RevViability

Documentation Index

Fetch the complete documentation index at: https://docs.revilico.bio/llms.txt

Use this file to discover all available pages before exploring further.

​

Why Use This Engine?

In the documentation below, we will use Revilico’s RevViability engine to predict how a drug will affect cancer cell viability at different concentrations without running a wet-lab assay. RevViability generates full dose-response curves, IC50 values, sensitivity classifications, and selectivity indices across a panel of cancer cell lines, enabling researchers to prioritize compounds and cell line selections before committing to experimental resources.

​

Background

Cell viability assays measure the fraction of living cells after drug treatment at a range of concentrations, producing a dose-response curve that is fit to the Hill sigmoid model to extract the IC50 (the concentration at which 50% of cells are killed). Running these assays experimentally across large compound libraries and diverse cell line panels is time-consuming and costly. The Genomics of Drug Sensitivity in Cancer (GDSC) project has generated over 467,000 drug-cell line viability measurements across 408 drugs and 978 cell lines, providing a large-scale training resource for predictive models. RevViability uses a GNN+MLP hybrid model trained on GDSC1, GDSC2, and DepMap CCLE 24Q4 data. The graph neural network encodes the molecular structure of the drug as a graph of atoms and bonds, learning structural features relevant to cytotoxicity. The MLP combines these drug features with genomic features of the cell line (gene expression, copy number, mutation status) to predict the dose-response relationship. The model achieves R-squared of 0.87 on held-out test data against experimental GDSC values. Predictions are available for 972 CCLE-mapped cell lines.

​

Dose-Response Modeling

The predicted dose-response relationship for each drug-cell line pair follows the Hill equation: $$v(c) = 1 - \frac{c^n}{c^n + \text{IC}_{50}^n}$$ Where $v(c)$ is the viability fraction at concentration $c$, $n$ is the Hill coefficient controlling curve steepness, and IC50 is the half-maximal inhibitory concentration. The curve is evaluated at 64 concentration points spanning the user-defined range. Exposure time correction: The base model IC50 corresponds to the GDSC standard 72-hour exposure. When the user selects a different exposure time $t$, the IC50 is corrected by: $$\text{IC}_{50}^{\text{eff}} = \text{IC}_{50}^{\text{base}} \times \left(\frac{72}{t}\right)^{0.4}$$ This correction reflects the empirical relationship between exposure duration and apparent potency. Biological variability: When replicates are requested, Gaussian noise calibrated to match GDSC inter-replicate coefficient of variation is added to the viability predictions, and IC50 confidence intervals are computed from the replicate distribution.

​

Sensitivity Classification

IC50 values are classified into four sensitivity categories based on clinically relevant concentration thresholds:

​

Selectivity Index

When multiple cell lines are profiled, the selectivity index quantifies how selective a drug is for the most sensitive cell line relative to the least sensitive: $$\text{SI} = \frac{\text{IC}_{50}^{\max}}{\text{IC}_{50}^{\min}}$$ A high SI indicates on-target selectivity (one lineage is significantly more sensitive than others), which is the expected signature of a mechanism-based inhibitor. A low SI indicates pan-cytotoxicity.

​

Media pH Simulation

RevViability models media acidification over the exposure time based on CO2 production and lactate accumulation from cellular metabolism. If pH drops below 6.8, a viability penalty factor is applied to correct for acidic stress-induced cytotoxicity that is independent of the drug mechanism. This provides more accurate predictions for long-duration assays at high cell densities.

​

GDSC Reference Comparison

For drugs in the GDSC training database, the predicted IC50 is compared against the published experimental reference value and the fold deviation is reported. A fold error below 2x indicates excellent agreement. Fold errors between 2x and 5x are within the range of expected inter-laboratory variability in experimental viability assays.

​

Batch Screening

RevViability supports screening up to 50 compounds simultaneously against any selection of the 972 available cell lines. Batch inputs accept compound name plus SMILES (one per line) or direct SMILES input. Results are returned as a sortable table of IC50 values and sensitivity classifications per compound per cell line, with CSV download for downstream analysis.

​

Plate Layout and Assay Configuration

The engine simulates a standard 96-well plate format. Users assign cell lines and concentrations to individual wells or groups of wells using serial dilution presets, row and column selectors, or manual address entry. Four-cell-line layouts automatically fill all 96 wells with a serial dilution for each of four cell lines across two rows each. Well-level annotations can be assigned for control labeling and audit trail purposes.

​

Running the Engine

​

Inputs

​

Outputs

• IC50 values: Exposure-corrected IC50 per cell line with 95% confidence interval
• Dose-response curves: 64-point Hill equation fits per cell line
• Sensitivity classification: Highly Sensitive, Sensitive, Intermediate, or Resistant badge per cell line
• Selectivity index: Ratio of maximum to minimum IC50 across all profiled cell lines
• AUC: Area under the dose-response curve (lower = more sensitive overall)
• GDSC comparison: Fold deviation from experimental reference IC50 where available
• Waterfall plot: Cell lines rank-ordered by IC50 with error bars
• Heatmap: Full 96-well plate colored by predicted viability
• Lineage plot: IC50 by tissue type showing tissue-level selectivity
• ln(IC50): Log-scale IC50 matching GDSC ln_ic50 column format for direct database comparison
• Protocol export: Plain-text lab notebook format with plate layout, parameters, and results
• CSV export: Full well-level data with concentrations, viability values, IC50, and annotations