Arc unveils Virtual Cell Challenge

Arc Institute has unveiled the Virtual Cell Challenge, which predicts the effects of gene silencing. This challenge requires context generalization capability to predict changes when a gene is silenced in a specific cell type.

Dataset and Modeling Challenge

• Data composition: It utilizes approximately 300,000 single-cell RNA sequencing profiles. Of the 220,000 cells, about 38,000 are unmanipulated unperturbed cells, which serve as a baseline for distinguishing experimental noise from real signal.
• Technical challenge: Because cells are destroyed during the RNA sequencing process, direct before-and-after comparison of the perturbation is not possible. Therefore, the key to modeling is separating biological heterogeneity from technical noise using the unperturbed cell population as a reference point.

STATE Baseline Model For participants, Arc provides the transformer-based STATE model as a baseline.

• State Embedding Model (SE): Generates rich semantic embeddings of cells to improve generalization performance across cell types.
• State Transition Model (ST): A transformer model using a Llama backbone that acts as a 'cell simulator,' taking the transcriptome of unperturbed cells or SE's embeddings as input and outputting the changed transcriptome upon gene silencing.