Task Navigator

Harnessing AI to Build Virtual Cells

VCHarness is an autonomous system for building perturbation-response models that predict how gene expression changes after genetic or chemical perturbations. Given a dataset, a cell context, and an evaluation objective, it combines multimodal biological foundation models, an AI coding agent, and Monte Carlo Tree Search (MCTS) to search for strong model designs with far less manual iteration. This page showcases the Essential dataset case study from the paper, focusing on the four classification tracks for HepG2, Jurkat, hTERT-RPE1, and K562.

Our Paper Our Blog

4 cell linesclassification tracks

618 nodescurrently indexed

Essential Benchmark

Essential is a CRISPR knockdown benchmark spanning HepG2, Jurkat, hTERT-RPE1, and K562. It contains 964,451 single cells, 56 GEM-group batches, and 133,784 unique batch-perturbation pairs, with 2,387 perturbed genes for HepG2, Jurkat, and hTERT-RPE1, and 2,054 for K562.[1][2][3]

Our task is to predict, for 6,640 target genes in each cell line, whether expression is significantly up-regulated, unchanged, or significantly down-regulated after perturbation.

Visible Cell Lines

Loaded from local task folders.

Total Leaves

Current frontier across all trees.

Deepest Search

Maximum observed path depth.

Total Split Samples

Train, validation, and test across Fold_0.

Task Entry

Harnessing AI to Build Virtual Cells

Choose a cell line