Blank Bio releases Ladon, a 3-billion-parameter RNA foundation model that predicts immunotherapy response in six held-out trials

Ladon moves transcriptomic foundation-model evaluation onto entire held-out clinical trials, testing generalization across cohorts rather than fit within one dataset. Transcriptomic models have relied on rich expression signatures that do not transfer, or fallen back on single-gene markers that sacrifice the modality; foundation models have long struggled to beat simple approaches such as differential expression, single-gene biomarkers or even clinical covariates. After fine-tuning on academic immunotherapy cohorts, Ladon predicts response from pre-treatment RNA-seq alone in six held-out trials spanning bladder, kidney, lung and liver cancer, significantly outperforming all baselines including PD-L1 expression, the most widely used single-gene immunotherapy biomarker, as well as published expression signatures.

The model carries transcriptomic representations in a 3-billion-parameter transformer encoder-decoder, and a self-supervised objective lets samples without outcome labels contribute to learning. Only a small fraction of samples carry outcome labels, so conventional supervised training cannot use the rest. Pretraining spans millions of RNA-seq samples across technical and biological variation; on large cancer cohorts including TCGA and CPTAC and on academic datasets, response and prognosis prediction shows consistent gains over standard baselines and other foundation models.

Ladon represents RNA-seq at the isoform sequence level, so it can handle differences in isoform usage across tissues and sequencing protocols. Most other foundation models read the transcriptome at a low-resolution gene level, while protocols such as droplet single-cell RNA-seq, single-nucleus RNA-seq and 3' capture bulk RNA-seq only read out gene-level expression. On ENCODE4 data pairing bulk RNA-seq and snRNA-seq from two organ sites of the same donor, Ladon reconstructs isoform expression from pseudo-bulked gene-level snRNA-seq approaching the fidelity of bulk RNA-seq, outperforming baselines such as the tissue reference average or imputation from nearest neighbours.

The demonstrated scope is prognosis and response prediction from pre-treatment bulk RNA-seq, construction of a predictive biomarker from external data, isoform-expression super-resolution, and cell-composition deconvolution across sequencing protocols; the POPLAR score was built retrospectively on a randomized phase 2 trial rather than validated prospectively.

Ladon is described as a floor on the performance of foundation models for bulk transcriptomics, with larger parameter counts, longer training and expanded data named as next directions; the predictive score on POPLAR borrows data from related cohorts, and the report stresses that this signal must be free of batch effects while noting the framework can extend to single-arm trials using a synthetic control arm.