Overview

MuSyC, originally published in Meyer CT, Wooten DJ, Paudel BB, et al. Cell Systems. 2019 and updated in Wooten DJ, Meyer CT, et al., under review, 2020, is a framework for calculating drug synergy which distinguishes between different types of synergistic interactions. Synergistic efficacy (beta) measures the changes in maximal effect over single agents due to the combination. Synergistic potency (alpha) measures the change in potency of one drug given the presence of the other drug. Importantly these types of synergy align with common clinical motives for treating diseases with drug combinations: improve outcomes by escalating effect (synergistic efficacy) and reduce off-target toxicity by minimizing doses (synergistic potency).

FAQ

How do I use this site?

See the Help page. Use is subject to Terms and Conditions.

Can you quantify for 3 drug case?

Currently, the MuSyC portal can only handle the two drug case. For collaborative inquiries of this nature, please email musyc@gmail.com.

What is the optimal dose for synergy?

Contrary to prior frameworks, MuSyC quantifies the synergy of a drug combination. Once the best combination from the screen is selected, users should look for a minimum dose that achieves the desired effect magnitude. In other words, dose optimization IS ALWAYS done based on the observed effect. MuSyC helps in identifying combinations for which the desired effect is achievable by the combination but not the single drugs (synergistic efficacy) or where the doses required to achieve that effect are lowered due to the drugs interacting (synergistic potency).

What if I only have 1 condition of my second drug or another binary condition (e.g. gene knockout) plus and minus a dose titration experiment of the first drug?

The MuSyC fitting algorithm currently handles such cases by assuming the binary condition to satisfy [drug2]->inf. In this case, the MuSyC equation reduces to a Hill equation with an EC50 defined by C1/alpha1. See Section 6 of Supplement in (Wooten DJ et al. 2020) for proof of this condition.

How do I separate batches?

Use a unique identifier in the optional "batch" column of the upload. Each batch will be self-contained and not sampled from for fitting dose-response surfaces from other batches.

What is the minimal number of samples I need/what is the optimal sampling strategy?

We have found the MuSyC framework to be fairly robust in a wide range of sample density and designs. See Figure S4 in Wooten et al. 2020 for complete analysis. However, the exact sampling design requirements are idiosyncratic to the noise profile of a particular assay; therefore, no universal standard exists. Typically, the Matrix (also called Checkerboard) sampling strategy is most robust at the cost of higher data density demands. For extremely limited sampling where the full dose-response profile of each single agent cannot be captured, we recommend using Highest Single Agent (HSA) at the max concentration of both compounds as HSA approximates synergistic efficacy in this condition. Subsequent screens can identify synergistically potent combinations from the hits by increasing the sampling.

What is the fitting algorithm used?

The current fit algorithm leveraged in the MuSyC portal is described in the Methods section of Wooten et al. 2020. We use a Monte Carlo algorithm as suggested by Motulsky and Christopoulos (Chapter 17, pg 104) for estimating asymmetric 95% confidence intervals of each parameter. Briefly, this is done by fitting all the data using standard non-linear least squares regression (TFR option in SciPy's curve_fit). Based on this optimal fit, noise is added to every data point proportional to the root mean square error of the optimal fit. The new "noise-added" data is then fit again to generate a new parameter set. This process is run 100 times and the 95% confidence intervals for all parameters are calculated from the ensemble.

Should I apply some data smoothing before fitting?

According to Motulsky and Christopoulos, data should not be smoothed before fitting because this can arbitrarily reduce the noise dispersion in non-linear ways resulting in noise-profiles that are not homoskedastic, a common assumption of most non-linear least-square optimizers.

Funding

CTM was supported by National Science Foundation (NSF) Graduate Student Fellowship Program (GRFP) [Award #1445197]; CFL and DJW were supported by the National Science Foundation [MCB 1411482 and MCB 1715826, respectively]; CFL and VQ were supported by the National Institutes of Health (NIH) [U54-CA217450 and U01-CA215845]; VQ was supported by NIH [R01-186193].