Biology subjects

Soto-Garcia, N.

Publications and source records attributed to Soto-Garcia, N..

4 recordsLinked to original sources

bioRxiv2026

Data-Centric Evaluation of Protein Function Prediction Pipelines

Performance estimates in protein function prediction depend not only on model choice but also on upstream decisions that define the learning problem. Using antioxidant protein classification as a controlled case study, we evaluated how dataset harmonisation, protein representation, redundancy control, and partitioning strategy affect protein machine learning pipelines. We integrated 18,804 records from 12 publicly available dataset entries into a curated consensus dataset of 4,193 protein sequences. One-hot encoding and six pretrained protein language model representations were evaluated as model inputs and as similarity spaces for redundancy reduction and distance-aware splitting. Representation choice substantially altered dataset geometry, retained dataset size, class balance, and downstream evaluation. At representation-specific p90 thresholds, one-hot encoding retained the complete dataset, whereas pretrained embeddings retained between 5% and 25% of sequences. Distance-aware partitioning reduced apparent performance relative to random splitting by up to 0.15 MCC before redundancy control, while this difference narrowed after similarity filtering. Selected configurations nevertheless maintained high performance under stricter evaluation, reaching an MCC of 0.84. These findings show that performance estimates should be interpreted as outcomes of complete data-centric workflows rather than isolated properties of predictive models.

bioinformatics↗

bioRxiv2026

MAOMAO: An Ontology-Guided FAIR Resource for Harmonized Peptide Toxicity Data

Peptide toxicity is a critical safety and developability parameter in peptide discovery and therapeutic development, yet relevant information remains fragmented across databases, literature resources, and curated datasets. Here, we present MAOMAO, an ontology-guided FAIR-oriented resource that integrates and harmonizes peptide toxicity data from 54 sources. MAOMAO contains 71,701 unique peptide sequences across seven toxicity-related endpoints, represented as 501,907 sequence endpoint combinations with endpoint-specific evidence states and explicit encoding of unavailable information. The resource combines standardized terminology, a hierarchical toxicity vocabulary, evidence-aware state resolution, provenance-aware metadata, 41 physicochemical descriptors, 10 protein language model representations, and a one-hot baseline. It provides endpoint-specific benchmark partitions across splitting strategies and random seeds, reusable with numerical representations. MAOMAO establishes a reusable framework for peptide toxicity research and data-driven toxicology.

pharmacology and toxicology↗

bioRxiv2025

Protein language models accelerate the discovery of Plastic-Degrading Enzymes

Plastic pollution presents a critical environmental challenge, necessitating innovative and sustainable solutions. In this context, biodegradation using microorganisms and enzymes offers an environmentally friendly alternative. This work introduces an AI-driven frame-work that integrates machine learning (ML) and generative models to accelerate the discovery and design of plastic-degrading enzymes. By leveraging pre-trained protein language models and curated datasets, we developed seven ML-based binary classification models to identify enzymes targeting specific plastic substrates, achieving an average accuracy of 89%. The framework was applied to over 6,000 enzyme sequences from the RemeDB to classify enzymes targeting diverse plastics, including PET, PLA, and Nylon. Besides, generative learning strategies combined with trained classification models in this work were applied for de novo generation of PET-degrading enzymes. Structural bioinformatics validated potential candidates through in-silico analysis, highlighting differences in physicochemical properties between generated and experimentally validated enzymes. Moreover, generated sequences exhibited lower molecular weights and higher aliphatic indices, features that may enhance interactions with hydrophobic plastic substrates. These findings highlight the utility of AI-based approaches in enzyme discovery, providing a scalable and efficient tool for addressing plastic pollution. Future work will focus on experimental validation of promising candidates and further refinement of generative strategies to optimize enzymatic performance.

bioengineering↗

bioRxiv2024

RUDEUS, a machine learning classification system to study DNA-Binding proteins

DNA-binding proteins are essential in different biological processes, including DNA replication, transcription, packaging, and chromatin remodelling. Exploring their characteristics and functions has become relevant in diverse scientific domains. Computational biology and bioinformatics have assisted in studying DNA-binding proteins, complementing traditional molecular biology methods. While recent advances in machine learning have enabled the integration of predictive systems with bioinformatic approaches, there still needs to be generalizable pipelines for identifying unknown proteins as DNA-binding and assessing the specific type of DNA strand they recognize. In this work, we introduce RUDEUS, a Python library featuring hierarchical classification models designed to identify DNA-binding proteins and assess the specific interaction type, whether single-stranded or double-stranded. RUDEUS has a versatile pipeline capable of training predictive models, synergizing protein language models with supervised learning algorithms, and integrating Bayesian optimization strategies. The trained models have high performance, achieving a precision rate of 95% for DNA-binding identification and 89% for discerning between single-stranded and doublestranded interactions. RUDEUS includes an exploration tool for evaluating unknown protein sequences, annotating them as DNA-binding, and determining the type of DNA strand they recognize. Moreover, a structural bioinformatic pipeline has been integrated into RUDEUS for validating the identified DNA strand through DNA-protein molecular docking. These comprehensive strategies and straightforward implementation demonstrate comparable performance to high-end models and enhance usability for integration into protein engineering pipelines.

bioinformatics↗