Deep Multitask Learning-Driven Discovery of New Compounds Targeting Leishmania infantum

Abstract

Visceral leishmaniasis caused by Leishmania infantum is a severe and often fatal disease prevalent in low- and middle-income countries. Existing treatments are hampered by toxicity, high costs, and the emergence of drug resistance, highlighting the urgent need for novel therapeutics. In this context, we developed an explainable multitask learning (MTL) pipeline to predict the antileishmanial activity of compounds against three Leishmania species, with a primary focus on L. infantum. Then, we screened ∼1.3 million compounds from the ChemBridge database by using these models. This approach identified 20 putative hits, with nine compounds demonstrating significant in vitro antileishmanial activity against L. infantum. Three compounds exhibited notable potencies (IC₅₀ of 1.05-15.6 μM) and moderate cytotoxicities (CC₅₀ of 32.4 to >175 μM), positioning them as promising candidates for further hit-to-lead optimization. Our study underscores the effectiveness of multitask learning models in virtual screening, enabling the discovery of potent and selective antileishmanial compounds targeting L. infantum. Incorporating explainable techniques offers critical insights into the structural determinants of biological activity, aiding in the rational design and optimization of new therapeutics. These findings advocate for the potential of multitask learning methodologies to enhance hit rates in drug discovery for neglected tropical diseases.

Figure 1

Bioactivity distribution of antileishmanial data in classification and regression data sets. (a) Bar graphs illustrating the number of actives and inactives in L. donovani, L. infantum, and L. amazonensis tasks. (b) Histograms illustrate the distribution of pIC₅₀ values for each task. Venn diagram showing the content overlap between tasks in the (c) classification and (d) regression data sets.

Figure 2

Similarity maps illustrating structural diversity and pIC₅₀ values across the L. donovani, L. infantum, and L. amazonensis tasks. Nodes represent individual compounds, and edges indicate similarity based on FragFP descriptors. Gray nodes correspond to untested compounds.

Figure 3

Schematic representation of the (a) MT-MPNN and (b) MT-DNN architectures and corresponding hyperparameters for the classification and regression models explored in this study.

Figure 4

Test set performance of multitask (a) classification and (b) regression models developed using random and scaffold splits.

Figure 5

Average test set (a) ACC, (b) recall, and (c) SP of multitask and single-task classification models developed by using random splitting. The error bar represents the SD over five splitting runs.

Figure 6

Average (a) r, (b) MAE, and (c) RMSE performances of multitask and single-task regression models developed by using random splitting. The error bar represents the SD over five splitting runs.

Figure 7

SHAP summary plot and global feature importance scores for ECFP4 features trained on the L. infantum task. Panel (a) represents features with positive contributions to activity, whereas panel (b) represents features with negative ones. The purple circle denotes the fingerprint’s center with a radius involving atoms denoted by yellow circles. The asterisk denotes a continuation of the structure. Each point represents a sample from the test set, where pink points indicate the presence of the fragment encoded by the bit, and blue points represent its absence.

Figure 8

Virtual screening pipeline used to identify novel compounds against L. infantum. The physicochemical filter was based on aqueous solubility (cLogS > −5.0) and lipophilicity (cLogP between 0.5 and 4.0). A structural novelty filter was applied using Tanimoto similarity with ECFP4 fingerprints to ensure that the virtual hits had no prior experimental activity reported against L. infantum.

Figure 9

Dose–response curves depicting the antileishmanial activity and cytotoxicity of the top three hit compounds.

Figure 10

Structure–activity relationship of central ring substitutions in LC-15 and known L. donovani 20S proteasome inhibitors with antileishmanial activity, utilizing global features prioritized through SHAP values. The bioassay data and chemical structures for the 20S proteasome inhibitors were retrieved from Wyllie et al. and Khare et al. Fragments highlighted in pink represent ECFP4 bits contributing positively to antileishmanial activity, while fragments shown in blue indicate features with negative contributions.