ABSTRACT
For large libraries of small molecules, exhaustive combinatorial chemical screens become infeasible to perform when considering a range of disease models, assay conditions, and dose ranges. Deep learning models have achieved state-of-the-art results in silico for the prediction of synergy scores. However, databases of drug combinations are biased toward synergistic agents and results do not generalize out of distribution. During 5 rounds of experimentation, we employ sequential model optimization with a deep learning model to select drug combinations increasingly enriched for synergism and active against a cancer cell line-evaluating only â¼5% of the total search space. Moreover, we find that learned drug embeddings (using structural information) begin to reflect biological mechanisms. In silico benchmarking suggests search queries are â¼5-10× enriched for highly synergistic drug combinations by using sequential rounds of evaluation when compared with random selection or â¼3× when using a pretrained model.
Subject(s)
Computational Biology , Neoplasms , Humans , Drug Synergism , Computational Biology/methods , Drug Combinations , Neoplasms/drug therapyABSTRACT
Epithelial-mesenchymal transition (EMT) endows stem cell-like properties to cancer cells. Targeting this process represents a potential therapeutic approach to overcome cancer metastasis and chemotherapy resistance. FiVe1 was identified from an EMT-based synthetic lethality screen and was found to inhibit the stem cell-like properties and proliferation of not only cancer cells undergoing EMT, but also more broadly in mesenchymal cancers that include therapeutically intractable soft tissue sarcomas. FiVe1 functions by directly binding to the type III intermediate filament protein vimentin (VIM) in a mode that induces hyperphosphorylation of Ser56, which results in selective disruption of mitosis and induced multinucleation in transformed VIM-expressing mesenchymal cancer cell types. Cell-based potency (IC50 = 1.6 µM, HT-1080 fibrosarcoma), poor solubility (<1 µM) and low oral bioavailability limits the direct application of FiVe1 as an in vivo probe or therapeutic agent. To overcome these drawbacks, we performed structure-activity relationship (SAR) studies and synthesized a set of 35 new compounds, consisting of diverse modifications of the FiVe1 scaffold. Among these compounds, 4e showed a marked improvement in potency (IC50 = 44 nM, 35-fold improvement, HT-1080) and cell type selectivity (19-fold improvement), when compared to FiVe1. Improvements in the potency of 4e, in terms of overall cytotoxicity, directly correlate with VIM Ser56 phosphorylation status and the oral bioavailability and pharmacokinetic profiles of 4e in mouse are superior to FiVe1. Successful optimization also resulted in potent and selective derivatives 11a, 11j and 11k, which exhibited superior pharmacological profiles, in terms of metabolic stability and aqueous solubility. Collectively, these optimization efforts have resulted in the development of promising FiVe1 analogs with potential applications in the treatment of mesenchymal cancers, as well as in the study of VIM-related biology.