Graph-Based Feature Selection Approach for Molecular Activity Prediction

Is fragment-based graph a better graph-based molecular representation for drug design? A comparison study of graph-based models

Graph Neural Networks (GNNs) have gained significant traction in various sectors of AI-driven drug design. Over recent years, the integration of fragmentation concepts into GNNs has emerged as a potent strategy to augment the efficacy of molecular generative models. Nonetheless, challenges such as symmetry breaking and potential misrepresentation of intricate cycles and undefined functional groups raise questions about the superiority of fragment-based graph representation over traditional methods. In our research, we undertook a rigorous evaluation, contrasting the predictive prowess of eight models-developed using deep learning algorithms-across 12 benchmark datasets that span a range of properties. These models encompass established methods like GCN, AttentiveFP, and D-MPNN, as well as innovative fragment-based representation techniques. Our results indicate that fragment-based methodologies, notably PharmHGT, significantly improve model performance and interpretability, particularly in scenarios characterized by limited data availability. However, in situations with extensive training, fragment-based molecular graph representations may not necessarily eclipse traditional methods. In summation, we posit that the integration of fragmentation, as an avant-garde technique in drug design, harbors considerable promise for the future of AI-enhanced drug design.

Reactivity Graph Yields Interpretable IgM Repertoire Signatures as Potential Tumor Biomarkers

Combining adaptive and innate immunity induction modes, the repertoire of immunoglobulin M (IgM) can reflect changes in the internal environment including malignancies. Previously, it was shown that a mimotope library reflecting the public IgM repertoire of healthy donors (IgM IgOme) can be mined for efficient probes of tumor biomarker antibody reactivities. To better explore the interpretability of this approach for IgM, solid tumor-related profiles of IgM reactivities to linear epitopes of actual tumor antigens and viral epitopes were studied. The probes were designed as oriented planar microarrays of 4526 peptide sequences (as overlapping 15-mers) derived from 24 tumor-associated antigens and 209 cancer-related B cell epitopes from 30 viral antigens. The IgM reactivity in sera from 21 patients with glioblastoma multiforme, brain metastases of other tumors, and non-tumor-bearing neurosurgery patients was thus probed in a proof-of-principle study. A graph representation of the binding data was developed, which mapped the cross-reactivity of the mixture of IgM (poly)specificities, delineating different antibody footprints in the features of the graph-neighborhoods and cliques. The reactivity graph mapped the major features of the IgM repertoire such as the magnitude of the reactivity (titer) and major cross-reactivities, which correlated with blood group reactivity, non-self recognition, and even idiotypic specificities. A correlation between an aspect of this image of the IgM IgOme, namely, small cliques reflecting rare self-reactivities and the capacity of subsets of the epitopes to separate the diagnostic groups studied was found. In this way, the graph representation helped the feature selection in its filtering step and provided reduced feature sets, which, after recursive feature elimination, produced a classifier containing 51 peptide reactivities separating the three diagnostic groups with an unexpected efficiency. Thus, IgM IgOme approaches to repertoire studies is greatly augmented when self/viral antigens are used and the data are represented as a reactivity graph. This approach is most general, and if it is applicable to tumors in immunologically privileged sites, it can be applied to any solid tumors, for instance, breast or lung cancer.