Loading…
Mutation effect prediction at protein–DNA and protein–RNA interfaces, combining frozen ESM-2 embeddings with an edge-aware atomic graph network.
No providers recorded yet. Browse all providers
A mutation that loosens p53's grip on its DNA response element and one that changes how FUS holds RNA are the same measurement — a shift in binding free energy at a protein–nucleic acid interface — but the tools built to predict them were not. Methods here have been split by substrate, fitted separately on a few hundred protein–DNA or protein–RNA mutations apiece, and several require FoldX energy terms computed per mutation before any prediction can be produced.
DeepPNI, from Indian Institute of Technology Bhilai, predicts ΔΔG — the change in binding free energy caused by a single amino acid substitution — for protein–DNA and protein–RNA complexes with one model. Each mutation is read twice. A 10 Å atomic neighbourhood around the mutated residue becomes a heterogeneous graph, encoded by an edge-aware relational graph convolutional network; the surrounding ±200 residues of sequence become a 1280-dimensional embedding from ESM-2. Concatenating the two gives the regressor both the local interface geometry and the evolutionary context that geometry alone omits.
The language model stays frozen: only the graph network and regression head are fitted, keeping trained capacity in proportion to 1754 experimentally measured mutations. What ships is one checkpoint applied as-is to new complexes, in an open Zenodo archive that also carries the code, the full curated training set and the five-fold cross-validation splits, alongside a public web server.
Nodes carry a 12-dimensional feature vector — atomic number, degree, formal charge, aromaticity, implicit hydrogen count, and a one-hot hybridization type — projected to 128 dimensions and passed through four RGCN layers using basis decomposition, batch normalization, dropout and residual connections; mean pooling yields a 128-dimensional graph embedding. That is concatenated with the mean-pooled 1280-dimensional ESM-2 embedding and pushed through a 512–256–64 MLP to a single unbounded scalar. The chosen MLP carries 8.74 × 10⁵ trainable parameters, the capacity that minimized the generalization gap in a sweep from 3.94 × 10⁵ to 1.97 × 10⁶. Training data are 1754 mutations across 439 PDB structures curated from the NABE database, after averaging duplicates and trimming outliers at a Z-score threshold of 1.5.
Five-fold cross-validation gives a Pearson correlation of 0.76 ± 0.03, MAE 0.52 ± 0.03 and RMSE 0.65 ± 0.03. Ablation places most of that signal in the language model — ESM-2 alone reaches 0.73, the graph alone 0.52 — with the combination improving both metrics. Across four blind ProNAB sets grouped by assay type, correlations run from 0.58 (filter binding) to 0.81 (gel shift). The most conservative figure is the structure-based split, where no PDB entry is shared between training and test folds: correlation falls to 0.47 ± 0.05, against 0.29 ± 0.06 for DeePNAP under the same split.
DeepPNI is a triage tool for mutations in structurally characterized complexes: ranking variants of uncertain significance in transcription factors and RNA-binding proteins, narrowing saturation-mutagenesis panels before committing to calorimetry or gel-shift assays, and checking engineered substitutions in nucleases and other nucleic-acid-binding enzymes. Input is a mutation string, PDB identifier and chain, and 100 mutations finish in under a minute on the deployed hardware — the scale at which deep-mutational-scanning and clinical variant datasets arrive.
The field of ΔΔG predictors for protein–nucleic acid interfaces is small and mostly predates protein language models; the useful result is that a frozen sequence encoder plus an interface-local graph beats hand-built energy features without substrate-specific models or licensed force-field software. The limits are equally plain. Output is one scalar for one single-point substitution — no multi-mutation, insertion or deletion support, no per-atom attribution, and no exposed embeddings for downstream reuse. Accuracy on wholly unseen scaffolds is modest, and the 0.90 correlation reported on the PremPDI benchmark reflects an overlap of 193 of its 219 mutations with the training set, which the authors flag themselves. Licensing is inconsistent across surfaces: the code archive carries an MIT licence, its Zenodo record is labelled CC BY 4.0, the journal article is CC BY-NC 4.0, and the arXiv preprint is stricter still at CC BY-NC-ND 4.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.