DBP-CanPred: a machine learning model for predicting cancer-causing mutations in DNA-binding proteins

Interpretable prediction of nucleic acid-binding proteins using a protein language model

Abstract Motivation Accurate identification of DNA-binding proteins (DBPs) and RNA-binding proteins (RBPs) is critical for elucidating transcriptional and post-transcriptional regulatory mechanisms. However, existing computational approaches often rely on inferred labels or domain-specific annotations, which limit the subsequent generalizability. Results This study aimed to introduce transformer-based classifiers for human DBPs and RBPs that rely solely on protein sequence information without engineered features or domain constraints. The models were implemented using ESM-2 with low-rank adaptation (LoRA) fine-tuning and trained on experimentally validated datasets, including chromatin immunoprecipitation sequencing (ChIP-seq) annotations for DBPs and eCLIP annotations for RBPs. Next, to evaluate biological relevance, we computed value-aware attention (VAT) scores aggregated across transformer layers to interpret model focus. In 20-fold cross-validation, the DBP model achieved an area under the receiver operating characteristic curve (AUROC) of 0.84 with a Matthews correlation coefficient (MCC) of 0.40, while the RBP model achieved an AUROC of 0.92 with an MCC of 0.46. Proteins predicted as nucleic acid-binding were enriched for known binding domains, and inspection of attention distributions revealed preferential focus on annotated functional regions rather than non-binding segments. These results demonstrate that attention-based protein language models can accurately identify nucleic acid-binding proteins directly from sequence data. Moreover, these models reveal biologically meaningful sequence determinants of binding, establishing an interpretable and scalable framework for proteome-wide characterization of protein–nucleic acid interactions. Availability and implementation Code is available on GitHub (https://github.com/CSB-hub/DRBP).

Genomic language model for predicting enhancers and their allele-specific activity in the human genome

Abstract Motivation Predicting and deciphering the regulatory logic of enhancers remains a significant challenge due to their complex sequence features and the absence of consistent genetic or epigenetic signatures that distinguish them from other genomic regions. Existing machine learning methods capture nucleotide composition but often fail to model sequence context effectively. Results We present DNABERT-Enhancer, a novel enhancer prediction method, by applying DNABERT pre-trained language model on the human genome. Using ENCODE registry of candidate cis-regulatory elements (cCREs), we curated a benchmark dataset, consisting of 21 926 enhancers of 201 bp length and 46 159 enhancers of 350 bp length, as positive instances. The best fine-tuned model achieved 88.05% accuracy and a Matthews correlation coefficient of 76% on an independent dataset. Genome-wide application identified 1 684 595 enhancer regions covering 26.65% of the human genome. By performing integrative analyses with DNABERT-based transcription factor models, we identify 2681 statistically significant loss-of-function and 1917 gain-of-function enhancer variants, which respectively alter the function of 1623 and 1247 ENCODE-cCRE enhancers. Similarly, we identify 4057 candidate de novo enhancers, created by 5464 gain-of-function variants. These genome-wide enhancer annotations and candidate genetic variants predicted by DNABERT-Enhancer provide valuable resources for genome interpretation in functional and clinical genomics studies. Availability and implementation DNABERT-Enhancer is freely available at https://github.com/DavuluriLab/DNABERT-Enhancer; Trained model predictions can be explored interactively via the web application at https://dnabert-enhancer-datarepo.streamlit.app/. The fine-tuned models are archived and citable through Zenodo (https://doi.org/10.5281/zenodo.19157566).

DeepPNI: a language- and graph-based model for mutation-driven protein–nucleic acid binding energetics

DeepPNI is a deep learning regression model that integrates sequence- and structure-based features to estimate mutation-induced changes in binding free energy in protein–nucleic acid complexes, developed using a comprehensive dataset of 1754 mutations spanning protein–DNA and protein–RNA complexes.

Bidirectional network hubs: NT-genes as candidate targets for partial cancer reversal

Reversing the tumor phenotype is a long-standing challenge in cancer biology. Although GRNs comprise thousands of genes, normal and tumor tissues occupy distinct, low-dimensional attractors in expression space, raising the possibility that targeting a few key genes could induce widespread transcriptional changes. We build on two previously developed concepts: 1) N- and T-markers (genes with exclusive expression intervals in normal or tumor samples, respectively), and 2) Gene Deregulation Networks (GDNs) – directed acyclic graphs inferred from expression data, in which a link from C to E indicates that a deregulation at C increases the probability of a deregulation at E. A subset of N and T-genes, namely, NT-markers, appear in both networks and may act as bridges for phenotype reversal. Using TCGA bulk RNA-Seq data from five cancer types we identified N-, T- and NT-genes based on statistically significant exclusive expression intervals. Discretized expression states (N-active, T-active, inactive) are then associated with normal-exclusive, tumor-exclusive and non-exclusive expression intervals. GDNs were constructed with the CChains algorithm using the Loevinger coefficient, followed by Reichenbach (common cause) and Mokken (transitivity) pruning. We introduced a quantitative model to predict intervention outcomes in a tumor using gene activation frequencies and two topological metrics: the composed coverage and the fraction of the T-network unreached by the reverse deactivation cascade (U). We also use a Glauber-like dynamical model to simulate interventions. We predict that pure T-gene interventions mainly alter the T-network, while pure N-gene interventions create a mixed normal-tumor state. In contrast, high-frequency NT-genes are predicted to simultaneously deactivate T-cascades and reactivate N-programs. The unreached fraction U varies across cancers: for perfect diagnostic panels, U ≈ 30% in PRAD but only ≈6% in LUAD. Escape probability also depends on tumor stage and on the spontaneous activation rate. NT-genes with high dual frequency are predicted optimal targets for partial phenotype reversal. The combination of composed coverage, unreached fraction and normal-state relevance provides a quantitative guide for designing multi-target therapies. The framework is general and yields testable predictions for intervention outcomes across cancer types.

Graph Neural Networks for Cancer Driver Gene Prediction: From Fundamentals to Applications.

Non-coding regulatory variants in breast cancer: Current prediction models and experimental gaps