Identifying an optimal perturbation to induce a desired cell state by generative deep learning

Controlling cell states is pivotal in biological research, yet understanding the specific perturbations that induce desired changes remains challenging. To address this, we present PAIRING (perturbation identifier to induce desired cell states using generative deep learning), which identifies cellular perturbations leading to the desired cell state. PAIRING embeds cell states in the latent space and decomposes them into basal states and perturbation effects. The identification of optimal perturbations is achieved by comparing the decomposed perturbation effects with the vector representing the transition toward the desired cell state in the latent space. We demonstrate that PAIRING can identify perturbations transforming given cell states into desired states across different types of transcriptome datasets. PAIRING is employed to identify perturbations that lead colorectal cancer cells to a normal-like state. Moreover, simulating gene expression changes using PAIRING provides mechanistic insights into the perturbation. We anticipate that it will have a broad impact on therapeutic development, potentially applicable across various biological domains.