Sergio Decherchi

Introduction Machine learning models trained on metabolomics data frequently show reduced performance when applied to independent cohorts because of domain shifts introduced by differences in analytical platforms, sample handling, storage conditions, and metabolite quantification procedures. Robust approaches for harmonizing metabolomics datasets across cohorts are therefore needed to improve model transferability and reproducibility. Objectives To evaluate the impact of matched samples on domain harmonization and cross-cohort prediction performance in NMR metabolomics data, and to compare multiple harmonization strategies for improving model generalization across independently acquired datasets. Methods A machine-learning framework was developed to classify ICD-10 code I50.9 (unspecified heart failure) using plasma NMR metabolomics data from the Estonian Biobank. Models were trained on a source dataset comprising 12,000 samples measured at Nightingale Health and evaluated on an independently measured destination dataset of 295 samples acquired on a different NMR platform. Harmonization approaches included within-domain normalization, Optimal Transport, Bridge mapping, and Empirical Bayes correction. The effect of matched samples was assessed using up to 108 duplicated biospecimens measured in both datasets. Additional validation was performed using independently generated human gestational diabetes and mouse diabetes metabolomics datasets from Metabolomics Workbench. Results Domain harmonization generally improved cross-cohort prediction compared with unaligned data, particularly when matched samples were available. Predictive performance increased as the number of matched samples used for calibration increased, although gains varied across methods. Nevertheless, simple within-domain normalization consistently provided strong performance and often matched or exceeded more complex alignment approaches. External validation using fully unmatched human and mouse metabolomics datasets demonstrated that harmonization strategies improved classification performance relative to raw data, supporting the broader applicability of the framework. Conclusion Matched samples substantially enhance domain harmonization and improve the transferability of metabolomics prediction models across analytical platforms and cohorts. However, robust normalization remains a strong baseline, and the benefits of more sophisticated alignment methods depend on the availability and representativeness of matched reference samples. These findings provide practical guidance for developing reproducible cross-cohort metabolomics machine-learning models while preserving biologically relevant signals.

Severe COVID-19 displays substantial inter-individual variability that cannot be fully explained by common genetic variants, suggesting a contribution of rare functional variation. We performed whole-genome sequencing in 191 Italian patients (96 moderate, 95 severe) to investigate rare coding variants (minor allele frequency EIF4G3, GTF2F2, LY96, COX6A1, and STK11, which are involved in transcriptional and translational regulation, innate immune signaling, and mitochondrial function. Unsupervised biclustering significantly stratified patients by severity, and a decision tree model highlighted the burden of DOCK10 variants as a key discriminator between phenotypes. Convergence between the network-based and machine learning results further supported the robustness of the identified signals. Overall, our integrative framework underscores the contribution of rare variants to COVID-19 severity and highlights the coordinated perturbation of interconnected molecular pathways as a determinant of clinical outcome.

Neural network potentials (NNPs), trained on quantum mechanical (QM) data, can deliver near QM-level accuracy while being much faster than QM calculations. In this work, we introduce a new software tool, Torsionator, which currently supports MACE, OBIWAN, and UMA NNPs for energy minimization of small molecules. Torsionator can analyze a set of selected conformers to parameterize the dihedral angles. These dihedral parameters can then be used for molecular mechanics calculations and simulations. We showcase Torsionator by performing dihedral scanning and parameterization of two representative antiviral compounds: favipiravir defluoro analog T1105 and emivirine. To thoroughly evaluate its robustness and general applicability, we benchmarked Torsionator using MACE on the TorsionNet500 dataset, comprising 500 chemically diverse small molecules with reference QM torsional profiles. These results indicate Torsionator as a practical, efficient, and scalable software tool for dihedral parameterization, enabling its routine integration into molecular simulation workflows.

The global spread of β-lactamase-mediated resistance poses a threat to β-lactam antibiotics. Boron-based β-lactamase inhibitors (BLIs) represent a promising class of reversible covalent inhibitors, yet the molecular basis of their recognition and dissociation remains poorly understood. Using Pseudomonas-derived cephalosporinase-3 (PDC-3) as a model, we employed enhanced sampling strategies with machine learning and steady-state kinetic assays to investigate the binding and unbinding dynamics of LP06, a boronate BLI. We identify three binding pathways, governed by hydrophobic recognition motifs and a conserved arginine anchor that together steer the ligand toward the precovalent state. Sequence alignment of nearly 7000 class C β-lactamases supports the conservation of these determinants, and structural analyses suggest that R349 may act as a shared anchoring point across serine β-lactamases. Additionally, hydrogen-bonding interactions were found to delay productive binding by stabilizing nonproductive conformations. Our findings provide fundamental insights into β-lactamase inhibition and establish design principles for next-generation β-lactamase inhibitors.

NanoShaper (NS) is a widely used tool that leverages ray tracing for molecular surface triangulation, pocket detection, and supports Poisson-Boltzmann equation solvers. By retaining the established methodological framework and implementing a targeted redesign, our approach achieves performance gains of up to 12.5× and memory reductions of up to 8×, enabling the triangulation of complexes containing millions of atoms on relatively modest computational architectures in a short time. The key innovation is a patch-based ray tracing algorithm that replaces the traditional ray-sweeping approach. By iterating over surface patches rather than rays, this method enhances cache performance and removes the need for memory-intensive grids for patch localization, yielding major reductions in memory usage. Further optimizations include the parallelization of the analytical solvent-excluded surface (SES) construction and the replacement of uniform grids and octrees with bilevel grids and/or compressed buffers. We also introduce an analytical ray-torus intersection scheme based on an exact quartic solution, which improves both accuracy and computational efficiency. The updated version (v1.5) additionally provides a public C++ API for seamless integration with external tools. The tool is available at https://gitlab.iit.it/SDecherchi/nanoshaper. Results on more than 1500 structures and on a multimillion atom complex (e.g., the H1N1 virus) confirm speed and accuracy achievements.

The global spread of β-lactamase–mediated resistance poses a severe threat to β-lactam antibiotics. Boronic acid derivatives have emerged as a promising class of reversible covalent inhibitors; however, the molecular determinants governing their recognition and dissociation remains poorly understood. Using Pseudomonas-derived cephalosporinase-3 (PDC-3) as a model, we combined enhanced sampling simulations with machine learning self-organizing maps to investigate the binding and unbinding dynamics of the potent boronic acid inhibitor LP06. Our analysis reveals three distinct binding pathways driven by hydrophobic recognition motifs and a conserved arginine anchor (R349), which collectively guide the ligand toward the pre-covalent state. Unexpectedly, hydrogen-bonding interactions were found to stabilize non-reactive conformations, delaying productive binding. Site-directed mutagenesis of R349, supported by steady-state kinetic assays across varying pH, confirmed its critical role in efficient catalysis. Comparative sequence and structural analyses of >6600 serine β-lactamases, demonstrate that this hydrophobic recognition and arginine anchoring mechanism is broadly conserved. These findings provide mechanistic insights into β-lactamase inhibition and establish design principles for next-generation inhibitors targeting β-lactamase inhibitors.

Analyzing molecular surfaces to predict functional sites and identify protein cavities for small molecule binding is essential in structural biology and drug discovery, particularly when targeting allosteric sites or designing PROTACs. Moreover, measuring properties like volume, surface area, and pockets' chemical descriptors helps in understanding protein function and improving drug development. Over the past decades, numerous surface and pocket-detection tools have been developed. While these tools provide valuable insights, they often require extensive postprocessing of text output files, making the analysis workflow cumbersome. To address this limitation, we introduce NanoShaperWeb, a web server that not only provides the computational capabilities of NanoShaper but also eliminates the need for manual text file processing through an intuitive web-based interface. Molecular surface and pocket-detection computations are performed remotely via a queue, with results visualized interactively and available for download. The application also delivers for each pocket DrugPred descriptors, enabling deeper insights into pocket features. By streamlining molecular analysis, this tool offers an efficient and accessible platform for researchers, supporting key stages of the drug design pipeline. The NanoShaperWeb tool is freely accessible online at https://nanoshaperweb.iit.it/ with no required registration.

In this study, we addressed the challenge of estimating binding free energies in complex biological systems of pharmaceutical relevance, including both protein-ligand and RNA-ligand complexes. As case studies, we examined the intricate binding of the drug Gleevec to Abl-tyrosine kinase and two ligands binding to the preQ1 RNA riboswitch. By refining our approach based on nonequilibrium steered molecular dynamics simulations and path-based collective variables, we tackled the specific difficulties posed by these systems. In particular, the Abl-Gleevec complex is characterized by significant system size and extensive conformational rearrangements of the protein, whereas the systems involving RNA are characterized by marked conformational flexibility. For the Abl-Gleevec system, our method produced binding free energy estimates closely aligned with experimental values, demonstrating its reliability. For the RNA-ligand complexes investigated, we found that the simpler water model TIP3P yields more accurate free energy estimates than the TIP4P-D model, offering practical insight for future research. In this case, the agreement with the experimental results is reasonable. Overall, this work underscores the effectiveness of the proposed path-based workflow in handling complex biomolecular systems with unique characteristics, enabling systematic binding free energy predictions across a variety of targets.

Speed and efficiency of codes for atomistic simulations can be improved through refactoring and tailoring for GPU architectures. This activity, however, comes with associated, often overlooked, costs, namely a reduced readability and flexibility upon optimization and a non-negligible development time. The first element becomes particularly cogent when who carries out the code GPU porting task is not the creator of the algorithm. In this manuscript we investigate these issues by developing and comparing a CUDA (Compute Unified Device Architecture) and an OpenACC version of the MaZe simulative engine, a recently proposed tool for first principles molecular dynamics with interactions computed at the Orbital Free Density Functional level. We developed in approximately the same amount of time the two code bases. Given that this code bears several computational bottlenecks, and given the development time restraints, we ultimately found that OpenACC leads to a code that is not only simpler to maintain, but also faster, as in the OpenACC code base more routines were optimized compared to CUDA.