Publications | INFevo OCOO-T Virtual Cell Model

Preprint · 2026

OCOO-T: A Simple and Scalable Virtual Cell Model for Transcriptional Perturbation Response Prediction

Danning JiangZheming AnYalong ZhaoLipeng LaiarXiv:2606.12838

Method summary

OCOO-T is an AI Virtual Cell model for transcriptional perturbation-response prediction. The paper uses flow matching to represent changes between gene-expression states as continuous dynamics, with a standard Transformer operating directly on continuous expression vectors.

Perturbation identity, dose, and cellular context are introduced through adaptive layer normalization and in-context tokens. The paper evaluates chemical, genetic, and cytokine perturbations on the Tahoe100M, Replogle-Nadig, and PBMC benchmarks.

Public benchmarks reported in the paper

Representative metrics across three perturbation settings

0.952

Tahoe100M · PDCorr

Chemical perturbation; the best result for this metric reported in Table 2.

0.833

Replogle-Nadig · PDS_cos

Genetic perturbation; the best result for this metric reported in Table 3.

0.760

PBMC · DEOver

Cytokine perturbation; the best result for this metric reported in Table 4.

Values are taken from the OCOO-T preprint. These metrics measure different properties and should not be interpreted outside their datasets, baselines, and experimental settings. Refer to the paper for complete results and limitations.

Reading guide

Key terms and metrics

AI Virtual Cell: A model system for computationally simulating cellular states and how they change under external perturbations.
PDCorr: A measure of how well a model recovers the structure of response differences across perturbations; higher is better.
DEOver: A measure of overlap between predicted and observed differentially expressed gene sets; higher is better.

Continue exploring

From a published model to the full discovery loop

See how OCOO models connect with automated experiments and autonomous scientific discovery.

Explore technology