CLC-Pred Synergy

predicts the probability of pairwise drug combinations exhibiting synergistic anticancer cytotoxicity against 45 cancer cell lines from the NCI-60 panel, requiring only the structural formulas of the two compounds as input. Cancer pharmacotherapy increasingly relies on drug combinations to overcome resistance and reduce adverse effects compared to monotherapy. However, the vast combinatorial space of possible drug pairs makes experimental screening prohibitively expensive and time-consuming. CLC-Pred Synergy addresses this by applying structure–activity relationship (SAR) modeling to predict synergy in silico, building on the predecessor tool CLC-Pred, which predicted cytotoxicity of individual compounds against hundreds of cell lines.

How It Works?

The models were trained on data from the NCI-ALMANAC database, which contains over 300,000 cytotoxicity experiments covering ~ 5 000 pairwise drug combinations tested against the NCI-60 panel of 60 human cancer cell lines. Chemical structures of compound pairs are encoded using PoSMNA (Pairs of Substances Multilevel Neighborhoods of Atoms) descriptors, which are a a pairwise extension of the MNA descriptor system developed specifically for drug–drug interaction modeling. A modified Naive Bayes classifier implemented in the PASS DDI software then learns structure–synergy patterns from these descriptors.

From 1,500 generated SAR models (covering 60 cell lines × 5 synergy metrics × 5 thresholds), 104 high-quality models with AUC > 0.70 were selected, corresponding to 45 cancer cell lines and achieving a mean AUC of 0.749. We used a rigorous leave-one-out compounds-out cross-validation scheme to validate the models. This scheme ensured that no compound from a test pair appeared in any training pair, thus genuinely simulating prediction for novel molecules.

Input:

The structural formulas of two compounds, entered either as SMILES, retrieved by drug name search or draw using chemical editor.

Output:

A ranked table of predictions across 45 cancer cell lines, with the following columns for each cell line / synergy model combination:

Five synergy scoring frameworks are covered simultaneously: Bliss, ComboScore, HSA, Loewe, and ZIP. Combinations where both Pa and Pi are low indicate the pair falls outside the applicability domain of the model.

Validation Against Approved Drug Combinations

The service was validated on 23 clinically approved synergistic drug pairs (for melanoma, breast, colon, kidney, lung, and ovarian cancers) not present in the training sets. CLC-Pred Synergy correctly predicted synergistic activity for 17 of 23 pairs (73.9%), including notably predicting the encorafenib–binimetinib combination against melanoma cell lines harboring the BRAF V600E/V600K mutation, consistent with its FDA approval in 2018.

Practical Use:

Within seconds, the application returns a ranked prediction table covering up to 45 NCI-60 cancer cell lines across five synergy models (Bliss, ComboScore, HSA, Loewe, and ZIP). A typical workflow would involve screening several candidate combination partners for a lead compound, ranking them by Pa–Pi across relevant cancer cell lines, and prioritizing the top-scoring pairs for subsequent in vitro cytotoxicity assays. This significantly narrows the experimental search space before any laboratory work begins.