Raúl Fernández-Díaz

Predicting the location of metal-binding sites in proteins is crucial for fundamental biological questions and biotechnological applications. Over the past decade, the rise in metal-bound protein structures in the Protein Data Bank, combined with advanced statistical models such as deep learning, has accelerated the development of metal-binding site prediction tools. Several approaches are now available, offering high-quality benchmarks and predictive performance. Our initial development in this area is BioMetAll, whose first version was based on backbone pre-organization. Here, we introduce its second version, featuring two major updates: 1) metal-specific scoring functions and 2) prediction using backbone geometry alone or in combination with first coordination sphere descriptors. Apart from demonstrating metal sensitivity and yielding better benchmarking results, this new version allows the assessment of the influence of considering the metal’s first coordination sphere versus backbone pre-organization on how metallic species bind to proteins.

The number of unimodal molecule representation constantly increases, and researchers investigate how to combine them. Intuitively, multimodal representations may provide complementary information and combining them promises better performance. In this work, we systematically explore how combining multiple molecular modalities affects the performance of downstream prediction tasks, providing a baseline for informed decision making. Our study covers 7 molecular modalities and combines them using intermediate and late fusion, and 2 neural network architectures (with or without using knowledge graphs). We conduct experiments with 3 benchmarks for drug-target binding affinity, and 22 molecule property prediction. In total, we train and evaluate over 1400 models. In summary, our results show that combining multiple modalities improve the performance provided that effective fusion strategies are chosen. Knowledge-enhanced representation learning further boosts model performance. Notably, we find that even the use of simple late-fusion approaches establishes state-of-the-art results for some tasks.

Artificial intelligence (AI) has seen transformative breakthroughs in the life sciences, expanding possibilities to interpret biological information at an unprecedented capacity. To maximize return on growing investments and accelerate progress, it is urgent to address long-standing research challenges arising from the rapid adoption of AI methods. We review the erosion of trust in AI outputs driven by poor reusability and reproducibility, and highlight their impact on environmental sustainability. Furthermore, we discuss the fragmented components of the AI ecosystem and lack of guiding pathways to support open and sustainable AI model development. In response, this Perspective introduces practical open and sustainable AI recommendations mapped to over 300 ecosystem components and provides guiding implementation pathways. Our work connects researchers with relevant AI resources, facilitating the implementation of sustainable, reusable and reproducible AI. Built upon community consensus and aligned to existing efforts, these outputs will aid future policy development and structured pathways for guiding AI implementation.

Bioactive peptides are an important class of natural products with great functional versatility. Chemical modifications can improve their pharmacology, yet their structural diversity presents unique challenges for computational modeling. Furthermore, data for standard peptides (composed of the 20 canonical amino acids) is more abundant than for modified ones. Thus, we set out to identify whether predictive models fitted to standard data are reliable when applied to modified peptides. To do this, we first considered two critical aspects of the modeling problem, namely, choice of similarity function for guiding dataset partitioning and choice of molecular representation. Similarity-based dataset partitioning is an evaluation technique that divides the dataset into train and test subsets, such that the molecules in the test set are different from those used to fit the model. Graphical Abstract

Bioactive peptides are an important class of natural products with great functional versatility. Chemical modifications can improve their pharmacology, yet their structural diversity presents unique challenges for computational modeling. Furthermore, data for standard peptides (composed of the 20 canonical amino acids) is more abundant than for modified ones. Thus, we set out to identify whether predictive models fitted to standard data are reliable when applied to modified peptides. To do this, we first considered two critical aspects of the modeling problem, namely, choice of similarity function for guiding dataset partitioning and choice of molecular representation. Similarity-based dataset partitioning is an evaluation technique that divides the dataset into train and test subsets, such that the molecules in the test set are different from those used to fit the model. Scientific contribution. We demonstrate, across four peptide function prediction tasks, that chemical fingerprint-based similarity measures outperform traditional sequence alignment-based metrics for partitioning standard peptide datasets, challenging conventional practice. We have also found that the choice of representation does not have a substantial impact in the standard to standard interpolation modelling scenario, but for modified to modified interpolation chemical language models are the best option. Despite a 50\% drop in performance, chemical fingerprints and a chemical language model (ChemBERTa-2) are the best choices in the more challenging extrapolation scenario (standard to modified). All code and data necessary for reproducing the experiments are available in Github (https://github.com/IBM/PeptideGeneralizationBenchmarks).

HemeFinder has been developed to predict heme-binding sites in natural and heme-dependent de novo enzymes. This tool relies on the structural and physicochemical characteristics of heme-binding sites, including shape, residue composition, and geometric descriptors. HemeFinder benchmarks more than 94% accuracy in identifying the correct heme location, considering the complete set of solutions, and 72% accuracy for the proper location with the correct iron-coordinating residues among the three best-ranked solutions. HemeFinder performs within seconds for monomeric systems and takes minutes for larger multimeric ones, demonstrating that its speed does not compromise its performance. An illustrative case of its potential is provided. HemeFinder is applied to the heme carrier protein 1 (HCP1), a transmembrane protein involved in heme recruitment in evolved organisms, for which no ligand-bound structures have been revealed. HemeFinder provides a relevant prediction of the binding of porphyrin and, when combined with protein-ligand docking, offers the first evidence of low-energy Heme-HCP1 complexes and unveils possible heme pathways. HemeFinder is an interesting, fast, and accurate tool for identifying heme-binding sites in proteins. Source code, documentation, and data are available at https://github.com/laura-tiessler/hemefinder and ESI.

Molecular modelling is a vital tool in the discovery and characterisation of bioactive peptides, providing insights into their structural properties and interactions with biological targets. Many models predicting bioactive peptide function or structure rely on their intrinsic properties, including the influence of amino acid composition, sequence, and chain length, which impact stability, folding, aggregation, and target interaction. Homology modelling predicts peptide structures based on known templates. Peptide–protein interactions can be explored using molecular docking techniques, but there are challenges related to the inherent flexibility of peptides, which can be addressed by more computationally intensive approaches that consider their movement over time, called molecular dynamics (MD). Virtual screening of many peptides, usually against a single target, enables rapid identification of potential bioactive peptides from large libraries, typically using docking approaches. The integration of artificial intelligence (AI) has transformed peptide discovery by leveraging large amounts of data. AlphaFold is a general protein structure prediction tool based on deep learning that has greatly improved the predictions of peptide conformations and interactions, in addition to providing estimates of model accuracy at each residue which greatly guide interpretation. Peptide function and structure prediction are being further enhanced using Protein Language Models (PLMs), which are large deep-learning-derived statistical models that learn computer representations useful to identify fundamental patterns of proteins. Recent methodological developments are discussed in the context of canonical peptides, as well as those with modifications and cyclisations. In designing potential peptide therapeutics, the main outstanding challenge for these methods is the incorporation of diverse non-canonical amino acids and cyclisations.

Motivation: Automated machine learning (AutoML) solutions can bridge the gap between new computational advances and their real-world applications by enabling experimental scientists to build their own custom models. We examine different steps in the development life-cycle of peptide bioactivity binary predictors and identify key steps where automation cannot only result in a more accessible method, but also more robust and interpretable evaluation leading to more trustworthy models. Results: We present a new automated method for drawing negative peptides that achieves better balance between specificity and generalization than current alternatives. We study the effect of homology-based partitioning for generating the training and testing data subsets and demonstrate that model performance is overestimated when no such homology correction is used, which indicates that prior studies may have overestimated their performance when applied to new peptide sequences. We also conduct a systematic analysis of different protein language models as peptide representation methods and find that they can serve as better descriptors than a naive alternative, but that there is no significant difference across models with different sizes or algorithms. Finally, we demonstrate that an ensemble of optimized traditional machine learning algorithms can compete with more complex neural network models, while being more computationally efficient. We integrate these findings into AutoPeptideML, an easy-to-use AutoML tool to allow researchers without a computational background to build new predictive models for peptide bioactivity in a matter of minutes. Availability and implementation: Source code, documentation, and data are available at https://github.com/IBM/AutoPeptideML and a dedicated web-server at http://peptide.ucd.ie/AutoPeptideML. A static version of the software to ensure the reproduction of the results is available at https://zenodo.org/records/13363975.

We have developed Hestia, a computational tool that provides a unified framework for introducing similarity correction techniques across different biochemical data types. We propose a new strategy for dividing a dataset into training and evaluation subsets (CCPart) and have compared it against other methods at different thresholds to explore the impact that these choices have on model generalisation evaluation, through the lens of overfitting diagnosis. We have trained molecular language models for protein sequences, DNA sequences, and small molecule string representations (SMILES) on the alternative splitting strategies for training and evaluation subsets. The effect of partitioning strategy and threshold depend both on the specific prediction task and the biochemical data type, for tasks for which homology is important, like enzymatic activity classification, being more sensitive to partitioning strategy than others, like subcellular localization. Overall, the best threshold for small molecules seems to lay between 0.4 and 0.5 in Tanimoto distance, for DNA between 0.4 and 0.5, and for proteins between 0.3 and 0.5, depending on the specific task. Similarity correction algorithms showed significantly better ability to diagnose overfitting in 11 out of 15 datasets with CCPart being more clearly dependent on the threshold than the alternative GraphPart, which showed more instability. Availability and implementation The source code is freely available at https://github.com/IBM/Hestia. The tool is also made available through a dedicated web-server at http://peptide.ucd.ie/Hestia.