UniPert-G2CP trains a model to predict how a molecule will change a cell's state
UniPert-G2CP trains a model to predict how a molecule will change a cell's state
On July 24, Cell published a paper on UniPert-G2CP. The model links the results of genetic experiments to cellular responses to small molecules, the chemical compounds that often serve as starting points in drug discovery.
A cell does not respond to a drug through a single switch. A molecule binds to a target protein and then alters the activity of many genes. The same drug can produce different responses in different cell types. Chemical screening therefore requires many separate experiments. For each molecule, researchers treat cells and measure which RNAs have changed.
A genetic screen provides a much broader map of these cellular states. In a common approach, CRISPR disables one gene at a time, after which researchers measure the cell's RNA profile. The resulting map shows the state produced by perturbing each gene. In May, TxPert predicted transcriptional responses to new genetic perturbations. UniPert-G2CP links such maps to the structures of small molecules. The Connectivity Map within the LINCS program already collects profiles of genetic perturbations and compound treatments. These profiles allow researchers to compare the signatures that genes and drugs leave in a cell.
UniPert-G2CP uses this map to search for chemical compounds. The model takes a target protein and a molecular structure, converts them into a shared representation, and compares that representation with a measured cellular profile. Genetic experiments show how a cell responds when a specific gene is disabled. Chemical experiments link similar profiles to actual molecules.
The authors tested this transfer using LINCS data covering 4 994 genes and 7 860 molecules across five cancer cell lines. When one fifth of the chemical measurements were retained for training, pretraining on genetic screens increased the mean correlation between the predicted and measured transcriptomes by 375,4%. This figure measures the agreement between two RNA profiles, not the number of drugs identified.
Finding compounds that alter age-related cellular states requires the same sequence of steps. Researchers must characterize the cellular effect of a protein target and then select molecules that could produce the desired profile. UniPert-G2CP provides a way to narrow this selection before candidates undergo laboratory testing.