Biology subjects

Guillen-Gosalbez, G.

Publications and source records attributed to Guillen-Gosalbez, G..

4 recordsLinked to original sources

bioRxiv2026

Bioprocess Modeling Made Easy With Hybrax

Hybrid models that embed neural networks in ordinary differential equations are increasingly popular in bioprocess modeling, but they are challenging to set up. Much of the effort goes into the code around the model: wet-lab data must be imported and pre-processed, feeds and sampling must be accounted for in the balance equations, and a training loop with cross-validation must be set up. Researchers usually write this code from scratch for each project. In addition, no widely adopted format for bioprocess data exists, and datasets are mostly published in bespoke spreadsheets or not at all. As a result, models and datasets from the literature are difficult to reproduce, reuse, or compare.We present Hybrax, an open-source Python framework that addresses these problems. It combines a flexible format for bioprocess data with a customizable end-to-end training pipeline for mechanistic and hybrid models. Hybrax derives the transport terms in the process dynamics automatically and provides the utilities needed for training, so that researchers can focus on developing the model itself. It also facilitates model sharing and reproducibility: each training checkpoint contains the model parameters, code, configuration, and training data required to run the model or retrain it from scratch. We demonstrate Hybrax by re-implementing two models from the literature, one trained on real-world E. coli data and one on synthetic HEK293 data, with performance comparable to the original models. Hybrax is available at https://github.com/hybrax/hybrax.

bioinformatics↗

bioRxiv2026

Integrating Metabolic Networks into Hybrid Bioprocess Models

The optimization and control of bioprocesses require robust in silico models that can accurately capture the complex and dynamic behavior of living cells. While hybrid models that combine machine learning with mechanistic equations have emerged as a powerful tools, they often require relatively large datasets and might yield inconsistent predictions that violate the stoichiometry of metabolism. In this study, we introduce FBA-Hyb, a multi-scale hybrid modeling framework that tightly integrates genome-scale metabolic networks via flux balance analysis (FBA) into its architecture. In our FBA-Hyb framework, artificial neural networks predict key FBA inputs (substrate uptake rates and cellular objectives) while a surrogate FBA module translates them into the metabolic fluxes that govern the bioprocess. A key novelty is that the FBA optimization step is replaced by a surrogate generated with symbolic regression, which encapsulates the FBA model into a compact analytical expression. This allows easy backpropagation through the integration of the neural controlled differential equationbased FBA-Hyb bioprocess model. We validated FBA-Hyb against a standard hybrid model (Std-Hyb) using two Escherichia coli fedbatch case studies. In the first study, FBA-Hyb achieved a 42 % average improvement in predictive accuracy (R2) during a leave-one-process-out cross validation. Crucially, FBA-Hyb maintains strict stoichiometric feasibility even during extrapolation. Meanwhile, an alternative approach based on standard architectures leads to stoichiometrically inconsistent solutions in 22 % of the cases analyzed. In the second case study, we demonstrate how FBA-Hyb effectively simulates unmeasured chemical species and discovers a metabolic shift in sulfate-limited regimes during bioprocessing. By providing a modular, biologically consistent, and computationally efficient architecture, FBA-Hyb offers a robust foundation for the next generation of bioprocess models and sustainable process optimization. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=81 SRC="FIGDIR/small/720062v1_ufig1.gif" ALT="Figure 1"> View larger version (28K): org.highwire.dtl.DTLVardef@16f011eorg.highwire.dtl.DTLVardef@b25b5borg.highwire.dtl.DTLVardef@18bd178org.highwire.dtl.DTLVardef@65274e_HPS_FORMAT_FIGEXP M_FIG C_FIG HighlightsO_LIFBA-Hyb integrates flux balance analysis (FBA) into hybrid bioprocess models. C_LIO_LISymbolic regression discovers a simple closed-form FBA surrogate model. C_LIO_LIThe FBA surrogate ensures accurate reaction stoichiometry. C_LIO_LIA neural network predicting the FBA objective keeps the model flexible. C_LIO_LIFBA-Hyb has superior capabilities and accuracy compared to the current standard. C_LI

bioengineering↗

bioRxiv2025

De novo design of peptides localizing at the interface of biomolecular condensates

The interface of biomolecular condensates has been shown to play an important role in processes such as protein aggregation and biochemical reactions. Targeted modulation of these interfaces could, therefore, serve as an effective strategy for engineering condensates and modifying aberrant behaviors. However, the molecular grammar driving the preferential localization of molecules at condensate interfaces remains largely unknown. In this study, we developed a computational pipeline that combines highthroughput coarse-grained simulations, machine learning, and mixed-integer linear programming to design peptides that selectively partition at the interfaces of specific condensate targets. Using this workflow, we designed and synthesized peptides that localize at the interface of three distinct condensates formed by different intrinsically disordered protein regions (IDRs). These peptides exhibit surfactant-like architectures, with one tail incorporated into the condensate and the other excluded from the dense phase. In all cases, the tail entering the condensates is enriched in aromatic residues, while the sequence of the excluded tail varies among the IDRs. For hnRNPA1- and LAF1-IDRs, the excluded tail is enriched in lysines and matches the net charge of the condensate-forming protein, promoting electrostatic repulsion. In the case of DDX4-IDR, which exhibits the lowest charge density, the excluded tail mainly consists of uncharged valine residues, which exhibit negligible interactions with the scaffold protein. These results highlight the importance of the net charge of the scaffold as a key physicochemical parameter for designing peptides with preferential interfacial localization. Overall, our pipeline represents a promising strategy for the rational design of interface-localizing peptides and the identification of the corresponding molecular grammar. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=73 SRC="FIGDIR/small/653111v1_ufig1.gif" ALT="Figure 1"> View larger version (21K): org.highwire.dtl.DTLVardef@eab174org.highwire.dtl.DTLVardef@2852d6org.highwire.dtl.DTLVardef@156960corg.highwire.dtl.DTLVardef@1952e1d_HPS_FORMAT_FIGEXP M_FIG C_FIG

biochemistry↗

bioRxiv2021

Functional-Hybrid Modeling through automated adaptive symbolic regression for interpretable mathematical expressions

Mathematical models used for the representation of (bio)-chemical processes can be grouped into two broad paradigms: white-box or mechanistic models, completely based on knowledge or black-box data-driven models based on patterns observed in data. However, in the past two-decade, hybrid modeling that explores the synergy between the two paradigms has emerged as a pragmatic compromise. The data-driven part of these have been largely based on conventional machine learning algorithm (e.g., artificial neural network, support vector regression), which prevents interpretability of the finally learnt model by the domain-experts. In this work we present a novel hybrid modeling framework, the Functional-Hybrid model, that uses the ranked domain-specific functional beliefs together with symbolic regression to develop dynamic models. We demonstrate the successful implementation of these hybrid models for four benchmark systems and a microbial fermentation reactor, all of which are systems of (bio)chemical relevance. We also demonstrate that compared to a similar implementation with the conventional ANN, the performance of Functional-Hybrid model is at least two times better in interpolation and extrapolation. Additionally, the proposed framework can learn the dynamics in 50% lower number of experiments. This improved performance can be attributed to the structure imposed by the functional transformations introduced in the Functional-Hybrid model.

bioengineering↗