Blank Bio is excited to release Ladon, a transcriptomic foundation model trained on millions of bulk RNA-seq samples. Ladon allows for robust representations of RNA-seq, outperforming baseline gene expression and foundation model methods in predicting prognosis and response on external validation cohorts.
The rapid acceleration of AI innovation in drug design and preclinical work has emphasized the biggest bottleneck in drug development: clinical trials. Clinical trials safeguard patients from inefficacious or unsafe drugs, but come at a tremendous cost and limit the speed at which we can advance medicine. Blank Bio believes clinical trial abundance is achievable without lowering the bar for patient safety by making trials far more likely to succeed.
Many decisions throughout the clinical development pipeline can be aided and accelerated by predictive models. From choosing the target indication to setting the patient enrollment criteria, AI models reduce the need to guess where a drug will work, so fewer trials fail for avoidable reasons. Underlying all of these decisions are two fundamental prediction problems: therapeutic response and patient prognosis. Accurately predicting these outcomes requires capturing heterogeneous patient biology, and in complex diseases like cancer, that heterogeneity is driven by molecular differences between patients.
We introduce Ladon, a state-of-the-art foundation model built for predicting therapeutic response and patient prognosis from transcriptomic data. Ladon is a 3-billion-parameter transformer encoder-decoder trained on millions of RNA-seq samples. Using a self-supervised objective, Ladon learns from every sample, not just the small fraction with outcome labels.
Transcriptomics is rich with biological information, but has historically lacked wider clinical adoption because of a fatal flaw: generalization. Transcriptomic data is sensitive to how it’s generated, while clinical decisions require stable signal across different trial sites, tissue contexts, and even sequencing protocols. An RNA model for patients is only as useful as its accuracy in unseen cohorts. This bar has forced the field to either use rich transcriptomic signatures that don’t transfer, or single-gene markers that sacrifice the modality’s potential.
This is the exact problem foundation models are built to solve. By pre-training on millions of samples spanning technical and biological variation, a foundation model learns to separate clinically transferable biology from the quirks of any one research study. We built Ladon as the first step towards turning 25 years of promise in cancer transcriptomics into better drug development decision-making.
In our technical report, we show that Ladon beats gene expression and foundation model baselines in predicting prognosis and response. Our results are reported on entire clinical trials and cohorts held out from training, directly testing generalization. We also demonstrate how Ladon can find biomarkers of treatment benefit to guide trial enrichment criteria, and how Ladon captures biology like cell composition and isoform usage. Finally, we find that Ladon shows consistent improvement as we scale the model. Below, we highlight several results.
Ladon Predicts Response in Held-Out Clinical Trials
Clinical trials are generalization problems with billion-dollar consequences. The decision on whether to advance drugs to the next phase, which indications to target, and which patients should receive an approved therapy comes down to one question: will this patient respond to the drug? Transcriptomics provides additional signal to help make these decisions.
The complication is that response models rarely get to train on the exact data context they are used in. Response signatures must typically be learned from another trial, another tissue, or cohorts collected in entirely different settings. Historically, it has been hard for foundation models to beat simple approaches such as differential expression, single-gene biomarkers, or even clinical covariates.
Ladon beats these long-standing baselines. We fine-tune Ladon on academic immunotherapy cohorts and test its ability to predict response based on pre-treatment RNA-seq alone in 6 held-out clinical trials across bladder, kidney, lung, and liver cancer. Ladon significantly outperforms all baselines, including PD-L1 expression, the most widely used single-gene immunotherapy biomarker. Ladon also outperforms published expression signatures, demonstrating that its pre-trained representations surface signals not accessible from raw expression.
Ladon’s ability to generalize extends beyond these trials. We assess Ladon’s capability in predicting response and prognosis across large cancer cohorts (TCGA, CPTAC) and academic datasets, where we again see consistent outperformance compared to standard baselines and other foundation models. These results are described in our full technical report, alongside our vision for how improving these capabilities can shorten clinical trials and improve their chance of success.
Read more in the technical report: Ladon predicts outcomes for unseen clinical trials
Ladon Finds Robust Predictive Biomarkers From External Data
One of the most consequential decisions in a drug program lies in choosing which patients to enroll. Clinical trials can use biomarkers, measurable indicators of a patient’s underlying biology, to enrich for those most likely to benefit from the drug. Finding a good biomarker greatly improves the chance of trial success, and billion-dollar oncology programs have lived and died by their biomarker decisions.
Predicting patient response, as above, is a strong starting point for biomarker discovery but, on its own, poses a subtle trap. Consider the counterfactual: if a patient responds to the proposed drug, would they have also responded to any other drug simply because they were healthier to begin with? Trials must prove treatment benefit over the standard of care, and predicting patient response in the treatment arm doesn’t tell you how they would fare in the control arm.
In biomarker terms, this is the distinction between a prognostic biomarker, which identifies patients who do well regardless of which trial arm they are assigned to, and a predictive biomarker, which identifies patients who specifically benefit from the drug relative to the control arm. Enrichment based on predictive biomarkers improves the odds of trial success, but many candidate biomarkers turn out to be prognostic, not predictive.
We use Ladon to develop a putative predictive biomarker for POPLAR, a randomized phase 2 immunotherapy trial in lung cancer. Using external data collected from academic cohorts and other immunotherapy trials, we construct a predictive score using causal machine learning methods that estimates whether a patient benefits from immunotherapy over chemotherapy.
We optimize the Ladon predictive score for progression-free survival (PFS) at 12 months while adjusting for baseline prognosis, and find that patients in the top half of Ladon scores benefit much more from treatment than control, whereas patients in the bottom half see less differential benefit. This indicates that the Ladon score behaves as a predictive (and not merely prognostic) biomarker. In contrast, PD-L1 expression and other published signatures perform worse as predictive biomarkers.
Our experimental setup reflects the reality of drug development — if we want to use a de novo predictive biomarker in a phase 3 trial, it must be established in the preceding phase 2 trial. These trials are typically small, meaning the ability to borrow data from related cohorts is desirable, but this signal must be free of batch effects. Ladon’s ability to effectively transfer in this context reinforces the robustness of its representations. In our technical report, we describe further experimental details and ways to extend this framework to single-arm trials using a synthetic control arm.
Read more in the technical report: Ladon predicts treatment benefit better than existing biomarkers
Ladon Understands Isoform Usage Across Tissues and Sequencing Protocols
Why does Ladon offer better representations? One of Ladon’s unique advantages is its ability to understand isoform usage across tissues and sequencing protocols, whereas most other foundation models read the transcriptome at a low-resolution gene level.
Isoforms are variants of a gene that create related but non-identical proteins with different functions. This allows the same gene to play different roles depending on the tissue context. In cancer, disrupted isoform usage and creation of cancer-specific isoforms impact therapeutic response and prognosis.
Ladon represents RNA-seq at the isoform sequence level, allowing it to natively incorporate the nucleotide sequence differences between isoforms and understand their functional impacts. However, not all transcriptomic sequencing protocols and platforms can actually measure isoform expression, instead only reading out gene-level expression — a coarser measurement than the underlying biology requires. These sequencing protocols include most droplet-based single-cell RNA-seq and single-nucleus RNA-seq (snRNA-seq), and 3’ capture bulk RNA-seq.
Ladon can leverage its understanding of isoform biology to perform “super-resolution” of isoform expression from coarse gene-level inputs. We demonstrate this using data from ENCODE4, which measured expression using paired bulk RNA-seq and snRNA-seq in two organ sites from the same donor. From low resolution pseudo-bulked gene-level snRNA-seq, Ladon reconstructs isoform expression approaching the fidelity of bulk RNA-seq. Across tissues, Ladon outperforms baselines such as taking the reference average from the tissue context or imputation from the sample’s nearest neighbours.
The ability to accept coarse gene-level input while retaining isoform-level insight allows Ladon to be applied in real-world settings that lack isoform measurements. For example, massive perturbational screens such as DRUG-seq are widely used to test drug effects but only read out gene-level expression. Practically, many biological datasets are stored as gene-level counts, even when the original sequencing protocol could resolve isoforms. Ladon can operate effectively in both settings.
In our technical report, we describe the full experimental details and perform additional experiments on Ladon’s performance on isoform super-resolution across tissues, and find that it also accurately predicts the impact of temporal and disease state shifts on isoform distributions.
Read more in the technical report: Ladon understands isoform dynamics across biological contexts
Ladon Recovers Cell Composition Signal Despite Cross-Assay Batch Effects
The composition of cell types within a bulk RNA-seq sample is one of the most informative signals for response and prognosis. Ladon’s representations excel at capturing this signal even when the cell-type references come from a different technology than the samples.
A tumor is a mixture of malignant and non-malignant cells, and its composition shapes how a patient fares. Immune infiltration is one example — immunotherapies work through the immune system, and infiltrating T cells predict who responds. The malignant cells themselves can also be a mix of molecular subtypes that progress and respond to treatment differently. Bulk RNA-seq measures all of these cells only in aggregate, but the aggregate still carries information about their proportions. Deconvolution aims to recover these hidden proportions.
Accurately deconvolving bulk RNA-seq into its constituent cell types requires a reference of what RNA expression in each cell type looks like. These references are now usually constructed from single-cell RNA-seq (scRNA-seq), whose technical differences from bulk RNA-seq outweigh the biological differences between samples. Transferring from one to the other is a difficult generalization challenge for any model.
Ladon bridges this gap by learning a representation shared by bulk samples and by synthetic bulks summed from arbitrary scRNA-seq references, and performs cell type deconvolution. We demonstrate this on a dataset of bulk RNA-seq from peripheral blood samples. Using an external scRNA-seq dataset, we build a classifier on top of Ladon embeddings and find that it outperforms strong existing baselines for cell type deconvolution, including classic deconvolution algorithms and published bulk RNA-seq cell type references.
Understanding the cell composition is critical for interpreting the mechanisms behind response and prognosis. Applying Ladon to a breast tumor atlas reference, we inferred the composition of pre-treatment biopsies from two additional trials, such as ARTEMIS and PBCP, and found that Ladon-derived cell composition predicts therapeutic response more accurately than classifiers built from the same set of genes used in commercial tests such as Oncotype DX and MammaPrint.
Because cell composition is directly interpretable, it provides insight into why Ladon predicts that a patient will respond, in contrast with an opaque black-box score. Furthermore, the ability to ingest scRNA-seq references without overfitting to batch effects allows this capability to be extended to diseases where isolating disease-relevant cell types for bulk RNA-seq is challenging or unknown. In our technical report, we explore deconvolution across more trials, show detailed results on the deconvolution based on scRNA-seq references, and visualize the compositional feature space learned by Ladon.
Looking Forward
Advances in contemporary AI are achieved through a predictable decrease in pre-training loss as model size and training compute increase. Scaling laws describe how these changes reduce pre-training loss, a measure of error on the model’s training task. This makes it valuable to identify objectives that both improve with additional compute and carry those improvements through to downstream tasks.
We assembled ClinBench, a patient-level transcriptomics benchmark covering survival and treatment response tasks. We find that lower reconstruction loss on Ladon’s pre-training objective is strongly correlated with higher aggregate performance on ClinBench. This suggests that improvements during pre-training capture information relevant to clinical prediction.
We believe Ladon is a floor on the performance of foundation models for bulk transcriptomics. These results give us a clear direction for further development. Increasing parameter count, training for longer, and expanding our training data are paths we are excited to explore as we continue improving Ladon’s predictions and exploring how generative model capabilities can further de-risk and accelerate clinical development.
Read more in the technical report: Ladon performance improves as we scale model size and compute
Conclusion
In this blog, we introduce Ladon and show how it effectively harnesses the biological signal in transcriptomics to make robust predictions of prognosis and treatment response in patients. We believe these are foundational prediction tasks that ground the capabilities needed to unlock clinical trial abundance. We’re building towards a future in which we equip drug developers and clinicians with tools like Ladon to run smaller, more efficient trials, with higher success rates. Ultimately, we think it’s important to build on the capabilities that we’ve highlighted today and increasingly shift our ability to de-risk clinical trial decision-making in silico. Over the next few months, this is what we’ll be focused on building. If you’re interested in using the model today and working closely with our team to adapt it to your most pressing problems, please reach out!
Thanks to Karina, Daniele, Ben and Manu for early feedback on this blog.
Header artwork by Visual Cinnamon.