Rationale prediction of drug combinations based on large-scale perturbation proteomics

The pharmaceutical industry is hindered by high failure rates, steep expenses, and sluggish drug development. Drug repositioning can alleviate these challenges by slashing costs, expediting development, and unveiling new targets for existing drugs, but standalone efficacy is often limited by low plasma concentrations. Drug combinations may enhance repositioning success, but the vast number of possible pairings is a significant obstacle. Current computational methods, which rely heavily on genomic and transcriptomic data, fall short in systematically identifying successful drug repositioning combinations. We propose integrating a drug repurposing library and post-perturbation proteomic data to enhance the success rate of drug combination repositioning.
In our study, we focused on three cell lines, namely NCI-H1299, HCT116, and 786-O. To identify effective drug combinations, we implemented a 2x7 concentration matrix approach, using 25 cancer-specific anticancer drugs as 'anchor' compounds, each at two optimized concentrations. Leveraging the molecular representation learning model, Uni-Mol, we carried out unsupervised clustering on 3000 repurposed drugs, and picked 192 library drugs, each set at seven concentration points across a 1000-fold range.
We conducted perturbed proteomic analysis of 3000 drugs, and 300 each of effective and ineffective drug combinations on these cell lines at two time points (6 and 24 hrs), using data-independent acquisition on Orbitrap Exploris™ 480 Mass Spectrometers. We generated over 40,000 viability measurements for over 4000 pairwise drug combinations across three cell lines—the largest dataset of its kind focusing on repurposed drugs for anticancer regimens. Combinations demonstrating shifts beyond Bliss in potency (ΔIC50 > 3) or efficacy (ΔEmax > 0.2) were classified as synergistic. Notably, less than 10% of the combination-cell line tuples exhibited synergy, underlining the rarity of effective combinations. Furthermore, synergistic drug combinations showed a strong correlation with tissue origin.
Next, we produced over 20,000 perturbed proteomic data, quantifying over 10,000 proteins with high confidence and technical reproducibility. Unsupervised clustering showed cell line-specific proteome responses to perturbations, highlighting the cell context-dependence of the perturbation proteome. We compared perturbed proteomic data between synergistic and ineffective drug combinations to uncover potential molecular mechanisms of synergy effects.
Utilizing baseline genomic data, perturbed proteomic data, and SMILE as training inputs, we developed a neural Ordinary Differential Equations model to predict the efficacy of independent drug combinations from a pool of repurposing drugs and anticancer drugs. The model achieved an AUC above 0.80. We further validated a few representative drug synergies in cell line models and explored the targets of the selected drug combinations by thermal proteome profiling experiments.