InsilicoCell selects molecules for laboratory experiments using several criteria at once
InsilicoCell selects molecules for laboratory experiments using several criteria at once
On August 28, the authors posted a preprint on InsilicoCell, a model pretrained on more than 88 million biological measurements with known outcomes across seven tasks. It combines predictions that are usually generated separately and produces a short list of molecules for laboratory testing. The authors demonstrated the approach in liver cancer cells and a model of pulmonary fibrosis.
Selecting a molecule for an experiment requires several decisions: whether it can alter the target cellular process, slow the growth of tumor cells, and preserve the viability of other cells. When these properties are assessed sequentially, a candidate may pass an early stage but fail at the next one. InsilicoCell combines several requirements into a single selection rule before the compound is purchased.
The same research group previously introduced GPS, a model that predicts changes in cellular gene activity from a compound's structure. InsilicoCell extends this approach. For different tasks, it uses the compound's structure and the cellular context. When predicting changes in gene activity, it also considers the dose and duration of exposure. Training on seven types of measurements allows the model to generate related predictions and apply them within a single selection process.
For the HepG2 liver cancer cell line, the authors specified three criteria: a molecule had to slow cell growth, reduce expression of the MYC gene, and inhibit the activity of c-Myc, a protein that regulates many genes. About 1.8 million compounds passed through these three filters, and five met all the criteria simultaneously. In cell experiments, all five reduced the amount of c-Myc and slowed HepG2 growth, while two molecules inhibited c-Myc activity.
For idiopathic pulmonary fibrosis, a disease that causes lung scarring, the model searched for molecules predicted to reduce three scarring markers in myofibroblasts, the cells that form scar tissue, while preserving fibroblasts, which are connective tissue cells. Five candidates were selected from 4.7 million molecules. Two compounds predicted to be toxic did damage the cells. Two of the remaining three reduced scarring markers in the cell model. One of these was then tested in six independent experiments using lung tissue slices from patients with fibrosis, where it reduced the levels of FN1 and COL1A1, proteins associated with scarring.
In this case, virtual screening addresses a specific laboratory task: before a compound is purchased, it identifies molecules predicted to meet several measurable criteria for experimental testing.