Abstract
Sequence-to-function analysis is a challenging task in human genetics, especially in predicting cell-type-specific multi-omic phenotypes from biological sequences such as individualized gene expression. Here, we present UNICORN, a new method with improved prediction performances than the existing methods. UNICORN takes the embeddings from biological sequences as well as external knowledge from pre-trained foundation models as inputs and optimizes the predictor with carefully-designed loss functions. We demonstrate that UNICORN outperforms the existing methods in both gene expression prediction and multi-omic phenotype prediction at the cellular level and the cell-type level, and it can also generate uncertainty scores of the predictions. Moreover, UNICORN is able to link personalized gene expression profiles with corresponding genome information. Finally, we show that UNICORN is capable of characterizing complex biological systems for different disease states or perturbations. Overall, embeddings from foundation models can facilitate the understanding of the role of biological sequences in the prediction task, and incorporating multi-omic information can enhance prediction performances.
Introduction
Predicting how specific DNA sequences translate into cellular behavior—the sequence-to-function challenge—remains a fundamental goal in human genetics. While genomic data is abundant, predicting cell-type-specific expression (e.g., distinguishing gene behavior in neurons versus immune cells) is difficult. Existing models like Enformer or Basenji2 primarily utilize "bulk" data, which obscures the unique signatures of individual cell types. Furthermore, single-cell sequencing, though precise, is characterized by high noise and sparsity. There is a critical need for a flexible framework capable of handling individualized genetic variations and multi-omic data while providing reliable uncertainty estimates.
Methods: The UNICORN Framework
The authors present UNICORN (UNIversal Cell expressiOn pRedictioN), a multi-task learning framework designed as a sophisticated "adapter." Rather than training a massive model from scratch, UNICORN leverages pre-trained foundation models to extract biological features.
The framework operates in two primary stages:
- Embedding Extraction: Biological sequences (DNA near transcription start sites or amino acid sequences) are processed through foundation models such as Enformer (genomics), ESM2 (proteins), and GPT-3.5 (functional descriptions).
- Lightweight Prediction: These high-dimensional embeddings are fed into a two-layer Multi-Layer Perceptron (MLP). This "adapter" approach allows for rapid training without full-parameter fine-tuning.
Key innovations include a multi-task learning objective that jointly predicts various modalities (e.g., RNA expression and ATAC-seq peaks), a custom loss function combining Pearson correlation and Poisson log-likelihood, and a "LossPred" module for uncertainty estimation.
Results and Performance
UNICORN demonstrates significant improvements in both accuracy and computational efficiency over existing baselines.
Computational Efficiency
A standout feature of UNICORN is its training speed. By utilizing frozen embeddings, the model achieves a training time of approximately 2 seconds per epoch, a massive reduction compared to the 2,242 seconds required by seq2cells.
Prediction Accuracy and Robustness
UNICORN consistently outperformed models like Enformer, Borzoi, and Decima in predicting gene expression at both the single-cell and "pseudo-bulk" levels. The model effectively handles experimental noise; when single-cell data is aggregated into cell types, correlation scores increase significantly, indicating that the model captures the underlying biological signal once experimental noise is reduced.
Multi-Modal Synergy and Personalization
The framework benefits from multi-modal synergy, where predicting RNA expression alongside ATAC-seq or protein levels improves overall accuracy. Additionally, by incorporating individual-specific Whole Genome Sequencing (WGS) data from the GTEx cohort, UNICORN captures personalized expression profiles more effectively than general reference-based models.
Discussion and Conclusion
UNICORN represents a significant step toward universal cellular expression prediction. Its ability to provide uncertainty scores offers a biological filter: genes with low uncertainty are often linked to essential metabolic processes, while high uncertainty identifies noisy or highly variable genes.
However, challenges remain. The model is subject to "reference bias" inherited from its foundation models and is sensitive to the inherent noise of single-cell data. Future work will focus on developing "individual-aware" foundation models and exploring architectures beyond Transformers to further bridge the gap between genotype and cellular phenotype.
Summary updated