Biology subjects

Yeoh, J. W.

Publications and source records attributed to Yeoh, J. W..

7 recordsLinked to original sources

bioRxiv2026

Comprehensive evaluation of LLM capabilities for interpretation and analysis of genome-scale metabolic models in metabolic engineering

Genome-scale metabolic models (GSMs) underpin pathway and strain engineering by linking genes to metabolic reactions and enabling system-level simulation of cellular fluxes and intervention effects, yet end-to-end analysis workflows remain fragmented, expert-demanding, and slow to adapt. Large language models (LLMs) could transform this landscape, lowering the barrier by explaining concepts, interpreting GSM files, and turning natural-language instructions into valid analysis code, thereby substantially mitigating the time, effort, and expertise required. However, their reliability for domain-specific tasks remains unexplored. Here, we delivered a systematic benchmark of four leading LLMs (GPT-4, Gemini, Claude, DeepSeek-R1) across four task areas central to metabolic engineering: domain knowledge, metabolic flux prediction, pathway construction, and flux optimization. For benchmarking, we introduced a standardized, rubric-based evaluation framework that uses multi-LLM automated scoring (an ensemble of LLM-as-a-judge assessments) and two distinct sets of nine task-tailored metrics (domain vs coding-focused tasks), rated on a 1-5 scale (up to 45 per task), covering scientific validity and code executability where applicable. Across tasks, we reveal consistent strengths (conceptual explanation, code synthesis) and critical failure modes (e.g., context window limitations, incorrect identifier assumptions, strain-dependent reasoning errors, and errors in domain-specific algorithms). In aggregate, DeepSeek-R1 led in domain tasks, narrowly edging GPT-4, Claude, and Gemini, demonstrating that conceptual biological logic remains highly invariant across architectures. In contrast, Gemini achieved the highest score for coding tasks, distinguished by functional execution and excelled in error handling, documentation, and readability, followed by GPT-4, Claude, and DeepSeek. We also evaluated LLM self-inspection capability by injecting subtle, consequential faults: a stoichiometric sign error causing mass imbalance and an omitted pathway reaction. We reveal that conversational "blind search" prompting completely fails to localize these network faults. Instead, robust error localization requires prompts reframed with domain-informed constraints that force the LLM to leverage tool-assisted code procedures, such as COBRApy mass-balance functions. Together, this work establishes an evidence-based baseline for LLM-enabled GSM analysis, providing actionable guidance for building reliable, automation-ready workflows for pathway and strain design. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=88 SRC="FIGDIR/small/730004v1_ufig1.gif" ALT="Figure 1"> View larger version (28K): org.highwire.dtl.DTLVardef@413e29org.highwire.dtl.DTLVardef@1581de2org.highwire.dtl.DTLVardef@11ef7borg.highwire.dtl.DTLVardef@1817653_HPS_FORMAT_FIGEXP M_FIG C_FIG

systems biology↗

bioRxiv2025

Machine learning-assisted enzyme engineering through ultra-high throughput sorting and large-scale sequence-function data generation

Machine learning (ML) shows great promise in protein engineering but has yet to be integrated with ultra-high throughput sorting (ultra-HTS) and NGS for large-scale sequence-function data generation to harness its capability to explore wider search space and more complex mutation events. Here, we introduce PUSDA, a framework that rapidly sorts mutant libraries into multiple performance groups with good accuracy and generates large-scale sequence-function data to power ML-driven protein design. As a demonstration, PUSDA generated over five million sequence-function data of an enzyme, with data processing revealing over 1.3 million unique enzyme mutants being sorted in a single day. With a trained ML model that achieved 93.52% accuracy, we further analysed combinatorial mutation events and applied a ratio-based selection approach to design novel enzyme sequences. Validation experiment demonstrated a 16.67-fold improvement in efficiency of identifying high-performance enzymes using PUSDA compared to using ultra-HTS alone. The designed novel enzyme achieved 8.23-fold increase in productivity compared to wild type. PUSDA lays a foundation to integrate ultra-HTS, NGS, and ML for future predictive enzyme engineering, offering a data-driven tool for accelerating breakthroughs in biotechnology.

synthetic biology↗

bioRxiv2023

Automated Microbial Classification System based on Deep Convolutional Neural Networks using Images from Colony Picker

Colony screening in single and multi-species environments is an essential step for microbiome studies. However, it possesses a high possibility of inaccurately classifying the species of interest and demands a high degree of manpower and time. Thus, automating the classification of microbes is beneficial to minimize the time and inaccuracy in the colony screening/picking step. Here, we developed an automated microbial classification system for five target species, based on deep Convolutional Neural Networks (CNN) using images captured by an automated robotic colony picker. Multiple possible scenarios of colony culture and diverse morphologies of colonies were examined in building the training and test datasets to train and validate the model and performance on real-life implementations. The final model trained using 60,000 training images, with 12,000 images per species and 3-fold cross-validation, achieved a test accuracy of 94.2% and a test loss of 0.154. Upon testing using a deployment dataset of 4,500 images (900 images per species) with different methods of applying cells onto the agar plate, high accuracy of up to 96.6% was obtained. Five evaluation metrics were implemented to evaluate diverse scenarios of the test data to justify the validity of the model in real-life applications. This model forms a foundation for the classification of more species through transfer learning in the future.

synthetic biology↗

bioRxiv2021

Highly Reversible Tunable Thermal-repressible Split-T7 RNA polymerases (Thermal-T7RNAPs) for dynamic gene regulation

Temperature is a physical cue that is easy to apply, allowing cellular behaviors to be controlled in a contactless and dynamic manner via heat-inducible/repressible systems. However, existing heat-repressible systems are limited and rely on thermal sensitive mRNA or transcription factors which function at low temperatures, lack tunability, suffer delays or overly-complex. To provide an alternative mode of thermal regulation, we developed a library of compact, reversible and tunable thermal-repressible split-T7 RNA polymerase systems (Thermal-T7RNAPs) which fuses temperature-sensitive domains of Tlpa protein with split-T7RNAP to enable direct thermal control of the T7RNAP activity between 30 - 42 {degrees}C. We generated a large mutant library with varying thermal performances via automated screening framework to extend temperature tunability. Lastly, using the mutants, novel thermal logic circuitry was implemented to regulate cell growth and achieve active thermal control of the cell proportions within co-cultures. Overall, this technology expands avenues for thermal control in biotechnology applications.

synthetic biology↗

bioRxiv2021

BMSS2: a unified database-driven modelling tool for systematic model selection and identifiability analysis

SummaryModelling in Synthetic Biology constitutes a powerful tool in our continuous search for improved performance with rational Design-Build-Test-Learn approach. In particular, kinetic models unravel system dynamics, enabling system analysis for guiding experimental designs. However, a systematic yet modular pipeline that allows one to identify the "right" model and guide the experimental designs while tracing the entire model development and analysis is still lacking. Here, we introduce a unified python package, BMSS2, which offers the principal tools in model development and analysis--simulation, Bayesian parameter inference, global sensitivity analysis, with an emphasis on model selection, and a priori and a posteriori identifiability analysis. The whole package is database-driven to support interactive retrieving and storing of models for reusability. This allows ease of manipulation and deposition of models for the model selection and analysis process, thus enabling better utilization of models in guiding experimental designs. Availability and ImplementationThe python package and examples are available on https://github.com/EngBioNUS/BMSS2. A web page which allows users to browse and download the models (SBML format) in MBase is also available with the link provided on GitHub. Supplementary InformationSupplementary data is available.

synthetic biology↗

bioRxiv2020

A ventral striatal-orexin/hypocretin circuit modulates approach but not consumption of food.

Reward-seeking involves the engagement and computation of multiple physiological and motivational parameters. The lateral hypothalamus (LH) is a necessary node in the circuits that control food-seeking and motivation. One group of cells that plays an important yet incompletely understood role in these processes are the orexin/hypocretin (OX/HT) neurons. OX/HT cells are located exclusively within the LH and are implicated in feeding, arousal, and reward-seeking behavior. Importantly, the role of OX/HT neurons in consummatory versus reward-seeking actions is not fully defined, nor are the circuits that control the activity of these neurons under different behavioral states. Here we show that OX/HT neurons respond in real time to food presentation and that this response is modulated by differences in metabolic state. We observed increased OX/HT neuron activity on approach to food, but this activity trended towards pre-approach levels by the start of the consummatory phase. Next, we studied ventrostriatopallidal (VSP) inputs to the OX/HT neurons. Using optogenetics and cell type-specific electrophysiology, we found that ventral pallidum inputs onto OX/HT neurons exert strong inhibitory (and weak excitatory) effects whereas the lateral nucleus accumbens shell provides weaker direct inhibitory connections with OX/HT neurons. These findings reveal that the activity of OX/HT neurons is strongly modulated by metabolic and hedonic state. Further, OX/HT neurons is primarily associated with food approach and that the effect of VSP-terminal output is to suppress OX/HT activity.

neuroscience↗

bioRxiv2018

Activation of lateral hypothalamic group III mGluRs suppresses drug-seeking following abstinence and cocaine-associated increases in excitatory drive to orexin/hypocretin cells

The perifornical/lateral hypothalamic area (LHA) orexin (hypocretin) system is involved in drug-seeking behavior elicited by drug-associated stimuli. Cocaine exposure is associated with presynaptic plasticity at LHA orexin cells such that excitatory input to orexin cells is enhanced, both acutely and into withdrawal. These changes may augment orexin cell reactivity to drug-related cues during abstinence and contribute to relapse-like behavior. Studies in hypothalamic slices from drug-naive animals indicate that agonism of group III metabotropic glutamate receptors (mGluRs) reduces presynaptic glutamate release onto orexin cells. Therefore, we examined the group III mGluR system as a potential target to reduce orexin cell excitability in-vivo, and tested whether activating these receptors could normalize orexin cell activity following cocaine and reduce cocaine-seeking elicited by drug-associated stimuli during abstinence. First, we verified that group III mGluRs regulate orexin cell activity in vivo by showing that intra-LHA infusions of the selective agonist L-(+)-2-Amino-4-phosphonobutyric acid (L-AP4) reduces Fos expression in orexin cells following 24h food deprivation. Next, we extended these findings to show that intra-LHA L-AP4 infusions reduced discriminative stimulus-driven cocaine-seeking following withdrawal. L-AP4 had no effect on general motor activity of sucrose self-administration. Finally, using whole-cell patch clamp recordings from identified orexin cells in orexin-GFP transgenic mice, we show that enhanced presynaptic drive to orexin cells persists for up to 14d into withdrawal and that this plasticity is normalized by L-AP4. L-AP4 had no effect on measures of postsynaptic plasticity in cocaine-exposed animals. Together, these data indicate that agonism of LHA group III mGluRs reduces orexin cell activity in-vivo and is an effective strategy to suppress cocaine-seeking behavior following withdrawal. These effects are likely mediated, at least in part, by normalization of presynaptic plasticity at orexin cells that occurs as a result of cocaine exposure.

neuroscience↗