovered in a virtual screening intended to find inhibitors (vide supra) [7] . At the time of the study, ChemDiv contained 220,000 compounds (https://www.chemdiv.com/; accessed March, 2023). The similarity search was done with two fingerprints of different designs: dictionary (MACCS Keys, 166-bits) and topological (Extended Connectivity Fingerprint, Radius 6, ECFP6) [8] . Similarity searching hits were selected from ChemDiv if they had a Tanimoto similarity [9] value equal to or above 0.80 (with MACCS keys) or equal to or above 0.50 (with ECFP6). For ECFP6, we considered a similarity value lower than MACCS keys because of the higher resolution of the former [10] ). Figure S2 in the Supplementary Material illustrates the general approach to identifying the DNMT1 enzyme activators reported in this study.

Selected computational hits were purchased from ChemDiv and experimentally tested at the Reaction Biology Company in an enzymatic activity methyltransferase assay using the HotSpotSM platform [11] . Our group has reported the methodology and results of this biochemical assay, including the identification of 7-amino alkoxy-quinazolines [3] . Briefly, HotSpotSM is a low-volume radioisotope-based assay that employs tritium-labeled AdoMet -3H-SAM- as a methyl donor. The test compounds diluted in dimethyl sulfoxide were added using acoustic technology (Echo550, Labcyte, San Jose, CA, USA) into an enzyme/substrate mixture in the nano-liter range. The reactions were started by adding tritium-labeled AdoMet and incubated at 30˚C. Total final methylations on the substrate (Poly dI-dC) were identified by a filter binding method implemented in Reaction Biology. Data analysis was conducted using Graphed Prism software (La Jolla, CA, USA) for curve fit. The enzymatic inhibition assays were carried out with the DNMT1 enzyme at a concentration of 25 nM, 1 μM of S-adenosyl methionine, SAM. The standard positive control was S-adenosylhomocysteine, SAH. The compounds were tested in 10-concentration IC50 (effective concentration to inhibit enzymatic activity by 50%) with a threefold serial dilution starting at 100 μM.

Five activators of the enzymatic activity of DNMT1 were used to build a pharmacophore hypothesis: L112-0668, L112-0728, L485-2669, L112-0731, and RG108-KQVY (Figure 1). These compounds were classified as high activators for their pronounced activation potential (higher than 150%) for the initial four compounds. Furthermore, RG108-KQVY was ascribed the classification of “highly activating.” Three nucleotide molecules were used to perform the pharmacophore model of the inhibitory compounds: decitabine, zebularine, and azacitidine. Another pharmacophore model was constructed using the experimentally confirmed hit molecules and queries that had greater than 400%: C066-2421, C350-0522, L112-0668, L112-0725, L112-0728, L112-0731, L485-2669, and L485-2777. The pharmacophore models were constructed using the Molecular Operating Environment (MOE) 2022.02 program with the pharmacophore elucidator and consensus function. The default settings were employed.

Table 1 summarizes the results of the similarity-based virtual screening of ChemDiv, including the percentage of enzymatic activity of DNMT1. The table also summarizes the results of the enzymatic activity of the compounds previously identified as activators during a docking-based virtual screening to identify inhibitors (vide supra) [7] . Table S1 in the Supplementary Material shows the results of the similarity values for each of the six hit compounds from the similarity searching.

Results of the similarity searching indicated that six out of the nine molecules selected from ChemDiv activated the enzymatic activity of DNMT1 in a biochemical assay. The molecules that turned out to be inactive, despite having a high Tanimoto coefficient value based on MACCS keys fingerprints (0.87 - 0.92), had low Tanimoto coefficient values based on ECFP6 fingerprint (0.16 - 0.24). This suggests that the ECFP6 fingerprint has a greater impact on assessing molecular similarity for virtual screening applications. Indeed, in previous benchmark studies, it has been noticed the efficiency of circular fingerprints such as ECFP6 [12] . This could be because ECFP6 fingerprints consider not only the presence or absence of chemical groups, as MACCS keys [8] , but also the connections between them, providing a more certain measure of similarity between molecules [13] . It is also why the similarity values computed with ECFP are generally lower than those computed with MACCS keys (166 bits), as they evaluate the entire molecule, including the connectivity of its chemical groups [13] .

The difference in activation percentage between the activators ranged from 191% to 672%. This depended on the molecule’s activation percentage, to which they had the highest similarity. It is important to mention that the molecules that had more activation percentage were those that had similarity with the molecules published by Martínez-Fernández [7] , the molecules that were considered as references for this virtual screening were the ones that are mentioned in Section 2.3. The molecules were classified in three categories based on their activation percentage: “activators” were those with over 100% activation, “high activators” greater than 150%, and “strong activators” with an activation

^aThe averages of the DNMT1 enzyme activation values are shown in units of percentage and the standard deviation of their measurements. SAH was used as a positive control. ^bCompounds used to construct the pharmacophoric hypothesis. ^cCompounds considered as queries in the similarity searching. ^dMolecules from ChemDiv identified from similarity searching of reference compounds using the Tanimoto coefficient and ECFP6 and MACCS keys (166-bits) fingerprints. The results of the similarity searching are summarized in Table S1. ^eMolecules used for the pharmacophore hypothesis with the high activators molecules.

percentage over 400%. Table S2 in the Supplementary Material summarizes the percentage of DNMT1 of selected hit compounds at different concentrations. Results indicated that, for some compounds such as L112-0729, L112-0716, L112-0706, and L112-0698, the activity was higher at lower compound concentrations. The overall increase in enzymatic activity at lower concentrations followed by a decrease in activity at even lower concentrations seem to be associated with the so-called biphasic response, wherein an automatic reduction or decrease in response occurs at high concentrations. The biphasic response results in an optimal range of stimulation for downstream effects [14] . Biphasic responses in signaling are frequently sought after, as they can act as a safeguard against excessive activation, overexpression, and the dominance of mutants [14] . In any case, it is remarkable that L485-2777 increases the activity of DNMT1 up to 627% at 11 μM. The activity of L112-0729 is also notable: 275% of activation at 11 μM. L112-0716, L112-0706, and L112-0698 increased the activity of DNMT1 between 261% - 393% at 5.6 μM (Table S2).

Figure 1(b) and Figure 1(c) shows the chemical structures of the activators, highlighting in blue the core (Bemis-Murcko) scaffolds [15] . The Figure indicates that a scaffold is common to all three categories of molecules according to the classification based on activity (vide supra). For example, the molecules with higher activity L485-2777 (672%) and the molecule L485-2761 (151.49%) share the same scaffold; however, the only difference between them lies in their substituent groups. L485-2777 has a methoxy group and a chloride, whereas L485-2761 has a methyl group and a chloride. Similarly, the molecules L112-0725, L112-0668, L112-0728, L112-0731, L112-0706, L112-0698, L112-0716, L112-0688 with activity percentages ranging from 121% to 525% share the same scaffold but have different substituents, suggesting that it is the substituents that make the difference in activation capacity due to their electronegativity, and their donor or acceptor capacity, among other factors [16] [17] . In contrast, the inhibitory compounds (Figure S1) and activators do not share scaffolds.

Figure 2 shows the preliminary pharmacophoric points of DNMT1 modulators, including activators (Figure 2(a)), strong activators (Figure 2(b)), and nucleoside inhibitors (Figure 2(c)). The figure shows reference activator compounds L112-0668 (Figure 2(a)) and L485-2777 (Figure 2(b)), and the DNMT1 inhibitor azacitidine. The figure highlights the principal difference between the pharmacophore models of the activators and nucleoside inhibitors: the activators have more aromatic/hydrophobic groups, one donor group but in different positions. Something that the inhibitors and activators have in common is that both of them have two acceptor groups in the ring at the same position. Furthermore, the highly activating compounds share only one hydrophobic group with a hydrogen acceptor group with the inhibitors, while they share four hydrophobic/aromatic groups and only one hydrogen acceptor group with the activators. One hypothesis of the differences and similarities between these two models is based on the binding site. According to Figure 1 and Figure 2, and Figure S1 in the Supplementary Material, most of the activating compounds are larger than the nucleoside inhibitors, and if they have different binding sites, they have different groups because they need other interactions with amino acids found in that cavity. This hypothesis will be explored with a structure-based pharmacophore model [18] [19] . Of note, a structure-based pharmacophore model of activators of DNMT1 is yet to be explored since the mechanism of activation is not yet understood at the molecular level.

A chemical multiverse is a group of multiple chemical spaces, each one defined by a given set of descriptors [20] . Therefore, multiple chemical representations provide a more comprehensive view of the coverage of the chemical space of compound sets. The chemical multiverse has been used to analyze the relationship and diversity of compound datasets [21] . The chemical multiverse of DNMT1 modulators is shown in Figure 3, described by ECFP6, MACCS Keys

(166 bits), and 2D pharmacophoric fingerprints (Fp2D). The visualization method was tSNE. The inhibitors used for the multiverse are nucleoside and non-nucleoside, shown in Figure S1 in the Supplementary Material [22] . The visual representation of the chemical multiverse indicated that activators and inhibitors share common structural features among themselves. The characteristics in common are atom groups, the connectivity between the atom groups, and pharmacophoric features. This is why, within the chemical space described by various fingerprints, activating molecules arrange into distinct clusters within the groups formed by DNMT1 inhibitors.

Compounds with the ability to increase the enzymatic activity of DNMT1 have been disclosed on a limited basis. Since in most previous virtual screenings and optimization campaigns, the inhibitors are pursued, the activators might be considered as negative results or non-interesting results. Recently, another research group disclosed two compounds (hybrids, small molecules-peptides) that, in a serendipitous finding, could activate the activity of DNMT1. Herein, we discuss a total of seventeen small molecules with various levels of activation of enzymatic activity. The most potent activator of the enzymatic activity of DNMT1 was L485-2777 (623% enzyme activation at 11 μM of compound concentration). All compounds were tested in a biochemical assay for DNMT1 activity under the same conditions. Analysis of the chemical multiverse indicated that the inactive, activators and inhibitors occupy distinct positions in the chemical space sharing, within each group, common atomic groups, connectivity, and pharmacophoric characteristics, as captured by the different fingerprints. A preliminary pharmacophore hypothesis indicated by the compounds employed in the model’s construction, it is notable that both inhibitory and activator compounds share three pharmacophoric features: one aromatic group and two hydrogen-bond acceptors. Conversely, the pharmacophore model of the highly activating molecules shares four hydrophobic points and one hydrogen-bond acceptor group with the model of the activating molecules. When compared to the model of the inhibiting molecules, it only shares one aromatic group with one hydrogen-bond acceptor. Activator compounds contain a higher quantity of hydrophobic groups in their structure, while nucleosidic inhibitors contain a higher number of hydrogen bond acceptor and donor groups. These differences are partly due to the fact that activators and inhibitors do not share common scaffolds. Given this observation, we posit the hypothesis that activating molecules typically possess a larger molecular size when contrasted with nucleoside inhibitors. However, among the different groups of activators evaluated (activators > 100%, super activators > 150%, and strong activators > 400%), they share common core structures. The primary distinctions among them lie in their functional groups. Additionally, they share pharmacophoric features, exhibiting fewer pharmacophoric points as the molecules become more activating. Consequently, the binding sites for these two categories of compounds are probably dissimilar. The mechanism of enzymatic activation of the small molecules at the structural level is not known. For that purpose, it is necessary to conduct a structure-based approach.

Perspective of this study is performing more replicates of the enzymatic assays to obtain a more reliable value of the percentage activation and the evaluation of the newly identified small molecules as activators of the enzymatic activity of other DNMTs (e.g., DNMT3A, 3B). Other perspective is to conduct cell-based assays. Examples of cell-based assays are the live-cell DNA methylation assays, immunofluorescence for DNA methylation [23] , protein-compound interaction assays [24] , flow cytometry assays [25] , and global methylation profiling assays [26] [27] .