EPEPDI

Transcription factors, repair enzymes, and replication machinery all work by gripping DNA, and a single amino acid substitution in the binding interface can loosen or tighten that grip enough to cause disease. Quantifying that shift — the change in binding free energy, ΔΔG, caused by a missense mutation — is the standard way to reason about which variants matter. Structure-based methods can estimate ΔΔG for protein-DNA complexes but require a solved or modeled structure of the complex, which limits how many variants can be screened in practice.

EPEPDI is a sequence-based deep learning framework for this task, developed by Xuan Yu at City University of Hong Kong and posted as a preprint in July 2025. It takes a wild-type protein sequence and a missense mutation and predicts the resulting change in protein-DNA binding free energy, with no structural input required. Rather than training a language model from scratch, EPEPDI draws frozen sequence representations from three existing protein language models — ESM-2, ProtTrans, and ESM-1v — and fuses them through a multi-channel network trained on a curated set of experimentally measured mutations.

Its distinguishing design choice is biological rather than architectural. Double-stranded DNA binding proteins (DSBs) and single-stranded DNA binding proteins (SSBs) engage their substrates through different chemistry: DSBs typically read the major groove of a rigid duplex, while SSBs stack aromatic residues against exposed bases of a flexible strand. Most prior ΔΔG predictors pool the two classes together. EPEPDI treats them as distinct problems, and the paper frames this separation as central to its reported gains.

#Key Features

• Structure-free ΔΔG prediction: Operates on amino acid sequence and the mutation alone, so variants in proteins with no experimentally solved protein-DNA complex can still be scored.
• DSB and SSB separation: Builds a curated dataset that explicitly labels double-stranded versus single-stranded DNA binding proteins and models the two classes separately, rather than assuming one set of features covers both.
• Multi-channel embedding fusion: Combines frozen representations from three pretrained protein language models in parallel channels, letting complementary encodings of the same sequence contribute to a single prediction.
• Entropy-based sequence truncation: An information entropy algorithm identifies 181 residues as the optimal input length, the point where amino acid contribution and entropy dynamics balance, improving computational efficiency alongside accuracy.
• Ablation-validated feature set: The contribution of each embedding channel and feature group is measured directly through ablation rather than assumed.

#Technical Details

EPEPDI is trained once on MPD276, a dataset of experimentally characterized missense mutations in DNA-binding proteins assembled from multiple published sources, and is applied to new mutations without per-target refitting. Each input sequence is encoded independently by ESM-2, ProtTrans, and ESM-1v; the resulting embeddings feed separate channels of a deep network whose outputs are merged to regress a single ΔΔG value. The 181-residue input length, chosen by the entropy analysis, keeps the embedding step tractable enough to screen large mutation sets.

On MPD276 under ten-fold cross-validation, EPEPDI reaches an average Pearson correlation coefficient of 0.755 between predicted and measured ΔΔG. On a held-out independent test set covering both DSBs and SSBs, correlation is 0.632 — a realistic gap between cross-validated and out-of-distribution performance that reflects how small and heterogeneous experimental protein-DNA binding datasets remain. Ablation studies in the paper are used to select the feature combination behind these numbers.

#Applications

The immediate use case is triage: ranking missense variants in DNA-binding proteins by predicted effect on binding, in settings where structures are unavailable or too numerous to model. That includes interpreting variants of uncertain significance in transcription factors and repair proteins, prioritizing candidates for biophysical measurement, and engineering DNA-binding domains toward altered affinity. Because it requires only sequence, it can be applied to SSB families and disordered binding regions that resist structure-based ΔΔG calculation.

#Impact

EPEPDI is a focused predictive head on top of general-purpose protein language models, and its main contribution is the argument that DSBs and SSBs should not be pooled when learning protein-DNA binding energetics — a distinction that generalizes beyond this particular architecture. Its limitations are those of the data: MPD276 is small by machine-learning standards, and the drop from 0.755 to 0.632 on independent tests indicates that generalization to unseen protein families is still constrained. The work remains a preprint awaiting peer review, and no code, trained weights, or web server has been released, so the reported results cannot currently be reproduced or the model applied by third parties.