GeneICL: A Tabular Foundation Model for Bulk Transcriptomics
GeneICL is a 4.2M-parameter tabular model built with transcriptomics-aware pretraining, not generic scale.
The paper says bulk gene-expression prediction is still hard because the data are high-dimensional, correlated, and often short on labels. GeneICL uses semi-synthetic pretraining from measured bulk expression profiles and a recurrent architecture, then runs inference without gradient updates. The authors report tests across 80 clinical outcome-prediction tasks, including classification, regression, and survival. They say it ranks best overall among the evaluated foundation models and tuned baselines, while using up to 387x fewer parameters and producing laptop-CPU predictions within seconds. ArXiv · AI/CL/LG's note
The paper says bulk gene-expression prediction is still hard because the data are high-dimensional, correlated, and often short on labels. GeneICL uses semi-synthetic pretraining from measured bulk expression profiles and a recurrent architecture, then runs inference without gradient updates. The authors report tests across 80 clinical outcome-prediction tasks, including classification, regression, and survival. They say it ranks best overall among the evaluated foundation models and tuned baselines, while using up to 387x fewer parameters and producing laptop-CPU predictions within seconds. ArXiv · AI/CL/LG's note
score 5