Design and training of integrated seq2func–transcriptomic systems

Construct and train a bidirectional system that integrates sequence-to-function models conditioned on cellular embeddings with single-cell transcriptomic foundation models, enabling joint reasoning about cis-regulatory changes, propagation to cell-wide state, and feedback to refine sequence predictions across diverse cellular contexts.

Background

Generalization across cellular contexts is limited when changes in trans factors and chromatin state alter the sequence–function mapping. The authors propose a bidirectional interface between seq2func models and transcriptomic foundation models to jointly reason about cis- and trans-regulatory dynamics under perturbation. Realizing this integration requires novel conditioning mechanisms, training procedures, and evaluation frameworks.

References

Designing and training such integrated systems remains an open challenge.

— Toward Interpretable and Generalizable AI in Regulatory Genomics  (2602.01230 - Nagai et al., 1 Feb 2026) in Section “Generalization Across Cellular Contexts”