"Frustratingly Easy" Domain Adaptation for Cross-Species Transcription Factor Binding Prediction

Mark Maher Ebeid, Ali Tugrul Balci, Maria Chikina, Panayiotis V. Benos, Dennis Kostka

TL;DR

Researchers have developed MORALE, a "frustratingly easy" framework for predicting where proteins bind to DNA across different species. By simply aligning the statistical moments (mean and covariance) of internal model representations, MORALE outperforms complex adversarial methods and can even beat models trained on a single species' own data.

What Problem Does This Paper Address?

Deep learning has revolutionized our ability to predict how DNA sequences regulate genes—specifically, where Transcription Factors (TFs) will bind. However, these models suffer from a "domain" problem: a model trained on human data often fails when applied to a mouse, and vice versa.

The challenge is that while the fundamental "biochemical rules" of TF binding are often conserved across millions of years of evolution, the "background noise" of the genome (like repetitive elements and varying GC content) changes rapidly. Standard models often overfit to these species-specific quirks rather than learning the universal regulatory grammar. Bridging this gap usually requires complex adversarial training, which is notoriously difficult to tune and can sometimes make model performance worse.

Methods

The authors introduce MORALE (Moment Alignment for Regulatory Logic Evaluation). Instead of using a complex "adversarial" branch to hide species identity from the model, MORALE uses a much simpler approach:

  1. Embedding Layer: DNA sequences are passed through a standard architecture (Convolutions + LSTM/GRU) to create a latent representation (an embedding).
  2. Moment Alignment: The model calculates the first moment (mean) and second moment (covariance) of these embeddings for both the source species (e.g., mouse) and the target species (e.g., human).
  3. The Loss Function: A penalty is added to the training process if the statistical moments of the two species don't match. This forces the model to find a "species-invariant" way to represent DNA.

This approach is "frustratingly easy" because it requires zero extra parameters and can be expressed in a closed mathematical form, unlike adversarial methods like Gradient Reversal Layers (GRL), which can add millions of parameters to a network.

Datasets

The study evaluated MORALE on two primary setups using ChIP-seq data (which identifies protein-DNA binding sites) from liver tissue:

  • Two-Species Case: Human and Mouse data for four TFs (CTCF, HNF4α, RXRA, and CEBPA).
  • Multi-Species Case: Data from five mammals—Human, rhesus macaque, mouse, rat, and dog—for four TFs (CEBPA, FOXA1, HNF4α, and HNF6).
  • Genomic Context: The authors specifically looked at performance within repetitive DNA elements (SINEs and LINEs), which are often the hardest regions to generalize across species.

Key Insights and Ideas

  • Simplicity Wins: MORALE consistently outperformed or matched the much more complex GRL adversarial approach. Crucially, while GRL sometimes caused "performance degradation" (performing worse than a model with no adaptation at all), MORALE remained stable.
  • The "More is Better" Effect: In the five-species test, a model trained with MORALE actually outperformed a model trained only on human data when predicting human binding sites. This suggests that by looking across many species, the model learns a "cleaner" version of the regulatory code that isn't visible in a single genome.
  • Better Interpretability: Using feature attribution (TF-MoDISCo), the researchers found that MORALE discovered DNA motifs that were much closer to the "ground truth" biological motifs than those found by adversarial models.
  • Repeat Performance: MORALE was particularly effective at handling species-specific repeats (like Alus in humans), which usually act as a major source of error in cross-species transfer.

Key Results

  • Performance Boost: In the two-species human-to-mouse transfer, MORALE achieved an average auPRC (Area Under Precision-Recall Curve) of 0.332, compared to 0.304 for the adversarial GRL and 0.31 for the baseline.
  • Human Superiority: In the five-species case, MORALE reached an auPRC of ~0.35 for CEBPA on human data, significantly higher than the ~0.22 achieved by a model trained only on human data.
  • Efficiency: MORALE achieved these results while keeping the model size at roughly 613k parameters, whereas the GRL approach ballooned the model to over 9.4 million parameters due to the added discriminator branch.

Limitations and Future Work

While MORALE is highly effective, the authors note a few areas for further exploration:

  • Tissue Specificity: This study focused exclusively on liver-essential TFs. It remains to be seen if the same "frustratingly easy" alignment works as well for TFs with more complex, multi-tissue binding patterns.
  • Architecture Agnosticism: While the authors tested CNN/RNN hybrids, future work could integrate MORALE into larger Transformer-based "foundation models" for genomics.
  • Causal Modeling: The authors suggest that domain adaptation is a necessary step toward building "causal" models of gene regulation that can truly predict the effects of mutations across different evolutionary backgrounds.

Summary updated

Abstract

Abstract Motivation Understanding how DNA sequence encodes gene regulation remains a central challenge in genomics. While deep learning models can predict regulatory activity from sequence with high accuracy, their generalizability across species—and thus their ability to capture fundamental biological principles—remains limited. Cross-species prediction provides a powerful test of model robustness and offers a window into conserved regulatory logic, but effectively bridging species-specific genomic differences remains a major barrier. Results We present MORALE, a novel and scalable domain adaptation framework that significantly advances cross-species prediction of transcription factor (TF) binding. By aligning statistical moments of sequence embeddings across species, MORALE enables deep learning models to learn species-invariant regulatory features without requiring adversarial training or complex architectures. Applied to multi-species TF ChIP-seq datasets, MORALE achieves state-of-the-art performance—outperforming both baseline and adversarial approaches across all TFs—while preserving model interpretability and recovering canonical motifs with greater precision. In the five-species transfer setting, MORALE not only improves human prediction accuracy beyond human-only training but also reveals regulatory features conserved across mammals. These results highlight the potential of simple yet powerful domain adaptation techniques to drive generalization and discovery in regulatory genomics. Crucially, MORALE is architecture-agnostic and can be seamlessly integrated into any embedding-based sequence model. Availability Code is available at https://github.com/loudrxiv/frustrating .
BibTeX
@inproceedings{ebeid2026frustratingly,
  author = {Mark Maher Ebeid and Ali Tugrul Balci and Maria Chikina and Panayiotis V. Benos and Dennis Kostka},
  title = {{"Frustratingly Easy" Domain Adaptation for Cross-Species Transcription Factor Binding Prediction}},
  booktitle = {RECOMB-CG},
  year = {2026},
  doi = {10.1007/978-3-032-26891-4\_19},
}
Figure