We thank all data authors (incl. Kory Lavine, Patrick Ellinor, and Norbert Hubner’s labs, among others) and enrolled patients. We acknowledge funding from DFG through CRC1550. All code available at github.com/saezlab/rehe...

This study was co-led by @ricoramirez.bsky.social and @jlanzer.bsky.social with supervision by @juliosaezrod.bsky.social and help by Jose Linares, and the group of Norbert Frey (Marco Steier and Ashraf Rangrez) who performed the experimental work.

By integrating transcriptomic datasets across cohorts and technologies, we provide a reference for examining multicellular aspects of heart failure.

An interactive platform allows users to explore gene-expression patterns and project new samples: tinyurl.com/bdwxdf6j

New patients can be positioned on the map:
In a cohort receiving LVADs (data from Kory Lavine’s lab), molecular changes in the map were consistent with clinical improvement, suggesting that the map may help characterize treatment responses.

This allowed us to distinguish genes driven by compositional changes from those primarily affected by molecular regulation.

We also updated the heart-failure transcriptional signature derived from bulk studies and combined it with our multicellular programs to examine how different cell types contribute to gene dysregulation.

The coordinator role of fibroblast was better characterized by a broad phenotypic shift rather than the accumulation of specific cell states.

We also uncovered a network of cell-type dependencies underlying these multicellular programs, with fibroblasts playing a central role, particularly coordinating with cardiomyocyte reprogramming (where we validated several ligand candidates consistent with this interaction).

Using Multicellular Factor Analysis, a patient-level integration method (doi.org/10.7554/eLif...), we constructed a transcriptional patient map summarizing multicellular gene-expression variation in heart failure along two main axes.

Across studies, gene-expression changes associated with heart failure showed a reproducible pattern in both bulk and single-nucleus data. Changes in cell-type composition were more variable, indicating that these two aspects of tissue remodelling may not always occur together.

We curated an extensive compendium of >1500 patients profiled w bulk or single-nuc transcriptomics, building on our previous work tinyurl.com/sn6dxu95. This data engineering effort enabled the comparison and integration of insights to generate a reference of Heart Failure.

It was also supported by projects: AI4FOOD-CM (Y2020/TCS6654), FACINGLCOVID-CM (PD2022-004-REACT-EU), INTER-ACTION (PID2021-126521OB-I00 MICINN/FEDER), HumanCAIC (TED2021-131787BI00 MICINN), PowerAI+ (SI4/PJI/2024-00062 Comunidad de Madrid and UAM), and Cátedra ENIA UAM-VERIDAS.

This work was supported through state funds approved by the State Parliament of Baden-Württemberg for the Innovation Campus Health + Life Science Alliance Heidelberg Mannheim.

This work was a nice collaboration with Ph.D. student Sergio Romero, supervised by professors Ruben Tolosana and Aythami Morales (UAM, Spain), who was visiting us for 3 months to work on this project with @pablormier.bsky.social and @martingarridorc.bsky.social.

🔧 We developed ScAPE using #Keras 3, so you can use #Tensorflow, #JAX or #PyTorch as backends
👉 github.com/scapeML/scape

We benchmarked ScAPE against:
🔹 Other winning challenge methods
🔹 TabPFN, a foundation model for tabular data

➡️ ScAPE matches or outperforms them, showing the value of simple, efficient baselines.

Despite its simplicity, ScAPE ranked among the top methods in the challenge.
It generalizes across new drug–cell combinations and offers a robust baseline for evaluating novel approaches.

✨ ScAPE (Single Cell Analysis of Perturbational Effects)

- Lightweight neural network (∼19M params)
- Uses only aggregated gene-level stats (robust + simple)
- Multi-task: predicts both significance (p-values) & effect size (fold-change)