DPAC: Prediction and Design of Protein-DNA Interactions via Sequence-Based Contrastive Learning

Leo Tianlai Chen, Rishab Pulugurta, Pranay Vure, Pranam ChatterjeebioRxiv 2025 · DOI

TL;DR

This paper introduces DPAC, a novel method that uses contrastive learning on protein and DNA sequences to predict and design their interactions. DPAC significantly speeds up the process compared to structure-based methods and shows promise in designing new proteins with specific DNA-binding capabilities.

What Problem Does This Paper Address?

Proteins that bind to DNA are fundamental to many biological processes, from gene regulation to DNA repair. Designing proteins that can bind to specific DNA sequences with high precision is crucial for advancing genomic technologies. Current state-of-the-art methods for this task heavily rely on 3D structural information. While powerful, these structure-based approaches face limitations in scalability and computational efficiency, especially for large-scale applications. Furthermore, they struggle with modeling conformationally flexible proteins like transcription factors, which are a critical class of DNA-binding proteins. This paper aims to overcome these limitations by developing a purely sequence-based method for predicting and designing protein-DNA interactions.

Methods

DPAC employs a "dual-tower" architecture, similar to CLIP, which leverages pre-trained language models for both proteins (ESM-2) and DNA (Nucleotide Transformer 2 or DNABERT-2). These models generate high-dimensional embeddings for input sequences. The core innovation lies in using a contrastive loss function to align these protein and DNA embeddings into a shared, high-dimensional latent space. This alignment allows the model to learn which protein-DNA pairs are likely to interact.

For designing new protein sequences, DPAC integrates a simulated annealing framework. This process guides the mutation and selection of protein sequences by leveraging the DPAC score (derived from the learned latent space) to optimize for DNA binding affinity. The simulated annealing helps explore the sequence space efficiently, aiming to find sequences that bind strongly to a target DNA.

Datasets

The authors trained and evaluated DPAC using data from the BioLip2 database, a curated collection of protein-DNA binding interactions. After initial filtering and processing, the dataset was split into:

  • Low Identity Set: Composed of 99 complexes, designed to challenge models by ensuring proteins and DNAs interact only within their respective clusters, thus preventing reliance on sequence similarity. This set contained 3,132 non-redundant entries.
  • High Identity Set: Created by randomly sampling 1,000 pairs from the remaining data, representing interactions with higher sequence similarity.
  • Total Training Dataset: 11,419 entries.

For binding affinity benchmarking, a subset of the PDBBind v2020 database was used, excluding NMR structures and focusing on experiments with identical assay conditions.

Key Insights and Ideas

  • Sequence-based is Sufficient: The success of DPAC demonstrates that high-quality prediction and design of protein-DNA interactions can be achieved using only sequence information, bypassing the need for computationally expensive 3D structures.
  • Contrastive Learning for Modality Alignment: Aligning protein and DNA embeddings in a shared latent space via contrastive learning is a powerful strategy for learning cross-modal interactions.
  • Simulated Annealing for Design: Integrating simulated annealing with the learned latent space provides an effective and principled way to search for novel protein sequences with optimized binding properties.
  • Robustness to Sequence Identity: The model shows a degree of robustness to variations in sequence identity, suggesting its potential for predicting interactions between novel or distantly related proteins and DNA.

Key Results

  • Prediction Performance: On the challenging low-identity set, DPAC achieved an AUC score of 0.591, outperforming state-of-the-art structure-based methods.
  • Efficiency: DPAC is significantly faster than structure-based methods. It can screen 2,000 samples in about 2 minutes, whereas RFAA requires approximately 336 hours for the same task. This translates to a massive speedup, making large-scale virtual screening feasible.
  • Design Capability: In silico tests showed that DPAC, combined with simulated annealing, could recover up to 20% of binding affinity in engineered protein sequences that had undergone significant disruption.
  • Binding Affinity Correlation: DPAC logits showed a strong correlation with experimentally measured binding affinities, with negative slopes (indicating correct ranking) in 80.8% of tested groups.

Limitations and Future Work

The authors acknowledge several areas for future improvement:

  • Low Identity Sequences: Enhancing performance on sequences with very low identity remains a challenge.
  • Post-Translational Modifications (PTMs): Incorporating PTMs could further improve DNA binding prediction.
  • Advanced Generative Methods: Exploring guided discrete diffusion and discrete flow matching models could lead to more sophisticated sequence design.

Despite these limitations, DPAC offers a significant advancement in sequence-based protein-DNA interaction modeling and design, paving the way for new applications in genomics and biotechnology.

Summary updated

Abstract

Abstract Interactions between DNA and proteins are pivotal in natural biological processes, and designing proteins that can bind to DNA with high specificity is crucial for advancing genomic technologies. Existing state-of-the-art models for both modeling and designing protein-DNA interactions primarily rely on structural information, facing limitations in scalability and efficiency for large-scale applications. Notable methods like AlphaFold 3 and RosettaTTAFold All-Atom exist, but they are inefficient and inherently struggle at modeling conformationally unstable proteins, such as transcription factors, which arguably represent the most important class of DNA-binding proteins. Here, we present DPAC 1 ( D NA- P rotein binding A lignment via C ontrastive learning), which leverages pre-trained protein and DNA language models via a contrastive loss to align the two modalities in a high-dimensional shared latent space. DPAC not only significantly accelerates the design process compared to current structure-based methods but also demonstrates a strong ability to differentiate real binders from non-binders. Our model achieves an AUC score of 0.591 on a low identity set, outperforming state-of-the-art structure-based methods. Additionally, DPAC integrates simulated annealing for the design of new protein sequences with optimized DNA binding affinity, successfully recovering binding affinity in engineered sequences by up to 20% in in silico tests. Our results highlight DPAC’s potential for facilitating the design and discovery of sequence-specific DNA-binding proteins, paving the way for advancements in genomic research and biotechnology applications.
BibTeX
@misc{chen2025dpac,
  author = {Leo Tianlai Chen and Rishab Pulugurta and Pranay Vure and Pranam Chatterjee},
  title = {{DPAC: Prediction and Design of Protein-DNA Interactions via Sequence-Based Contrastive Learning}},
  year = {2025},
  doi = {10.1101/2025.05.14.654102},
  howpublished = {bioRxiv},
}
Figure