Machine Learning Predicts RNA-Protein Interactions Using Language Embeddings and Sequence Features

Computational biologists at the Indraprastha Institute of Information Technology in New Delhi have developed a sequence-based machine learning framework that pairs dinucleotide autocorrelation descriptors with large protein language model representations to predict interactions between long non-coding RNAs and proteins. Training on a non-redundant benchmark derived from the NPInter v5.0 database comprising 103,549 interacting pairs and an equal number of non-interacting pairs, the resulting CatBoost model achieved an area under the receiver operating characteristic curve of 0.929 on the training data and 0.692 when evaluated against an independent, homology-aware validation set containing 24,825 positive and 24,825 negative pairs. By integrating Dinucleotide Autocorrelation features for RNA molecules with ESM-2-t30 embeddings for protein sequences, the computational system addresses persistent generalization challenges in molecular recognition. Alongside the primary predictive architecture, the investigators constructed an end-to-end two-tower convolutional neural network and released a publicly accessible web server and standalone software suite named lncrnaPI to facilitate proteome-wide and genome-wide interaction scanning from raw sequence files.

Long non-coding RNAs represent a distinct class of non-coding RNA transcripts exceeding 200 nucleotides in length that exhibit limited protein-coding potential yet serve as indispensable regulators across diverse cellular mechanisms. In living systems, these molecular transcripts carry out their biological functions predominantly by binding specific proteins to assemble functional ribonucleoprotein complexes. Well-characterized interactions illustrate this regulatory paradigm: the non-coding RNA XIST coordinates mammalian X-chromosome inactivation by recruiting Polycomb Repressive Complex 2, whereas the transcript MALAT1 associates with serine and arginine-rich proteins to modulate alternative splicing pathways. In addition to regulating transcription and chromatin remodeling, these molecular interactions govern intracellular signaling and organize subcellular compartments. When such associations undergo aberrant dysregulation, they are frequently implicated in severe human pathologies, including oncogenesis and neurodegenerative disorders. Understanding the global interactome that connects RNA molecules to protein targets is therefore critical for clarifying basic cellular biology and establishing foundations for targeted therapeutic interventions.

Historically, researchers have relied on laboratory experimental methods to map ribonucleoprotein interactions across cellular systems. High-throughput immunoprecipitation techniques, including traditional RNA immunoprecipitation and ultraviolet cross-linking immunoprecipitation variants such as PAR-CLIP and iCLIP, have successfully cataloged thousands of physical binding events. Despite their foundational importance in generating high-confidence interaction maps, these laboratory assays present operational obstacles: they remain labor-intensive, require substantial quantities of cellular starting material, and incur prohibitive experimental costs that limit exhaustive genome-scale investigations. To alleviate these experimental constraints, computational biologists previously developed algorithmic predictors based on classical machine learning classifiers, utilizing handcrafted features such as k-mer frequencies, secondary structures, and physicochemical properties. Subsequent methodologies incorporated heterogeneous biological network topologies or ensemble learning configurations, while contemporary efforts have applied graph convolutional networks, graph autoencoders, and graph transformers. However, a major impediment among existing predictive tools has been their reliance on outdated interaction databases such as NPInter v2.0, limited software availability, or dependencies on auxiliary data such as gene expression profiles and interaction networks.

To overcome the architectural and data limitations of prior computational systems, the investigators developed a multi-stage filtering protocol starting from 2,596,183 experimentally validated interactions in NPInter v5.0. Restricting the curation to Homo sapiens long non-coding RNA and protein pairs yielded an initial set of 488,982 records. RNA sequences were filtered to retain transcripts between 200 and 20,000 nucleotides, while protein records containing non-standard amino acid codes were excluded to preserve downstream extraction consistency, refining the curated set to 370,242 pairs. To prevent sequence redundancy from creating inflated performance evaluations, the team executed a strict homology-aware partitioning procedure. Functional similarity was calculated using the Jaccard similarity index across Pfam and Gene Ontology annotations derived via InterProScan, classifying proteins sharing a Jaccard index of 0.3 or greater as functionally similar, with an analogous Rfam analysis applied to RNA transcripts. Sequences were clustered using CD-HIT at an 80 percent identity threshold for RNA and 40 percent for proteins, assigning entire clusters exclusively to training or validation pools to ensure that similar sequences never crossed the split.

Because interaction pairs were retained in the validation partition only when both interacting molecular partners were independently assigned to the validation set, any interactions whose partners fell across different subsets were discarded to eliminate leakage. This procedure yielded 103,549 positive training pairs and 24,825 positive validation pairs. To generate balanced evaluation sets, non-interacting negative pairs were constructed by randomly pairing entities strictly within their respective training or validation partitions, excluding any combinations observed in the positive database. This generated matching negative counts of 103,549 training pairs and 24,825 validation pairs, establishing a total benchmark of 256,748 pairs. The investigators extracted diverse feature spaces, calculating k-mer distributions and dinucleotide correlation metrics using Nfeature alongside amino acid and dipeptide compositions from Pfeature. Protein evolutionary conservation was captured through Position-Specific Scoring Matrices constructed using the POSSUM toolkit and PSI-BLAST against SwissProt. Contextual representations were generated using modern transformer models, applying the bidirectional DNABERT-2 model to RNA sequences and extracting embeddings from pre-trained ESM-2 models, specifically the 30-layer ESM2-t30 and 33-layer ESM2-t33 architectures.

Before evaluating advanced machine learning architectures, the authors implemented classical alignment baselines using the Basic Local Alignment Search Tool to evaluate whether sequence homology alone could resolve molecular interaction pairs. Querying validation sequences against training databases demonstrated that most validation pairs generated no significant alignment hits due to substantial sequence divergence. For the subset of pairs producing detectable alignments, a best-match strategy correctly predicted interaction status for 10,236 of 16,547 queries to achieve 61.68 percent accuracy, while a voting-based consensus approach classified 9,908 of 15,497 pairs correctly for 63.93 percent accuracy. In parallel, preliminary sequence analyses demonstrated that protein amino acid composition provided a modest discriminative signal: leucine was depleted in interacting proteins relative to non-interacting proteins (6.76 percent versus 7.88 percent; 60 percent accuracy), whereas glycine was enriched (10.03 percent versus 8.49 percent; 59 percent accuracy). In contrast, nucleotide compositions in RNA sequences exhibited near-chance predictive accuracies between 50 and 51 percent, indicating that simple nucleotide frequencies cannot distinguish functional interaction partners.

To build robust predictive models, the researchers trained several parallelized algorithms, including Random Forest, Extra Trees, XGBoost, and CatBoost. Systematic hyperparameter tuning of the CatBoost model was conducted with the Optuna framework across 100 trials using five-fold stratified cross-validation on the training set, optimizing tree depth, iterations, learning rate, and regularization metrics against the area under the receiver operating characteristic curve. The resulting CatBoost classifier integrating Dinucleotide Autocorrelation features with ESM2-t30 protein embeddings delivered the strongest overall predictive capacity, reaching a training accuracy of 0.856, an F1 score of 0.714, a Matthews correlation coefficient of 0.862, and a training area under the curve of 0.929. When applied to the homology-aware validation set, this model achieved an accuracy of 0.564, an F1 score of 0.163, a Matthews correlation coefficient of 0.368, and a validation area under the curve of 0.692. Alternative combinations using Nucleotide Repeat Index or Distance Distribution features paired with ESM2-t30 reached slightly lower validation areas under the curve of 0.684 and 0.683, respectively.

In parallel with feature-engineered classifiers, the research team constructed an end-to-end deep learning framework based on a two-tower convolutional neural network that processes integer-encoded sequences through low-dimensional embedding layers without manual feature engineering. The RNA tower maps nucleotides to four-dimensional embeddings before processing through two sequential convolutional blocks containing 128 and 256 filters, respectively, with max-pooling and dropout. The protein tower maps amino acids to twenty-dimensional embeddings followed by analogous convolutional blocks with 128 and 256 filters. Global max pooling extracts fixed-dimensional vectors from both towers, which are concatenated and routed into dense layers with ReLU activation and a final sigmoid classification unit. To resolve optimal input lengths, the investigators conducted percentile ablation across training sequence distributions. Truncating sequences at the 99th percentile—corresponding to 12,913 nucleotides for RNA and 2,390 amino acids for proteins—yielded the best validation performance, producing an area under the curve of 0.674, a Matthews correlation coefficient of 0.248, an accuracy of 0.618, and an F1-score of 0.547 while retaining 99.0 percent of RNA and 98.9 percent of protein sequence lengths.

To investigate the biological mechanisms informing model decisions, the authors analyzed validation data across CatBoost variants using Shapley Additive Explanations. In a hybrid model pairing Dinucleotide Autocorrelation with structural property descriptors, protein physical characteristics accounted for 54.1 percent of overall feature importance, while RNA dinucleotide metrics contributed 45.9 percent. Structural coils and intrinsically disordered protein regions exerted the strongest predictive impact, showing a mean absolute Shapley value of 0.49, followed by hydrophobic character at 0.44, buried surface area at 0.43, and hydroxylic amino acid content at 0.42. Beeswarm distributions confirmed that high values for these flexible and exposed protein properties consistently drove positive interaction calls. Analysis of RNA dinucleotide features revealed that the model relies on base-step geometry parameters including Rise, Tilt, and Twist, alongside cross-covariance combinations restricted to narrow windows of three to four nucleotides. When analyzing deep contextual language model embeddings, the decision landscape inverted, with DNABERT-2 RNA embeddings driving 65.8 percent of predictions compared to 34.2 percent from ESM-2 protein embeddings.

Evaluating the optimized lncrnaPI model against six functional, sequence-input baseline tools on the identical validation benchmark demonstrated distinct operating characteristics across existing methodologies. Several historical predictors exhibited extreme classification skews: LPI-BLS demonstrated 0.985 sensitivity but only 0.036 specificity (area under the curve of 0.589), whereas LPI-MFF displayed 1.000 specificity but only 0.001 sensitivity (area under the curve of 0.589). Baseline tools AR-LPI, FMSRT-LPI, GHCT, and MFH-LPI recorded validation areas under the curve of 0.651, 0.500, 0.645, and 0.625, respectively. In contrast, lncrnaPI achieved the highest overall validation area under the curve of 0.692 and the highest Matthews correlation coefficient of 0.163, alongside 0.260 sensitivity, 0.867 specificity, 0.672 precision, and 0.564 accuracy. To support rapid high-throughput screening where extracting large language model embeddings is computationally prohibitive, the team established a lightweight classifier based on ten selected amino acids and four nucleotide frequencies that achieved a validation area under the curve of 0.601.

Despite these computational advances, the researchers documented several methodological limitations that temper interpretation of the findings. Because experimentally confirmed non-interacting pairs are unavailable in biological databases, generating negative examples through partner shuffling introduces the risk of including undiscovered, latent biological interactions. Furthermore, while the homology-aware splitting protocol minimized obvious sequence redundancy and shared protein families, it did not impose an absolute cold-start partition, meaning residual evolutionary relationships could still persist between training and validation sets. Truncating long sequences at the 99th percentile in convolutional architectures may also remove functionally important terminal domains or regulatory motifs in exceptionally long biomolecules. Finally, the authors emphasized that the models operate purely as sequence-based computational estimators and do not explicitly simulate physiological context, RNA tertiary folding, protein conformational transitions, local molecular abundance, or binding competition within intact living cells, necessitating downstream laboratory validation before confirming novel predicted interactions.

Subject of Research: Sequence-based machine learning framework for predicting long non-coding RNA and protein interactions

Article Title: Prediction of lncRNA-protein interacting pairs using LLM embeddings based on evolutionary information

Article References: Choudhury, S., Bajiya, N., & Raghava, G. P. S. (2026). Prediction of lncRNA-protein interacting pairs using LLM embeddings based on evolutionary information. Frontiers in Bioinformatics, 6, 1837747. https://doi.org/10.3389/fbinf.2026.1837747

Image Credits: AI Generated

DOI: 10.3389/fbinf.2026.1837747

Keywords: lncRNA-protein interactions, machine learning, CatBoost, protein language models, ESM-2 embeddings, DNABERT-2, dinucleotide autocorrelation, two-tower convolutional neural network, NPInter database, homology-aware validation, SHAP interpretability, bioinformatics tools, lncrnaPI web server, ribonucleoprotein complexes, sequence-based prediction, structural bioinformatics, computational biology