The E. coli WGS dataset utilized in this study is openly available and was collected from GitHub (2025), and Moradigaravand et al. (2018). Both these datasets have the susceptible and resistant information of the WGS of the E. coli K-12 strain. Due to the common association between these mutations and increased antibiotic resistance in both environmental and clinical contexts, the double-mutated E. coli genome dataset was chosen for its practical and clinical importance. The genetic variety of resistant strains is reflected in this dataset, which captures changes linked to resistance to several classes of antibiotics. This provides valuable insights into the complicated processes of resistance. Previous studies have mostly concentrated on single mutations or resistance genes specific to individual isolates; this work fills a significant void by shifting the focus to double mutations.

We employed two datasets of E. coli in this study, which included WGS, SNP, and resistance-susceptible data for four antibiotics: gentamicin (GEN), cefotaxime (CTX), ciprofloxacin (CIP), and ceftazidime (CTZ), as shown in Table 1. The first dataset has 809 E. Coli strains, which are generated by Ren et al. (2022b). Clinical samples from both humans and animals were used to get the isolates. Using the VITEK^® 2 system (bioMérieux, Nurtingen, Germany), antimicrobial susceptibility testing was carried out, and results were evaluated by EUCAST criteria. The proportion of isolates resistant to CTX, GEN, CTZ, and CIP is 23%, 44%, 34%, and 45% in that sequence.

It is observed from Figure 3 that the dataset utilized in our work has a high imbalance ratio of resistance-susceptible strains for GEN and CTZ antibiotics, with a slight improvement in the case of inconsistency for CTX and CIP antibiotics. There is a significant imbalance ratio in susceptible (S): resistant (R) of 1:3 and 1:2 in the case of GEN and CTZ antibiotics, respectively. While considering all four antibiotics, the ratio of S: R is 1:2 (1,188:2048).

Data quality is the key to various AI model performances (Nayak et al., 2022; Nayak et al., 2023; Swain et al., 2023). Ren, Y. et al. developed the dataset (GitHub, 2025) to preprocess the raw WGS data; it uses BWA-MEM, and clean reads were mapped to the E. coli reference genome (E. coli K-12 strain, MG1655) after low-quality reads were filtered using fastp (v0.23.2) (Chen et al., 2018). By extracting reference and variant alleles and combining isolates according to reference allele positions, single-nucleotide polymorphisms (SNPs) were found using bcftools (v1.14) (Danecek et al., 2011; Li and Durbin, 2009). Preserving alleles that were found to be variations in more than half of the samples and creating an SNP matrix. One-hot encoding transformed the matrix into a binary format for further ML analysis.

This phase is the most crucial and contributes the most toward the model’s performance (Nayak et al., 2024; Mohanty et al., 2023). We utilized the dataset developed by GitHub (2025) for our study. Hence, we restructured the dataset to meet our study objective. The one-hot encoding in the original data ranges from 1-4, while in our study, we modified this to 0.25-1. In addition, we aim to study the effect of this one-hot encoding on computational costs.

To identify E. coli strains that have gained resistance, we created the state-of-the-art aiGeneR 3.0 model; this model is based on DL and ML. The approach we have created is multi-staged and uses modern techniques to boost accuracy and robustness. Beginning with processed Next-Generation Sequencing (NGS) WGS data (GitHub, 2025), which offers a solid foundation for comprehensive inquiry, we use it as our primary source. To prepare the dataset, we use the quality control (QC) pipeline that we developed before (Nayak et al., 2022). In the final phase, highly-trained deep neural networks (LSTM) and Linear regression (LR) are used to reliably identify susceptible and resistant bacteria and to forecast the likelihood of multi-drug resistance in any given strain that shows resistance to any of the four antibiotics under investigation. The design and execution of aiGeneR 3.0 are depicted in Figure 4. To describe the association between gene regulation variables and resistance extent, we used Linear Regression with least-squares optimization to reduce prediction errors and identify resistance-associated markers. Linear Regression provided a baseline prediction framework for deep learning model comparison, making the aiGeneR 3.0 pipeline robust in multidrug resistance categorization. Ultimately, a comprehensive evaluation of its efficacy using a predetermined set of assessment measures ensures the model’s reliability. Biological confirmation also adds credence to its real-world utility. The aiGeneR 3.0 model is an all-inclusive and potent tool that could change the game when finding E. coli antibiotic resistance.

Step 1 Read the data

The primary step in using aiGeneR 3.0 is careful data preprocessing. The genomic sequences of different strains of bacteria are encoded into numerical formats that are suitable for input into neural networks. Enabling further calculation usually involves converting categorical genetic data into a numerical format using methods like one-hot encoding (Dahouda and Joe, 2021). The proposed architecture consists of a total of eight layers, and four types of layers are utilized, out of which three are dense, two are dropout, and one each for regularization, flattening, and softmax make up the model’s architecture. The first, second, and third dense layers contain 64, 64, and 32 neurons, respectively. Comparably, our work employs many values for the dropout and regularization layers. The different regularization values experimented with in our work are 0.01, 0.001, and 0.0001, and the dropout rates are 0.25, 0.5, 0.7, and 0.9. Our proposed local architecture of the LSTM model with other added layers utilizing the random search is shown in Figure 5.

The main objective of this study is to examine and compare the effectiveness of our proposed aiGeneR 3.0 model with different parameters to achieve the best classification accuracy for identifying the resistant strains utilizing the WGS E. coli NGS data. Thus, we adopt several changes during the implementation phase to the parameters of aiGeneR 3.0. We discuss some key phases of our experiments in the following sub-sections and the updation in several parameters of our implemented model, among which a few are shown in Table 2.

During the initial development phase, we are refining the aiGeneR 3.0 model, which employs an LSTM architecture, for our analysis. A softmax layer was incorporated to facilitate classification tasks. A learning rate of 0.01 and a dropout rate of 0.25 were utilized to optimize the training process and mitigate overfitting. To conduct a comprehensive assessment of the model’s performance, we partitioned the dataset into two distinct subsets: the training set and the testing set. The ratio of the train-test split was 70:30. Furthermore, a K-fold cross-validation technique was employed with K = 3 to assess the model’s ability to generalize. Through iterative training and evaluation on various subsets of the dataset, we successfully enhanced the accuracy and dependability of the model in identifying resistant strains in the data.

In this implementation phase, we continued our research by iteratively improving the aiGeneR 3.0 model by changing several critical hyperparameters. We adjusted the learning rate to 0.001 to address overfitting and raised the dropout rate to 0.5. These changes should promote more regularization. To keep the assessment process consistent, we partitioned the dataset at an 80:20 train-test split ratio. We also used a K-fold cross-validation method with K = 5 to strengthen our model evaluation and thoroughly examine its generalizability capabilities, which improved the validation procedure. The model’s training dynamics were fine-tuned using these improvements so that it could better use features from the NGS data to identify resistant bacteria.

During this experiment phase, we kept tweaking the hyperparameters of the aiGeneR 3.0 model to make it even better. Now, we are trying to find the sweet spot by gradually adjusting the model weights during training with a learning rate of 0.0001. We raised the dropout rate to 0.7 to improve model regularization and alleviate overfitting worries; this should lead to more diverse and resilient learned representations. To keep things uniform throughout the assessment, we kept the train-test split ratio at 90:10. To further evaluate the model’s efficacy across different data subsets, we also used a more stringent K-fold cross-validation method with K = 10. This broadened the scope of our validation strategy.

We consistently obtained the best performance metrics with an 8:2 train-test ratio throughout all stages of our model, aiGeneR 3.0, as shown in Table 2. This ratio consistently produced the best outcomes, even with changing parameter combinations during the different phases. After training on 80% of the data and testing on the remaining 20%, our model showed exceptional accuracy, precision, sensitivity, and specificity. This strategy ensured that generalization and model complexity were balanced, enabling reliable performance on several splits of the datasets. Furthermore, at each step, regularization strategies, dropout rates, and K-fold cross-validation were methodically investigated to improve the performance of our models. Notably, the 8:2 train-test ratio was a stable base for attaining optimal outcomes across all of our implementations, even though changes to the parameters affected the model’s behavior.

In addition to the above experiments, we also implemented our aiGeneR 3.0 in several other phases, with the model parameters fine-tuned. We also take the different train-test splits to the above experiments and add various other possible dropout rates. However, we observe different model matrices with each of these implementation phases of our aiGeneR 3.0 and consider the best performance, which is described in sections 5 (results) and 6 (discussions).

Power analysis is a statistical method employed to ascertain the minimal sample size necessary for a study to attain a specified degree of statistical power (Nayak et al., 2024; Jamthikar et al., 2020). Power analysis is essential in the realm of deep learning models as it allows for the estimation of the required sample size to effectively detect significant impacts or disparities in the model’s performance while maintaining a desired level of confidence.

We conducted a power analysis to determine the minimum sample size needed to calculate a population proportion with precision and accuracy. The experiments were conducted utilizing the methodology described in (Jamthikar et al., 2019; Skandha et al., 2020). The formula for sample size calculation, represented by the symbol Sn, is shown in Equation 1: S n =   z * 2 × p ∼ 1 ‐ p ∼ MoE 2   (1)

In this context, MoE represents the margin of error, p ∼ denotes the estimated proportion of the feature in the population, and z* refers to the Z-score linked to the relevant confidence level. The MoE2 was calculated by using half of the width of the confidence interval. We selected a ratio of 0.5 and a confidence level of 95% for our experiment. The power analysis is conducted using MedCalc (Medcalc, 2025) and demonstrates that the study has a sample size (809) that exceeds the required amount to achieve the intended degree of statistical power and correct classification. The minimum sample size for the dataset utilized is 271 (68 are susceptible and 203 are resistant), which is smaller than the available data.

The confusion matrix, a real-to-anticipated-class matrix with multiple evaluation standards, is the primary target of performance parameters. TP and FP stand for true positives and false positives, respectively, in the confusion matrix. Similarly, TN and FN represent true negatives and false negatives, respectively. There are four types of predictions: TP, which accurately predicts that samples with resistance will be resistant; TN, which accurately predicts that samples without resistance will be susceptible; FP, which inaccurately predicts that susceptible samples will be resistant; and FN, which inaccurately predicts that resistant samples will be susceptible.

Measures such as accuracy (Acc), precision (Pre), specificity (Spe), sensitivity (Sen), F1-score (F1), Matthews Correlation Coefficient (MCC), and area under the curve (AUC) are some of the classification performance measures that were studied in this study. The total number of input samples divided by the number of valid predictions is the “processor, ranging from 1 to 4 (dataset A), while the second dataset consists of one-hot encoding. We call the recall the percentage of positive observations that were projected to be positive compared to the total number of positive observations. F1 is the weighted mean of precision and recall. All the model metrics are calculated based on the following equations (Equations 2–7). Accuracy  Acc = T P + T N T P + F P + T N + F N (2) Precision  Pre = T P T P + F P (3) Sensitivity  Sen = T P T P + F N (4) Specificity  Spe = T P T N + F N (5) F 1  Score = 2 *   Precision * Recall Precision + Recall (6) MCC = T P × T N − F P × F N T P + F P T P + F N T N + F P T N + F N (7)

Our proposed aiGeneR 3.0 model’s experimental results show potential in predicting multi-drug resistant (MDR) in E. coli strains. We considered the strain that resists more than two antibiotics to be in the multidrug-resistant category. We used the prediction model of logistic regression (Vermeiren et al., 2007) to estimate the percentage of bacteria resistant to four frequently given antibiotics: CIP, CTX, CTZ, and GEN. The percentage of resistance to each antibiotic is determined by counting the number of antibiotics that the strain is resistant to; the range is 0.25 for resistance to one antibiotic and 1 for resistance to all four antibiotics. The performance of our deployed resistant prediction model achieves a prediction accuracy of 98%, and the other model metrics for all the studied models are shown in Table 7.

The experimental result for MDR prediction witnessed a 98% accuracy rate; our model demonstrated exceptional predictive performance and resilience in detecting variations in MDR. The model’s lowest mean squared error (MSE) during training (0.00054) was found during the model performance evaluation, demonstrating how closely the predicted resistance percentages matched the actual values. Moreover, the high R-squared value of 0.9940 indicates that our model may explain a considerable amount of variability in the resistance percentages across strains. The model’s accuracy for predicting levels of resistance is further demonstrated by the root mean squared error (RMSE) of 0.02327.

Our model’s active predictive capacity was tested using fresh data, and its MSE of 0.00051 confirmed its generalizability and dependability in practical settings. Together, these findings highlight the precision and effectiveness of our suggested aiGeneR 3.0 model in identifying multi-drug resistance in E. coli strains, providing crucial data for directing antibiotic treatment procedures and battling antibiotic resistance. As we obtained the best version of our proposed aiGeneR 3.0 with a moderate learning rate, with an increase in training data (80%), we intended to keep this for multi-drug identification. In addition to this, the model parameters like learning rate, CV, train-test split, and dropout rate of aiGeneR 3.0 are 0.001, 5, 8:2, and 0.5, respectively.

The comparison of classification accuracy on test data and all conceivable train-test splits on the used dataset is shown in Table 8. According to the PE (section 4), the objective is to monitor the effect of data size on the model’s performance. As the proportion of training data increases, the accuracy of the aiGeneR 3.0 classification model rapidly increases. It is observed that the model achieves its highest accuracy of 92% when the train-to-test split ratio is 80:20, as shown in Table 8.

During our experiments, we observed that the studied AI models require more training data for generalization compared to our proposed aiGeneR 3.0 model. If we set the performance threshold as classification accuracy, then our aiGeneR 3.0 takes only 70% (567 unseen cases) of the data to achieve this trademark. Similarly, all of the implemented ML and DL models take 80% of the unseen data to obtain the generalization standard. The generalization of our aiGeneR 3.0 requires 10% less unseen data to obtain the best classification accuracy, proving that our model can be generalized by utilizing fewer strains than all other studied models.

The matrix shown in Figure 9 has significant diagonal dominance, indicating that the model predicted the proper class with few misclassifications. Most GEN-resistant strains (142 of 154) were correctly classified, with only a few CTZ and CTX misassignments. The model also predicted CIP, a smaller class, with great accuracy (148 out of 156 properly classified), demonstrating its class imbalance resilience. The confusion matrix yielded class-wise measurements. Each class has good precision, sensitivity (recall), and specificity, indicating that the model minimized false positives and negatives. Minority class CIP had good sensitivity and specificity, showing that the model did not underperform on underrepresented categories, a major antimicrobial resistance prediction difficulty.

The model for GEN has lower specificity (76%) than sensitivity (93%), suggesting reliable identification of susceptible strains but a little probability of under-detection of resistant isolates. CTZ had 91% sensitivity and 85% specificity, recognizing resistant bacteria with minimal false-positive rates. CTX and CIP had a stable finding, with sensitivity and specificity exceeding 92%, indicating robust class classification. CIP’s sensitivity (93%) and specificity (94%) were the best, detecting resistant bacteria and identifying vulnerable ones. These results show that aiGeneR 3.0 consistently supports better specificity while maintaining excellent sensitivity, ensuring reliable detection of minority resistant strains without inflating misleading resistance predictions. The confusion matrix confirms that aiGeneR 3.0 demonstrated balanced predictive performance among antibiotics, with ∼92% accuracy, 92% precision, 91% sensitivity, 95% specificity, 91% F1-score, and 87% MCC.

The Receiver Operating Characteristic (ROC) curve is an essential metric for evaluating the effectiveness of a classification model. In this study, we conduct a performance analysis of our suggested aiGeneR 3.0 model in comparison to other studied models, with a significance level of p = 0.001. K-5 cross-validation is employed to determine the variation in the accuracy of each model as the quantity of training data changes.

The ROC performance of all the studied models is shown in Figure 10. The proposed model, aiGeneR 3.0, has achieved a significant milestone by achieving a strong Area Under the Curve (AUC) value of 98.48%. Nevertheless, when compared to other classification models, the ROC value of the 1-D CNN is the lowest (93.47%). Compared to the previously developed CNN TL model and despite the hurdles posed by the imbalance and small dataset, our proposed aiGeneR 3.0 achieves the highest AUC value in the identification of the resistant strains.

In the validation phase of our aiGeneR 3.0 model, we utilize an openly available and highly imbalanced dataset (Moradigaravand et al., 2018). The detailed characteristics of the dataset are summarized in Table 9. There is a high imbalance in the susceptible-to-resistant ratio in all four antibiotics taken for validation of our aiGeneR 3.0 model. It can be seen from the table that the ratio is very high in the case of CTZ and GEN (≈1:7) there is a slight increase in the ratio for CTX and CIP (≈1:4). We perform the model validation in two different phases first, we have considered four different datasets based on four individual antibiotics and secondly, prepare the dataset by combining all the four antibiotics into one dataset.

We tested the efficacy of our proposed model on the four individual antibiotics considered for our experiments in the publicly available data, and aiGeneR 3.0 holds the consistency and remains the best performer in terms of classification accuracy, specificity, and sensitivity. In Table 10, we summarize the performance of aiGeneR 3.0 on individual datasets.

It is observed from the above table that, during validation of aiGeneR 3.0 with individual antibiotics data, we obtained a higher classification accuracy in the case of CTX, and this is due to the higher strain ratio compared to the other three antibiotics datasets. aiGeneR 3.0 achieves the second-highest classification accuracy in the case of CTZ (90%), followed by CIP and GEN (89%).

Similarly, while we tested aiGeneR 3.0 along with other studied models on the dataset that combines all four antibiotics, we observed that aiGeneR 3.0 achieves the highest classification accuracy (90%), as shown in Table 11. The sensitivity and specificity of aiGeneR 3.0 are 97% and 76%, respectively, which is the highest among all the studied models, and this is due to the one-hot encoding we adopt in our study. In addition to this, the SVM, aiGeneR 1.0, CNN TL, CNN ensemble, and ResNet-50 achieve a remarkable sensitivity of more than 90%, whereas CNN TL and ResNet-50 achieve a specificity just higher than 70%. The studied model metrics of the validation phase are shown in Figure 11. However, in the validation phase, we observed that the ResNet-50, CNN ensembled, and TL performed better than aiGeneR 1.0. This is due to the effectiveness and automated feature extraction techniques with these models compared to our previously developed aiGeneR 1.0, which is based on traditional feature selection techniques. This validation outcome may provide insight into the use of automated and effective feature selection techniques, especially DL, for future resistant strain identification.

The performance of the designed model on every conceivable train-test split and the comparison of classification accuracy on test data were also explored in this study. The learning model is impacted by the amount of training data, which also helps the model generalize effectively to new data. Using a dataset with various train-test splits, we assess our suggested model, aiGeneR 3.0, and the four other classifiers employed in this investigation. It has been noted that while other models require a more significant number of cases for generalization, aiGeneR 3.0 requires a small number of cases. This section thoroughly explains how data size affects our suggested model. Figure 12 displays the comparison of classification accuracy on test data as well as all conceivable train-test splits on the utilized dataset.

It can be observed from the figure that the ML models require more training samples to obtain generalization in classification accuracy than the DL models. While we compare the top three ML (SVM + DT + KNN) with the top three DL (CNN TL + aiGeneR 1.0 + aiGeneR 3.0) models, there is a significant difference in the train-test split for learning models to achieve their best results. Compared to the top three DL models, the top three ML models take 65.9% more data to be generalized. The other DL models studied in this work take a range of 55%–75% unseen cases to obtain their generalization. In addition to this, our proposed aiGeneR 3.0 requires 70% (567 cases) of data to generalize and obtain a stable classification accuracy. This performance outcome of aiGeneR 3.0 showcases the model’s generalization ability with a very small number of unseen data, which leads to its chances of better performance with real-time data.

Our cutting-edge study reveals that aiGeneR 3.0 is an excellent resource for identifying strains of antibiotic resistance; it can handle imbalanced and constrained datasets with ease. Through the utilization of sophisticated DL algorithms, aiGeneR 3.0 achieves better classification accuracy, as shown in Table 12.

By minimizing variance and keeping feature scaling consistent, this method stabilizes the training process, which in turn produces smoother gradients and avoids problems like bursting gradients (dos Santos and Papa, 2022). We found that our studied DL models performed much better when we used a one-hot encoding range of 0.25-1 rather than 1–4. With a 2% increase in specificity and a 1% improvement in precision, our empirical data demonstrated better accuracy and generalizability on both the validation and test sets. In addition, our model was able to generalize well to a different dataset (Moradigaravand et al., 2018), which further proves how effective and resilient this proposed normalized feature range is for the learning of the deployed DL models.

The validation confirms the edge of aiGeneR 3.0, demonstrating its capacity to surpass rivals with small input data. Its processing cost is minimized, and its simple design gives it the ability to run on typical personal computers and laptops, ensuring better classification accuracy. Furthermore, a comprehensive power analysis reveals aiGeneR 3.0’s capacity to surpass the desired number of training cases, underscoring its potential for further refinement and expansion. Additionally, the significant AUC value of 98.48% shows the potential of our aiGeneR 3.0 toward its adaptability and learning capacity with imbalanced datasets. Overall, our research shows that aiGeneR 3.0 is an innovative breakthrough that will change how we diagnose diseases and, more generally, not just when identifying strains of antibiotic resistance.

We designed the cutting-edge aiGeneR 3.0 model, a DL-based AMR analysis tool, to use double-mutated gene data to predict multi-drug resistance and detect antibiotic resistance strains without prior knowledge of known ARGs. The powerful DL model, LSTM, and LR combination in aiGeneR 3.0 advances AMR analysis, particularly multi-drug prediction. The primary notable accomplishments of our aiGeneR 3.0 framework are as follows:

• We proposed aiGeneR 3.0, an AI model with a simple and robust architecture that can handle imbalances and small genomics data.

This proposed aiGeneR 3.0 model has the ability to identify E. coli bacteria that are resistant to antibiotics, which could be useful for antimicrobial stewardship initiatives. The approach has the potential to lower the usage of inefficient medicines and limit the use of broad-spectrum antibiotics by offering early insights about resistance profiles, which could support antibiotic selection that is tailored to individual patients. Subject to medical validation, the fast prediction capacity suggests real-time clinical decision support. Hospital monitoring systems can benefit from aiGeneR 3.0, which could have applications beyond individual patient care and bolster initiatives to combat new resistance tendencies. However, before these applications are widely used in ordinary practice, more multicenter clinical trials and prospective validation must be conducted.

This highly motivated study focuses on identifying resistant strains utilizing imbalances and small datasets. Data augmentation can be used to further this study and potentially improve model performance. However, because it is medically incorrect, experts do not advise using this approach (the augmentation of medical data). Better model metrics might be obtained if the model were trained using synthetic data. Further study could address a few biases in our model, such as (i) smaller studies are found related to our work, (ii) SNP filtering threshold applied during preprocessing, which may have influenced the set of variants included for model training (iii) the use of data augmentation, (iv) comparisons with other ML and recently trending DL models like deep network with an attention mechanism. (v) a summary of the benchmarking studies and (vi) no remarks regarding the clinical validation (Paul et al., 2022; Eskofier et al., 2016; Hu et al., 2021).

In addition to the above, AiGeneR 3.0 has a few limitations despite its better predictive accuracy. First, despite balancing and preprocessing, the datasets had class imbalances that potentially bias predictions toward the majority class. The model may learn dataset-specific artifacts instead of generalizable biological patterns when training on short or noisy datasets, increasing the risk of overfitting.

This work focuses on applying DL and AI models to resistant strain identification and classification. The proposed aiGeneR 3.0 is now considered a benchmark in the field of AMR analysis due to its great improvement in detecting resistant strains. The aiGeneR 3.0 model performs highly compared to earlier research (Ren Y. et al., 2022) on resistant strain identification. Furthermore, cross-validation and unseen implementations show the system’s endurance, domain flexibility, and capacity to function well in domains other than the one on which it was trained. In extension, the application of other DL models, especially transformer architecture with attention mechanism, and Dl model hyper-parameter optimization may be adopted to validate their efficacy in identifying multi-drug resistance in double-mutated genome sequences.

Despite aiGeneR 3.0’s capabilities, small or imbalanced datasets risk overfitting, when the model learns training data-specific patterns or noise rather than generalizable correlations. Even high accuracy on the training set may not guarantee accurate predictions on unseen data. Imbalanced class distributions bias the model toward majority classes, making unusual antibiotic-resistant organisms harder to detect. Additional variability can worsen model performance in real-world deployment. Different laboratory techniques, sample preparation methods, sequencing platforms, and batch effects introduce systematic variances that may not be in the training data. Sequencing errors or missing data can skew input features, and variable sample distributions across populations may produce patterns the model has not learnt, raising misclassification risk. Parameters and techniques like Cross-validation, data augmentation, and regularization (dropout, weight decay, and early stopping) need to be tested in a wider range. Finally, ongoing retraining with new datasets adapts the model to changing data distributions, ensuring robustness and reliability in clinical or laboratory contexts. Future research should use explainable AI methods like feature attribution or pathway-level analysis to identify resistance-predicting genes or biological markers. For clinical implementation, the model needs to be verified across larger, multi-center cohorts to account for sequencing procedures, sample handling, and patient demographics. To maintain accuracy and dependability in clinical operations, rigorous benchmarking, seamless software interfaces, real-time processing, and constant retraining with new resistance data are needed.