Multimodal Graph Neural Network and Natural Language Processing Model for Antileishmanial Molecule Screening

Neglected Tropical Diseases, such as leishmaniasis, suffer from a lack of new treatments due to limited interest from the pharmaceutical industry, motivating scalable computational strategies for virtual screening of therapeutic candidates, aiming to reduce the investment barrier in drug discovery. This work presented an integrative Deep Learning framework for the screening of compounds against multiple Leishmania species, combining self-supervised molecular representation learning, GPU-accelerated construction of chemical similarity networks, multimodal feature fusion within a custom "MoleculeSAGE" architecture, and semi-supervised active learning. The molecular node embedding fused three complementary views under learnable weights: a graph neural encoder applied to the atom-level molecular graph, a fixed-length fingerprint learned by a Tanimoto-optimized autoencoder, and, for labeled bioactivity data, a bioassay embedding. This embedding combined PubMedBERT encodings of assay descriptions, restructured from free text into categorical fields (e.g., target species, readout method) by a locally run large language model (Qwen 3.5 9B), with a species one-hot vector. The resulting graph comprised 296,646 molecular nodes, 22,050 of them labeled ChEMBL bioactivities and 274,596 unlabeled virtual-library compounds, connected by 6,170,742 edges. Labeled compounds were partitioned via a species-stratified Murcko scaffold split into 17,666 training, 2,193 validation, and 2,191 test compounds. On the test set, the model achieved an MCC of 0.5163, F1-score of 0.7431, and accuracy of 0.7589, with per-species F1 ranging from 0.8476 for L. major to 0.5455 for the smaller L. braziliensis subset (n=51). Benchmarked against a Random Forest baseline under a shared k-NN applicability-domain, MoleculeSAGE achieved higher out-of-domain F1 in all 30 grid configurations tested and degraded less on average, with F1 falling from 0.777 to 0.658 versus 0.768 to 0.601 for the RF, a roughly 29% smaller drop. Together, these results position the framework as an open, scalable pipeline for accelerating antileishmanial drug discovery in an underfunded research area.