How they use Nebius AI Cloud

Compugen searches for immuno-oncology drug targets using massive single-cell and tumor multi-omics datasets. Its Unigen platform integrates multi-omics, spatial and single-cell data to identify new drug targets and resistance pathways. It began collaborating with Nebius through an Academy prototype. Academy graduates worked with Compugen specialists on a TLS classification project using scGPT, scGPT-Spatial and Nicheformer across 14 cohorts and hundreds of thousands of cells. Nebius supplied the project team with a special GPU Cloud grant.

Compugen then used Nebius AI Cloud and NVIDIA GPUs to train a production-grade research model that predicts spatial immune features, including tertiary lymphoid structures, from tumor data. Nebius describes training large models, including single-cell RNA transformers, with models up to 3B parameters running in BF16/FP32. Its life-sciences page also reports fine-tuning on proprietary cancer datasets and training times of 20 seconds to 20 minutes per epoch for 100K-cell RNA-seq models.

Compugen’s research includes a variant designed to infer these features from expression data without spatial coordinates. Nebius reports a best TLS prediction result of 0.94 ROC-AUC without spatial coordinates. Compugen described integrating the models into its Unigen discovery pipelines as a next step. Nebius also describes its infrastructure as helping Compugen integrate AI into that pipeline.