In a recent study published in Engineering, researchers developed a Salmonella Antimicrobial Resistance Predictive Large Language Model (SARPLLM). The model not only predicts antimicrobial resistance (AMR) by Salmonella but also provides a visual representation of pan-genomic investigations.
Image Credit: TopMicrobialStock/Shutterstock.com
Introduction
Salmonella causes widespread foodborne diseases that concern public health. The rise in antimicrobial-resistant forms of Salmonella increases the concern. Salmonella genetic mutations and excessive antimicrobial use have led to the emergence of antimicrobial-recalcitrant variants.
Thus, antimicrobial resistance jeopardizes antibiotic effectiveness and raises disease burdens worldwide. Developing effective AMR prediction methods could improve clinical decision-making and patient outcomes.
Previous LLMs are prone to overfitting during training, resulting in poor predictive outcomes. Furthermore, current methods of Salmonella resistance prediction include source codes, toolkits, and experimental flowcharts that challenge resistance prediction and analysis.
About the study
In the present study, researchers developed SARPLLM, a platform that uses large language modeling to predict antimicrobial resistance among Salmonella organisms. The platform integrates web technology with knowledge graphs that predict resistance and visually indicate results.
The researchers obtained the United States Centers for Disease Control and Prevention (CDC) data on antimicrobial-resistant Salmonella typhimurium from the National AMR Monitoring System Now website.
They analyzed 1,167 samples’ antimicrobial susceptibility testing (AST) and whole-genome sequencing (WGS) data, which the National Center for Biotechnology Information (NCBI) provided. They assigned gene annotations from gene assemblies in the WGS data.
The researchers constructed three matrices from Salmonella gene annotations, representing antimicrobial resistance, Salmonella genes, and accessory genes. The AMR matrix indicated whether a sample was resistant or not.
The Salmonella typhimurium gene matrix comprised pan-genome data, from which core genes were eliminated to obtain accessory genes. Multiple-sequence alignment (MSA) from Salmonella whole-genome sequencing data revealed core gene alignments.
Feature selection for model development involved two steps: a chi-square nonparametric test followed by the Conditional Mutual Information Maximization (CMIM) algorithm. The nonparametric test addressed inaccurate antimicrobial resistance predictions resulting from the analysis of high-dimensional spaces.
The maximization algorithm revealed the relative significance of each Salmonella resistance gene for every antimicrobial. Feature selection using AST and WGS data highlighted accessory genes strongly correlating with AMR based on single-nucleotide polymorphisms (SNPs).
SARPLLM development comprises three procedural steps: prompt engineering, data modeling with fine-tuning, and antimicrobial resistance prediction. Prompt engineering involved prompts asking SARPLLM for resistance predictions based on genetic features in the sequences.
Data consolidation converted core and accessory gene features into computational language descriptions. Core SNP values of G, A, T, and C indicated the lack of guanine, adenine, thymine, or cytosine bases in the SNP locus of the Salmonella sample, respectively. Accessory gene features, designated values 0 or 1, indicated gene absence or presence in the sample.
The researchers converted the datasets into sentences that SARPLLM can recognize using Qwen2 LLM. To do so, they listed the antimicrobials followed by resistance features.
Low-rank adaptations (LoRA) fine-tuned the Qwen2 data to yield model outputs as 1 or 0, indicating resistance or sensitivity, respectively. The quantum sensing-based QSMOTEN algorithm reduced the time complexities in computing inter-sample distances to logarithmic levels of scalability.
The team screened Salmonella genes against five antimicrobials: ceftriaxone, cefoxitin, ampicillin, augmentin, and chloramphenicol.
To assess efficacy, they compared SARPLLM with other AMR prediction models, including random forest classifiers (RF), logistic regression (LR), eXtreme gradient boosting (XGBoost), support vector machines (SVM), resistance prediction neural networks (RPNN), and multilayer perceptron (MLP). Three-fold cross-validation with four repeated tests ensured reliability.
Results
The study demonstrates that the SARPLLM platform outperforms other AMR predictive methods and is robust in predicting Salmonella AMR.
The QSMOTEN algorithm swiftly and accurately computes distances (and similarities) between Salmonella samples, indicating the promising potential of quantum computing approaches in resistance prediction.
The SARPLLM web-based platform is user-friendly. In addition to online AMR predictions based on Salmonella gene features uploaded by users, SARPLLM displays pan-genomic analysis results and gene information as knowledge graphs for data visualization and easier understanding. Moreover, users can easily download the analytical data.
Based on the study findings, the SARPLLM platform integrates LLM and quantum technology to simplify the analysis of high-dimensional information and overcome the sample imbalance issue of WGS procedures for accurate prediction of Salmonella AMR.
The platform helps users visualize and download pan-genome results for Salmonella, making Salmonella study easy and convenient. SARPLLM’s commendable performance against five antimicrobial datasets having missing values indicates excellent generalization and robustness.
However, the platform has a few limitations. Large language modeling requires high-quality data, and quantum computing algorithms need high-performance simulators.
Future studies could enhance multi-source data integration with domain knowledge to improve the accuracy of LLM predictive platforms for Salmonella AMR and develop reliable quantum hardware.
Researchers Develop STITCHR to Insert Whole Genes Without Unwanted Mutations