Biology subjects

Oyarzun, D. A.

Publications and source records attributed to Oyarzun, D. A..

At least 19 recordsLinked to original sources

bioRxiv2025

Multiobjective design of growth media with genome-scale metabolic models and Bayesian optimization

The optimization of culture media is critical for improving the efficiency and cost of cellular production systems. Traditional approaches often rely on extensive experimental trials or statistical methods, which can be both costly and time-consuming. Here, we present gsMOBO, a computational approach for media design that integrates genome-scale metabolic models with multiobjective optimization. Our method employs Flux Balance Analysis coupled with a top-layer Bayesian optimization routine for efficient exploration and optimization of nutrient combinations across high-dimensional spaces. We show that our method finds optimal medium formulations along a Pareto front balancing growth, production and cost of medium components. We illustrate the approach in models of Escherichia coli engineered to produce antibody fragments, as well as Bacillus subtilis strains that synthesize cyclic lipopeptides. Our results show that gsMOBO identifies media compositions and Paretooptimal trade-offs consistent with prior experimental work. Our method provides a broadly applicable tool for the design of cost-effective and productive culture media, and offers a route to accelerate medium development in biomanufacturing.

synthetic biology↗

bioRxiv2025

Effective sequence-to-expression prediction for membrane proteins using machine learning and computational protein design

The recombinant expression of integral membrane proteins is notoriously challenging. One way to address this challenge is via computational genotype-to-phenotype models that determine how particular sequence features correlate with protein expression levels. However, the potential of such approaches is yet to be fully realised, at least partly because so few expression datasets are available. Here, we study the sequence-to-expression relationships of a library of 12,248 variants of a specific membrane protein derived from combinatorial computational design. The major advantage of this approach lies in the controlled sequence diversity explored in design, making this new dataset directly compatible with lightweight off-the-shelf bioinformatic tools. The expression phenotype of the entire library is assessed in the widely-used recombinant host Escherichia coli. We employed a relatively small dataset of [~]2000 phenotyped sequences to train a sequence-to-expression predictor using supervised machine learning, which achieved high classification accuracy on held-out test sequences. This model was then used to infer the expression of >10,000 unmeasured sequences, and validation of the top predictions of both high and low expressers achieved 100% success rate. Using tools from explainable AI, we identified specific sequence positions and substitutions that are most important in dictating cellular expression levels. This analysis was validated by model-guided protein engineering that achieved an 8-fold increase in the purification yield of a poorly-expressing variant. Our results show that, at least for this controlled dataset, straightforward and interpretable machine learning can reveal the intrinsic sequence code for membrane protein expression.

bioengineering↗

bioRxiv2025

Representation learning of single-cell time-series with deep variational autoencoders

Single-cell technologies have led to many insights on the function of individual cells within populations. Time-series data, in particular, have become increasingly adopted to uncover the molecular basis of heterogeneity in processes such as gene regulation, growth adaptation, and drug resistance. Yet the analysis of single-cell trajectories often requires manual and application-dependent techniques to extract meaningful features that correlate with the biological processes of interest. Here, we employ representation learning to automatically encode time-series into a low-dimensional feature space that can be used for downstream prediction tasks. We trained deep variational autoencoders on cell length data from Escherichia coli cells growing in a "mother machine" device that allows tracking single cells over long periods. We show that the learned representations preserve the structure of the data across various growth media. Using the pretrained model in tandem with supervised learning, we encoded fresh time-series from cells exposed to single and combination antibiotics, which achieved excellent classification accuracy across drug treatments. Moreover, we demonstrate that the learned embeddings can be used to accurately classify datasets from different laboratories and growth conditions without retraining. This demonstrates that the autoencoder extracts meaningful information with promising generalization power. Our results highlight the potential of representation learning for the analysis of single-cell responses under chemical perturbations and growth conditions.

microbiology↗

bioRxiv2025

Optimization of regulatory DNA with active learning

Many biotechnology applications rely on microbial strains engineered to express heterologous proteins at maximal yield. A common strategy for improving protein output is to design expression systems with optimized regulatory DNA elements. Recent advances in high-throughput experimentation have enabled the use of machine learning predictors in tandem with sequence optimizers to find regulatory sequences with improved phenotypes. Yet the narrow coverage of training data, limited model generalization, and highly nonconvex nature of genotype-phenotype landscapes can limit the use of traditional sequence optimization algorithms. Here, we explore the use of active learning as a strategy to improve expression levels through iterative rounds of measurements, model training, and sequence sampling-and-selection. We explore convergence and performance of the active learning loop using synthetic data and an experimentally characterized genotype-phenotype landscape of yeast promoter sequences. Our results show that active learning can outperform one-shot optimization approaches in complex landscapes with a high degree of epistasis. We demonstrate the ability of active learning to effectively optimize sequences using datasets from different experimental conditions, with potential for leveraging data across laboratories, strains or growth conditions. Our findings highlight active learning as an effective framework for DNA sequence design, offering a powerful strategy for phenotype optimization in biotechnology.

synthetic biology↗

bioRxiv2025

Accurate prediction of gene deletion phenotypes with Flux Cone Learning

Predicting the impact of gene deletions is crucial for biological discovery, biomedicine, and biotechnology. For example, identifying lethal deletions is key for new cancer therapies or antimicrobial treatments that bypass drug resistance. In biotechnology, non-lethal deletions are a powerful strategy to redirect chemical flux toward production of high-value compounds for the food, energy, and pharmaceutical sectors, using genetically engineered cells as an alternative to petrochemicals. Owing to the cost and complexity of large-scale deletion screens, there is a growing interest in computational models that can leverage such data for predictive modelling. Here, we present Flux Cone Learning, a general framework for predicting the impact of metabolic gene deletions on many phenotypes of interest. Flux Cone Learning is based on high-dimensional Monte Carlo sampling of the metabolic space in tandem with supervised learning of fitness scores from deletion screens. This strategy enables the method to learn correlations between the geometry of the metabolic space and a target deletion phenotype. We demonstrate best-in-class predictive accuracy for metabolic gene essentiality in organisms of varied complexity (Escherichia coli, Saccharomyces cerevisiae, Chinese Hamster Ovary cells), outperforming the gold standard predictions of Flux Balance Analysis. The method does not rely on an optimality principle and thus can be applied to a range of organisms where cellular objectives cannot be encoded as an optimization task. We showcase the versatility of Flux Cone Learning in other phenotypes by training a predictor of small molecule production from deletion screening data. Flux Cone Learning provides a widely applicable framework for phenotypic prediction and lays the groundwork for the development of metabolic foundation models across the kingdom of life.

systems biology↗

bioRxiv2025

Identification of drug candidates against glioblastoma with machine learning and high-throughput screening of heterogeneous cellular models

Glioblastoma multiforme (GBM) is an aggressive primary brain tumour that presents significant treatment challenges due to its complex pathology and heterogeneity. The lack of validated molecular targets is a major obstacle for discovering new therapeutic candidates, with no new effective GBM therapies delivered to patients in over two decades. Here, we report the identification of compounds that target the GBM stem cell survival phenotype. Our approach employs machine learning (ML) predictors of cell survival trained on high-throughput, image-based, phenotypic screening data for 3,561 compounds, at multiple concentrations, across a panel of six heterogeneous, patient-derived, GBM stem cell lines. We computationally screened more than 12,000 compounds spanning various chemical classes. Experimental validation of ML-identified candidates across the GBM stem cell lines led to the identification of three compounds with activity against the GBM phenotype. Notably, one of our validated hits, the Hsp90 inhibitor XL888, displayed targeted elimination of all six GBM stem cell lines with IC50 in the nanomolar range. The other two compounds, which displayed broad activity across multiple GBM cell lines with distinct cell line sensitivities, offer routes for future personalised medicine campaigns. Our work demonstrates the use of phenotypic screening in tandem with ML can effectively identify therapeutic leads for personalised treatments in highly heterogeneous indications with few known molecular targets.

cancer biology↗

bioRxiv2024

Single-cell metabolic oscillations are pervasive and may alleviate a proteome constraint

Biological rhythms not only coordinate cellular activity with external signals, but may also enable internal coordination. The metabolic cycle in budding yeast is perhaps the most well-studied example. Historically researchers have investigated this cycle in populations growing in chemostats, but more recently time-lapse microscopy has revealed single-cell oscillations in the redox state of enzyme cofactors and in ATP levels. How to relate the results of these two types of assays is however unclear. Here we report single-cell rhythms too in intracellular pH and show that oscillations in the redox state of flavin molecules occur in auxotrophic and prototrophic strains, in nutrients favouring respiration or fermentation, and in deletion mutants for which oscillations in chemostats are either unobservable or disrupted. To explain the pervasiveness of these rhythms, we postulate that cells generate oscillations to alleviate a proteome constraint - amino acids cells use for one class of enzymes are unavailable for others. Using flux balance analysis with an enzyme-constrained genome-scale metabolic model, we show that, with a finite proteome, sequential synthesis of biomass components typically generates a shorter doubling time than synthesising components in parallel. Our results suggest that the metabolic cycle drives growth and is potentially widespread because all cells grow within a proteome constraint.

systems biology↗

bioRxiv2024

Risk-averse optimization of genetic circuits under uncertainty

Synthetic biology aims to engineer biological systems with specified functions. This requires navigating an extensive design space, which is challenging to achieve with wet lab experiments alone. To expedite the design process, mathematical modelling is typically employed to predict circuit function in silico ahead of implementation, which when coupled with computational optimization can be used to automatically identify promising designs. However, circuit models are inherently inaccurate which can result in sub-optimal or non-functional in vivo performance. To mitigate this issue, here we propose to combine Bayesian inference, Thompson sampling, and risk management to find optimal circuit designs. Our approach employs data from non-functional designs to estimate the distribution of the model parameters and then employs risk-averse optimization to select design parameters that are expected to perform well given parameter uncertainty and biomolecular noise. We illustrate the approach by designing robust adaptation circuits and genetic oscillators with a prescribed frequency. The proposed approach provides a novel methodology for the design of robust genetic circuitry.

synthetic biology↗

bioRxiv2024

Integrating MHC Class I visibility targets into the ProteinMPNN protein design process

ProteinMPNN is crucial in many protein design pipelines, identifying amino acid (AA) sequences that fold into given 3D protein backbone structures. We explore ProteinMPNN in the context of designing therapeutic proteins that need to avoid triggering unwanted immune reactions. More specifically, we focus on intra-cellular proteins that face the challenge of evading detection by Cytotoxic T-lymphocytes (CTLs) that detect their presence via the MHC Class I (MHC-I) pathway. To reduce visibility of the designed proteins to this immune-system component, we develop a framework that uses the large language model (LLM) tuning method, Direct Preference Optimization (DPO), to guide ProteinMPNN in minimizing the number of predicted MHC-I epitopes in its designs. Our goal is to design proteins with low MHC-I immune-visibility while preserving the original structure and function. For our assessment, we first use AlphaFold to predict the 3D structures of designed protein sequences. We then use TM-score, that measures the structural alignment between the predicted design and original protein, to evaluate fidelity to the original protein structure. We find our LLM-based tuning method for constraining MHC-I visibility is able to effectively reduce visibility without compromising structural similarity to the original protein.

bioinformatics↗

bioRxiv2024

Low-dimensional representations of genome-scale metabolism

Cellular metabolism is a highly interconnected network with thousands of reactions that convert nutrients into the molecular building blocks of life. Metabolic connectivity varies greatly with cellular context and environmental conditions, and it remains a challenge to compare genome-scale metabolism across cell types because of the high dimensionality of the reaction flux space. Here, we employ self-supervised learning and genome-scale metabolic models to compress the flux space into low-dimensional representations that preserve structure across cell types. We trained variational autoencoders (VAEs) on large fluxomic data (N = 800, 000) sampled from patient-derived models for various cancer cell types. The VAE embeddings have an improved ability to distinguish cell types than the uncompressed fluxomic data, and sufficient predictive power to classify cell types with high accuracy. We tested the ability of these classifiers to assign cell type identities to unlabelled patient-derived metabolic models not employed during VAE training. We further employed the pre-trained VAE to embed another 38 cell types and trained multilabel classifiers that display promising generalization performance. Our approach distils the metabolic space into a semantically rich vector that can be used as a foundation for predictive modelling, clustering or comparing metabolic capabilities across organisms.

systems biology↗

bioRxiv2024

Large scale analysis of predicted protein structures links model features to in vivo behaviour

Rapid advancements in protein structure prediction methods have ushered in a new era of abundant and accurate structural data, providing opportunities to analyse proteins at a scale that has not been possible before. Here we show that features derived solely from predicted structures can be used to understand in vivo protein behaviour using data-driven methods. We found that these features were predictive of in vivo protein production for a set of designed antibodies, enabling identification of high-quality designs. Following on from this result, we calculated these features for a diverse set of {approx}500,000 predicted structures, and our analysis showed systematic variation between proteins from different organisms to such an extent that the tree of life could be recapitulated from these data. Given the high degree of functional constraint around the chemistry of proteins, this result is surprising, and could have important implications for the design and engineering of novel proteins.

bioinformatics↗

bioRxiv2024

Modelling dynamic host-pathway interactions at the genome scale

Pathway engineering offers a promising avenue for sustainable chemical production. The design of efficient production systems requires understanding complex host-pathway interactions that shape the metabolic phenotype. While genome-scale metabolic models are widespread tools for studying static host-pathway interactions, it remains a challenge to predict dynamic effects such as metabolite accumulation or enzyme overexpression during the course of fermentation. Here, we propose a novel strategy to integrate kinetic pathway models with genome-scale metabolic models of the production host. Our method enables the simulation of the local nonlinear dynamics of pathway enzymes and metabolites, informed by the global metabolic state of the host as predicted by Flux Balance Analysis (FBA). To reduce computational costs, we make extensive use of surrogate machine learning models to replace FBA calculations, achieving simulation speed-ups of at least two orders of magnitude. Through case studies on two production pathways in Escherichia coli, we demonstrate the consistency of our simulations and the ability to predict metabolite dynamics under genetic perturbations and various carbon sources. We showcase the utility of our method for screening dynamic control circuits through large-scale parameter sampling and mixed-integer optimization. Our work links together genome-scale and kinetic models into a comprehensive framework for computational strain design.

synthetic biology↗

bioRxiv2024

DNA representations and generalization performance of sequence-to-expression models

The growing demand for biological products drives many efforts to maximize expression of heterologous proteins. Advances in high-throughput sequencing can produce data suitable for building sequence-to-expression models with machine learning. The most accurate models have been trained on one-hot encodings, a mechanism-agnostic representation of nucleotide sequences. Moreover, studies have consistently shown that training on mechanistic sequence features leads to much poorer predictions, even with features that are known to correlate with expression, such as DNA sequence motifs, codon usage or mRNA secondary structures. However, despite their excellent local accuracy, current sequence-to-expression models can fail to generalize predictions far away from the training data. Through a comparative study across datasets in Escherichia coli and Saccharomyces cerevisiae, here we show that mechanistic sequence features can provide gains on model generalization, and thus improving their utility for predictive sequence design. We explore several strategies to integrate one-hot encodings and mechanistic features into a single predictive model, including feature stacking, ensemble model stacking, and geometric stacking, a novel architecture based on graph convolutional neural networks. Our work casts new light on mechanistic sequence features, underscoring the importance of domain-knowledge and feature engineering for accurate prediction of protein expression levels.

synthetic biology↗

bioRxiv2023

Integration of graph neural networks and genome-scale metabolic models for predicting gene essentiality

Genome-scale metabolic models are powerful tools for understanding cellular physiology. Flux balance analysis (FBA), in particular, is a popular optimization approach for predicting metabolic phenotypes under genetic and environmental perturbations. In model microbes such as Escherichia coli, FBA has been successful at predicting essential genes, i.e. those genes that impair survival when deleted. A central assumption in this approach, however, is that both wild type and deletion strains optimize the same fitness objective. While the optimality assumption may hold for the wild type metabolic network, deletion strains are not subject to the same evolutionary pressures and knock-out mutants may steer their metabolism to meet other objectives for survival. Here, we present FlowGAT, a hybrid FBA-machine learning strategy for predicting essentiality directly from wild type metabolic phenotypes. The approach is based on graph-structured representation of metabolic fluxes predicted by FBA, where nodes correspond to enzymatic reactions and edges quantify the propagation of metabolite mass flow between a reaction and its neighbours. We integrate this information into a graph neural network that can be trained on knock-out fitness assay data. Comparisons across different model architectures reveal that FlowGAT predictions for E. coli are close to those of FBA for several growth conditions. This suggests that gene essentiality can be accurately predicted by exploiting the network structure of metabolism, without additional assumptions beyond optimality of the wild type. Our approach demonstrates the benefits of combining the mechanistic insights afforded by genome-scale models with the ability of deep learning models to extract patterns from complex data.

systems biology↗

bioRxiv2023

Machine learning for optimization of multiscale biological circuits

Recent advances in synthetic biology have enabled the construction of molecular circuits that operate across multiple scales of cellular organization, such as gene regulation, signalling pathways and cellular metabolism. Computational optimization can effectively aid the design process, but current methods are generally unsuited for systems with multiple temporal or concentration scales, as these are slow to simulate due to their numerical stiffness. Here, we present a machine learning method for the efficient optimization of biological circuits across scales. The method relies on Bayesian Optimization, a technique commonly used to fine-tune deep neural networks, to learn the shape of a performance landscape and iteratively navigate the design space towards an optimal circuit. This strategy allows the joint optimization of both circuit architecture and parameters, and hence provides a feasible approach to solve a highly non-convex optimization problem in a mixed-integer input space. We illustrate the applicability of the method on several gene circuits for controlling biosynthetic pathways with strong nonlinearities, multiple interacting scales, and using various performance objectives. The method efficiently handles large multiscale problems and enables parametric sweeps to assess circuit robustness to perturbations, serving as an efficient in silico screening method prior to experimental implementation.

synthetic biology↗

bioRxiv2022

P1 bacteriophage-enabled delivery of CRISPR-Cas9 antimicrobial activity against Shigella flexneri

The discovery of clustered, regularly interspaced, short palindromic repeats (CRISPR) and the Cas9 RNA-guided nuclease provides unprecedented opportunities to selectively kill specific populations or species of bacteria. However, the use of CRISPR-Cas9 to clear bacterial infections in vivo is hampered by the inefficient delivery of cas9 genetic constructs into bacterial cells. Here, we use a broad-host-range P1-derived phagemid to deliver the CRISPR-Cas9 chromosomal-targeting system into Escherichia coli and the dysentery-causing Shigella flexneri to achieve DNA sequence-specific killing of targeted bacterial cells. We show that genetic modification of the helper P1 phage DNA packaging site (pac) significantly enhances the purity of packaged phagemid and improves the Cas9-mediated killing of S. flexneri cells. We further demonstrate that P1 phage particles can deliver chromosomal-targeting cas9 phagemids into S. flexneri in vivo using a zebrafish larvae infection model, where it significantly reduces the bacterial load and promotes host survival. Our study highlights the potential of combining a P1 bacteriophage-based delivery with the CRISPR chromosomal-targeting system to achieve DNA sequence-specific cell lethality and efficient clearance of bacterial infection.

microbiology↗

bioRxiv2022

Discovery of new senolytics using machine learning

Cellular senescence is a stress response characterised by a permanent cell cycle arrest and a proinflammatory secretome. In addition to its tumour suppressor role, senescence is involved in ageing and promotes many disease processes such as cancer, type 2 diabetes, osteoarthritis, and SARS-CoV-2 infection. There is a growing interest in therapies based on targeted elimination of senescent cells, yet so far only a few such senolytics are known, partly due to the poor grasp of the molecular mechanisms that control the senescence survival programme. Here we report a highly effective machine learning pipeline for the discovery of senolytic compounds. Using solely published data, we trained machine learning algorithms to classify compounds according to their senolytic action. Models were trained on as few as 58 known senolytics against a background of FDA-approved compounds or in late-stage clinical development (2,523 in total). We computationally screened various chemical libraries and singled out top candidates for validation in human lung fibroblasts (IMR90) and lung adenocarcinoma (A549) cell lines. This led to the discovery of three novel senolytics: ginkgetin, oleandrin and periplocin, with potency comparable to current senolytics and a several hundred-fold reduction in experimental screening costs. Our work demonstrates that machine learning can take maximum advantage of existing drug screening data, paving the way for new open science approaches to drug discovery for senescence-associated diseases.

cancer biology↗

bioRxiv2022

Prediction of gene essentiality using machine learning and genome-scale metabolic models

The identification of essential genes, i.e. those that impair cell survival when deleted, requires large growth assays of knock-out strains. The complexity and cost of such experiments has triggered a growing interest in computational methods for gene essentiality prediction. In the case of metabolic genes, Flux Balance Analysis (FBA) is widely employed to predict essentiality under the assumption that cells maximize their growth rate. However, this approach implicitly assumes that knock-out strains optimize the same objectives as the wild-type, which excludes cases in which deletions cause large changes in cell physiology to meet other objectives for survival. Here we resolve this limitation with a novel machine learning approach that predicts essentiality directly from wild-type flux distributions. We first project the wild-type FBA solution onto a mass flow graph, a digraph with reactions as nodes and edge weights proportional to the mass transfer between reactions, and then train binary classifiers on the connectivity of graph nodes. We demonstrate the efficacy of this approach using the most complete metabolic model of Escherichia coli, achieving near state-of-the art prediction accuracy for essential genes. Our approach suggests that wild-type FBA solutions contain enough information to predict essentiality, without the need to assume optimality of deletion strains.

systems biology↗