Biology subjects

Goteti, A.

Publications and source records attributed to Goteti, A..

4 recordsLinked to original sources

bioRxiv2025

AstraBIND: Graph Attention Network for Predicting Ligand Binding Sites

Predicting ligand binding sites is central to computational biology and drug discovery. Existing machine learning approaches either use protein sequence, structure, or both. While structure-based deep learning models typically outperform sequence-based methods, they often require high computational cost or ligand-specific data, forcing a trade-off between accuracy and scalability. We present AstraBIND, a lightweight graph neural network that bridges this gap by integrating protein sequence, structure (experimental or predicted), and homology information to predict ligand classes and binding residues within minutes. The model employs a GATv2 architecture with 0.9 M parameters, trained on {inverted question}250 000 curated protein-ligand complexes across 16 ligand categories. By encoding residue-level features and spatial geometry through graph attention, AstraBIND identifies binding residues and ligand types while maintaining structural consistency. In benchmarking, AstraBIND achieved a weighted macro-F1 of 0.47 across all ligand classes, with top performance for nucleotides (F1 = 0.79), porphyrins (0.74), and cofactors (0.73). Case studies, including p53 and CRFR1, demonstrate robust pocket localization for diverse proteins. Combined with its minimal inference time and broad ligand coverage, AstraBIND enables rapid in-silico screening and integration into laboratory workflows. Together with other Astra ML models (1; 2), it represents a step toward real-time protein design and validation pipelines. Astra models are available at https://www.orbion.life.

bioinformatics↗

bioRxiv2025

AstraPTM2: A Context-Aware Transformer for Broad-Spectrum PTM Prediction

Post-translational modifications (PTMs) are covalent changes in proteins after biosynthesis that shape stability, localization, and function. While numerous computational tools exist for PTM site prediction, most struggle to handle full-length proteins without truncating context, focus on only a limited number of PTM types, and perform unevenly on rare modifications. We present AstraPTM2, a transformer-based model that predicts 39 distinct PTM types on full-length sequences. By combining ESM-2 embeddings, AlphaFold2-derived structural features, and protein-level descriptors, AstraPTM2 captures both short-range motifs and long-range dependencies. Training uses a three-stage curriculum and adaptive focal loss to balance rare and common PTMs, followed by per-label affine calibration and optimized thresholds for well-calibrated probabilities. In hold-out tests, AstraPTM2 achieves AUROC = 0.99 and macro-F1 = 59% across 39 PTM types, with particularly strong performance on rare motif-driven PTMs such as O-linked glycosylation and sumoylation. Results are available through the Orbion web platform, which offers synchronized 2D and 3D visualizations, dual prediction modes (calibrated and exploratory), and reproducible exports to support downstream experimental planning. AstraPTM2 can be accessed at https://www.orbion.life.

bioinformatics↗

bioRxiv2025

AstraROLE & AstraSUIT: Multi-Task Annotation Models for Functional Profiling of Proteins

AbstractMost in-silico protein characterisation tools focus on only one aspect of protein function, forcing researchers to use multiple models or to bypass computational checks. Here we introduce AO_SCPLOWSTRAC_SCPLOWROLE2 and AO_SCPLOWSTRAC_SCPLOWSUIT2, two transformer-based, multi-task annotators that deliver an integrated functional profile in a single pass. A 1,351-dimensional input (ESM-2 CLS embeddings plus physicochemical Orbion enrichments) is mapped by a 512-unit encoder and task-specific linear heads: four in AO_SCPLOWSTRAC_SCPLOWROLE2 (EC class, GO term, molecular pathway, protein category) and nine in AO_SCPLOWSTRAC_SCPLOWSUIT2 (cofactor group, specific cofactor, domain, host, membrane association type, transmembrane helix number, subcellular localization, quaternary category, quaternary stoichiometry). Models were trained on 730k UniProt proteins with stratified 70/15/15 splits; class-weighted BCE and Optuna hyper-parameter search countered imbalance. On hold-out sets the heads reached macro F1=0.84-0.98 and MCC=0.85-0.98. Highest scores were seen for cofactor binding (0.98), membrane association type (F1=0.97) and top-level EC number (0.96); GO term classification was hardest (0.85). Against recent comparators (incl. DeepGOPlus and TargetP 2.0), the Astra models matched or exceeded performance, especially on metal-ion binding and cofactor binding. Additional tests on three novel proteins not included in initial dataset showed good predictions for most labels, underscoring the potential for hypothesis generation. Overall, AO_SCPLOWSTRAC_SCPLOWROLE2 and AO_SCPLOWSTRAC_SCPLOWSUIT2 supplied fast, state-of-the-art multi-label protein annotation within one unified model network.

bioinformatics↗

bioRxiv2025

AstraPTM: Context-Aware PTM Prediction Model for Large-Scale Proteins

Post-translational modifications (PTMs) are critical molecular events that pro-foundly influence protein stability, localization, and function. While numerous computational tools exist for PTM site prediction, most struggle with handling large proteins and rely on separate models for each modification. To address these challenges, we introduce AstraPTM, a transformer-based framework that predicts 25 PTMs in a single pass. By leveraging advanced protein embeddings (ESM2) and training on a high-coverage dbPTM dataset, AstraPTM captures both short-range sequence motifs and long-range interactions across proteins without a sequence length limitation. AstraPTM combines a binary classification module--indicating whether a residue is modified--with a multi-label module that pinpoints specific PTM types. This dual approach achieves high accuracy on well-represented PTMs (e.g., phospho-rylation, glycosylation) while maintaining sensitivity for rarer modifications. In benchmarks against existing methods such as MusiteDeep and MIND-S, AstraPTM demonstrates competitive or superior performance, demonstrating AUC-ROC above 99% for well-represented modifications, underscoring its versatility for proteome-wide annotation. Beyond prediction, the models capacity to handle full-length proteins offers a powerful resource for researchers investigating PTM crosstalk and disease pathways, ultimately bridging the gap between large-scale omics data and targeted biomedical applications.

bioinformatics↗