Benchmarking Machine Learning Models in Bioinformatics: Best Practices and Pitfalls

Machine learning benchmarking in bioinformatics is the process of systematically comparing model performance on defined prediction tasks using standardized data splits, evaluation metrics, and statistical tests. For students, researchers, analysts, and life-science professionals, a proper benchmark answers a specific question: which algorithm, feature set, and data handling strategy produces the most reliable predictions for a given biological problem. Without a rigorous benchmarking protocol, published performance claims become difficult to interpret, reproduce, or extend to new datasets. This article outlines a standardized workflow for benchmarking machine learning models in bioinformatics, covering data preparation, validation design, metric selection, statistical comparison, and reporting standards, with attention to common pitfalls such as data leakage and multiple testing.

The Role of Benchmarking in Bioinformatics Research

Bioinformatics research increasingly depends on machine learning to interpret high-dimensional biological data. Applications range from predicting drug sensitivity in cancer cell lines to annotating cell types in single-cell RNA sequencing data and estimating genetic risk for complex diseases. In each case, researchers must choose among many available algorithms, feature representations, and training strategies. A benchmark provides the evidence needed to make those choices defensibly.

The need for systematic comparison is well documented across biological domains. A study of synthetic lethality prediction methods benchmarked 12 recent machine learning approaches and found that all methods improved significantly when data quality was enhanced, for example by excluding computationally derived labels from training and sampling negative examples based on gene expression [5]. This finding shows that benchmark results depend heavily on data preparation decisions, beyond algorithm choice.

Similarly, a comprehensive comparison of machine learning algorithms for drug sensitivity prediction using the Genomics of Drug Sensitivity in Cancer cell line panel trained random forests, neural networks, boosting trees, and elastic nets across 179 anti-cancer compounds [6]. The study compared results on statistical performance, runtime, and interpretability, and demonstrated that complex models benefit from optimized dimensionality reduction while standard models with fewer features can remain competitive [6]. These outcomes reinforce that benchmarking must evaluate the full modeling pipeline, including the final algorithm.

Benchmarking also serves a practical role in clinical and translational settings. A benchmark suite built on the eICU critical care dataset defined four prediction tasks including mortality prediction, length of stay estimation, patient phenotyping, and decompensation risk, comparing clinical gold standards with baseline and deep learning models across roughly 73,000 patients [16]. Public benchmarks of this type allow researchers to replicate results and build on prior work, which accelerates progress in fields where data access is otherwise restricted.

Core Principles of Reliable Benchmarking

Define the Biological Question Before Selecting Tools

A benchmark begins with a precise statement of the prediction task. The task definition determines which datasets are appropriate, which metrics are meaningful, and which algorithms are eligible for comparison. For example, predicting RNA secondary structure from sequence data differs fundamentally from predicting drug response from cell line molecular profiles. Each task carries its own data distribution, evaluation conventions, and failure modes.

The task definition should specify the input features, the target variable, the unit of prediction, and the intended use of the model. A model intended for clinical risk assessment requires different validation than a model intended for hypothesis generation. Researchers should document these decisions before running experiments to avoid post hoc rationalization of results.

Select Representative Datasets

Dataset selection determines whether benchmark results generalize beyond the specific samples used. Public repositories such as the European Bioinformatics Institute training resources and the National Center for Biotechnology Information data resources provide access to curated biological datasets [1][2]. These sources support reproducible research because the data are available to other investigators who wish to verify or extend benchmark findings.

When selecting datasets, consider the diversity of samples, the quality of labels, and the relevance to the intended application. A benchmark built on a single homogeneous cohort may produce optimistic results that do not hold in external populations. Conversely, a benchmark built on highly heterogeneous data may obscure genuine algorithmic differences. The synthetic lethality benchmarking study demonstrated that label quality matters: methods performed better when computationally derived labels were excluded from training data [5]. Researchers should audit label provenance before including any dataset in a benchmark.

Control for Confounding Factors

Confounding occurs when an apparent relationship between features and outcomes is actually driven by an unmeasured variable. In bioinformatics, common confounders include batch effects, sequencing depth, sample preparation differences, and population structure. Machine learning models can exploit these confounders to achieve high training performance while failing in real-world applications.

The microbiome cancer research literature highlights this concern. Untargeted omics techniques require careful selection and use to avoid biases, and the application of machine learning to microbiome data carries specific pitfalls related to data processing and interpretation [21]. Researchers should test whether benchmark conclusions hold after adjusting for known confounders or after training on data from multiple batches or centers.

Designing the Benchmarking Workflow

Data Splitting Strategies

The way data are divided into training, validation, and test sets is the most consequential decision in any benchmark. The standard approach is to hold out a portion of the data for final evaluation, use another portion for validation during model development, and train on the remainder. However, biological data often contain structure that complicates simple random splitting.

Related samples, such as cells from the same patient or sequences from the same gene family, can create information leakage if they appear in both training and test sets. The synthetic lethality study examined diverse data splitting scenarios and found that performance varied substantially depending on how splits were constructed [5]. Researchers should consider split strategies that respect biological groupings, such as splitting by patient, by cell line, or by sequence family.

Nested cross-validation provides a more rigorous alternative to a single train-test split. In nested cross-validation, an outer loop partitions the data into folds for performance estimation, and an inner loop performs model selection within each outer training fold. This approach produces unbiased performance estimates when model selection is part of the pipeline. The lesion-symptom mapping study used nested cross-validation to evaluate machine learning models for predicting language outcomes in stroke survivors, comparing parcellation schemes, neuroimaging modalities, and model types [13][15].

Cross-Validation Considerations

K-fold cross-validation is the most common validation strategy in bioinformatics benchmarks. The data are divided into k folds, and the model is trained on k minus one folds and evaluated on the remaining fold, repeating the process so each fold serves as the test set once. The choice of k affects the bias-variance tradeoff in performance estimates. Smaller k produces faster experiments but higher variance in estimates. Larger k produces more stable estimates but requires more computation.

For small datasets, leave-one-out cross-validation may be appropriate, though it is computationally expensive for large samples. For datasets with hierarchical structure, such as multiple samples per patient, grouped cross-validation ensures that all samples from one patient appear in the same fold. This prevents the model from learning patient-specific patterns that would not generalize to new patients.

The RNA secondary structure prediction review compared 15 methods and found that deep learning algorithms outperformed shallow learning and traditional methods when the data distribution in the test set resembled the training distribution [12]. However, when predicting structures for new RNA families, the advantage of deep learning disappeared, and its performance became inferior or equal to that of other methods [12]. This finding underscores that validation strategy must match the intended deployment scenario. If the model will encounter new biological families, the benchmark must include held-out families in the test set.

Feature Selection and Dimensionality Reduction

Biological datasets often contain far more features than samples. Genomic prediction studies routinely analyze datasets where markers greatly outnumber training lines. The genomic prediction benchmark across 18 traits in six plant species found that feature selection prior to model training was critical for artificial neural networks when markers greatly outnumbered training lines [7]. Dimensionality reduction methods can also improve model performance, but the choice of method interacts with the choice of algorithm.

The drug sensitivity prediction study compared nine dimensionality reduction approaches and found that complex machine learning models benefited from optimized dimensionality reduction strategies [6]. However, standard models using considerably fewer features could still achieve superior performance [6]. This result suggests that researchers should benchmark feature representations and dimensionality reduction methods as part of the modeling pipeline, treating them as variable components instead of fixed preprocessing steps.

Feature selection must be performed inside the cross-validation loop to avoid leakage. If features are selected using the entire dataset before splitting, the test set information influences model training, producing inflated performance estimates. This error is common and can be difficult to detect after the fact.

Metric Selection and Interpretation

Choosing Appropriate Evaluation Metrics

The choice of evaluation metric should reflect the biological and practical goals of the prediction task. Classification tasks commonly use accuracy, precision, recall, F1 score, and area under the receiver operating characteristic curve. Regression tasks use mean squared error, mean absolute error, and correlation coefficients. Ranking tasks use metrics such as normalized discounted cumulative gain or precision at k.

The synthetic lethality study evaluated methods on both classification and ranking tasks, recognizing that different applications require different performance criteria [5]. A method that performs well for classification may perform poorly for ranking, and vice versa. Researchers should select metrics that match the intended use of the model.

For imbalanced datasets, accuracy can be misleading. If 95 percent of samples belong to one class, a model that always predicts the majority class achieves 95 percent accuracy without learning anything useful. Precision, recall, and F1 score provide more informative assessments in this setting. The single-cell annotation study reported accuracy, precision, and F1 scores for eight machine learning models and found that XGBoost achieved high accuracy and F1 scores while simpler models like logistic regression and naive Bayes performed worse [14].

Comparing Against Baselines

A benchmark is only meaningful when compared against appropriate baselines. The drug sensitivity prediction study provided strategies for assessing model performance compared with a simple baseline model and measuring the trade-off between models of different complexity [6]. A baseline might be a simple linear model, a majority-class predictor, or an existing published method.

The lesion-symptom mapping study compared machine learning models against traditional mass univariate statistics, which do not account for multidimensional relationships between variables [13][15]. This comparison established whether machine learning added value beyond established approaches. Without a baseline comparison, a benchmark cannot demonstrate that a new method improves upon the status quo.

The late-onset Alzheimer's disease study compared representative machine learning models for predicting disease risk from genetic variation data and found that the best models achieved approximately 72 percent area under the ROC curve [8]. This absolute performance level provides context for interpreting the relative performance of different models. A benchmark should report both absolute performance and relative comparisons.

Statistical Testing of Performance Differences

Observed differences in performance between models may arise from chance, particularly when datasets are small or noisy. Statistical tests help determine whether observed differences are likely to reflect genuine algorithmic differences. Common approaches include paired tests such as the Wilcoxon signed-rank test or the paired t-test applied to per-fold or per-sample performance values.

Multiple testing is a serious concern in benchmarking. When many models are compared across many datasets or tasks, the probability of finding at least one significant difference by chance increases. The critical care benchmark defined four prediction tasks and compared multiple models on each task [16]. Researchers should apply multiple testing corrections, such as the Bonferroni correction or the Benjamini-Hochberg procedure, when conducting many comparisons.

The genomic prediction study found that no single algorithm performed best across all species and trait combinations, but ensemble predictions combining results from multiple algorithms performed consistently well [7]. This finding suggests that researchers should be cautious about declaring a single winner from a benchmark. Performance differences may be small, dataset-specific, or dependent on hyperparameter choices.

Common Pitfalls in Benchmarking

Data Leakage

Data leakage occurs when information from the test set influences model training. This can happen through feature selection performed on the full dataset, through normalization statistics computed on the full dataset, through duplicate or near-duplicate samples appearing in both training and test sets, or through temporal leakage when predicting future events using models trained on data that include the future.

The synthetic lethality study demonstrated the impact of data quality on benchmark results, showing that excluding computationally derived labels from training improved all methods [5]. This is a form of label leakage, where the training labels contain information derived from the test set or from external sources that would not be available in real applications.

To prevent leakage, all preprocessing steps must be fitted on the training data only and then applied to the test data. This includes scaling, imputation, feature selection, and dimensionality reduction. Researchers should document the preprocessing pipeline precisely so that others can verify that no leakage occurred.

Hyperparameter Tuning on the Test Set

Using the test set to select hyperparameters inflates performance estimates. The test set should be used exactly once, at the end of the benchmarking process, to produce the final performance estimate. Hyperparameter selection should occur within the training data using a validation set or inner cross-validation.

The genomic prediction study found that hyperparameter selection was necessary for all non-linear algorithms [7]. This requirement creates a risk of overfitting the validation process itself if hyperparameter search is extensive. Nested cross-validation addresses this problem by performing hyperparameter selection within each outer training fold, producing an unbiased estimate of the performance of the model selection process.

Ignoring Class Imbalance

Biological datasets frequently have imbalanced class distributions. Disease cohorts may contain far fewer cases than controls. Rare cell types may be underrepresented in single-cell data. Models trained on imbalanced data may achieve high overall accuracy while performing poorly on the minority class.

The single-cell annotation study found that all models struggled with classifying intermediate-stage cells, highlighting challenges in distinguishing transitional cell states [14]. This difficulty would be masked by overall accuracy metrics if the intermediate cells were rare. Researchers should report per-class performance metrics and consider class-balanced evaluation approaches.

Overinterpreting Feature Importance

Machine learning models can identify features associated with outcomes, but these associations do not necessarily imply causation. The enzyme classification study used SHAP analysis to identify molecular weight as the primary discriminative feature for classifying enzyme classes [19]. This finding is useful for understanding the model but should not be interpreted as evidence that molecular weight directly determines enzyme function.

Feature importance measures can also be unstable across different training runs or datasets. Researchers should assess the stability of feature importance rankings before drawing biological conclusions. The Alzheimer's disease study noted that systematic machine learning model selection provides the opportunity to identify new genetic markers potentially associated with the disease [8], but such markers require independent validation.

Practical Implementation Steps

Step 1: Define the Benchmark Protocol

Write a benchmark protocol before running experiments. The protocol should specify the datasets, the prediction tasks, the data splitting strategy, the preprocessing pipeline, the algorithms to compare, the hyperparameter search ranges, the evaluation metrics, and the statistical tests. This protocol serves as a preregistration that reduces the risk of selective reporting.

Step 2: Prepare and Audit the Data

Download the data from public repositories and verify its integrity. The National Center for Biotechnology Information provides access to a wide range of biological datasets [2]. Check for missing values, inconsistent labels, and duplicate samples. Document the data provenance and any filtering steps applied. If the data come from multiple sources, check for batch effects that could confound the benchmark.

Step 3: Implement the Preprocessing Pipeline

Implement preprocessing steps as a pipeline that can be applied consistently to training and test data. Fit all preprocessing parameters on the training data only. This includes scaling parameters, imputation values, and feature selection masks. Save the fitted preprocessing objects so that the pipeline can be applied to new data in the future.

Step 4: Run the Benchmark

Execute the benchmark experiments according to the protocol. Record all experimental settings, including random seeds, hyperparameter values, and computational resources. The R programming language provides extensive support for machine learning and bioinformatics analysis, with thousands of packages available for tasks ranging from machine learning to transcriptome data analysis [22]. Other languages and frameworks may also be appropriate depending on the research context.

Step 5: Analyze and Report Results

Compute performance metrics for each model and each task. Apply statistical tests to compare models. Examine failure cases to understand the limitations of each approach. Report results in a standardized format that includes the dataset description, the preprocessing steps, the model configurations, the performance metrics, and the statistical analyses.

Records and Measurements for Benchmarking

Documentation Requirements

Reproducible benchmarking requires detailed documentation of every step in the analysis. The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable [4]. Applying these principles to benchmark data and code enables other researchers to verify results and build upon them.

The critical care benchmark study made its source code publicly available so that anyone could replicate the results and build upon the work [16]. This practice should be standard for bioinformatics benchmarks. Documentation should include the exact versions of all software packages, the random seeds used, and the computational environment.

Version Control and Environment Management

Use version control for code and configuration files. Record the versions of all software dependencies. Containerization tools can help ensure that the computational environment is reproducible. The benchmark should be rerunnable by other researchers with reasonable effort.

Performance Logging

Log performance metrics for every model and every fold of cross-validation. This allows researchers to examine the variability of performance across folds and to apply statistical tests that account for the paired nature of the data. Raw predictions should be saved so that alternative metrics can be computed without rerunning the experiments.

Common Failure Patterns in Benchmarking Studies

Winner-Takes-All Conclusions

Benchmarks that declare a single best algorithm often fail to account for the variability of performance across datasets and tasks. The genomic prediction study found that no one algorithm performed best across all species and trait combinations [7]. Researchers should report the distribution of performance across datasets instead of focusing only on aggregate statistics.

Inadequate Baseline Comparisons

Benchmarks that compare new methods against other new methods, without including simple baselines, can overstate the value of complex approaches. The drug sensitivity prediction study showed that standard models with fewer features could outperform more complex models [6]. Simple baselines provide a reference point for interpreting the performance of sophisticated methods.

Ignoring Computational Costs

Performance metrics such as accuracy and F1 score do not capture the computational cost of training and evaluating models. The drug sensitivity prediction study compared results regarding statistical performance, runtime, and interpretability [6]. Researchers should report training time, inference time, and memory usage alongside predictive performance.

Confusing Correlation with Causation

Machine learning models can identify predictive patterns without revealing causal mechanisms. The lesion-symptom mapping study identified key brain regions associated with language outcomes, but these associations require careful interpretation [13][15]. Researchers should avoid making causal claims based solely on predictive model performance.

Limitations of Benchmarking

Dataset-Specific Conclusions

Benchmark results are specific to the datasets and tasks evaluated. A model that performs well on one dataset may perform poorly on another. The RNA secondary structure study found that deep learning methods performed well when the test data resembled the training data but lost their advantage on new RNA families [12]. Researchers should be cautious about generalizing benchmark conclusions beyond the evaluated conditions.

The Cold-Start Problem

Many biological prediction tasks involve cold-start scenarios where the model must make predictions for entities not represented in the training data. The synthetic lethality study found that methods had limitations in realistic scenarios such as cold-start independent tests and context-specific synthetic lethality [5]. Benchmarks should include cold-start evaluations when the intended application involves new entities.

Evolving Data Landscapes

Biological data resources change over time. New sequencing technologies, updated genome annotations, and expanded clinical cohorts can alter the characteristics of benchmark datasets. The drug combination therapy review highlighted the importance of data integration and described publicly available databases used as benchmarks [10]. Researchers should periodically reassess whether existing benchmarks remain relevant.

Safety and Regulatory Context

Data Privacy and Sharing

Biological data often contain sensitive information about individuals. The National Institutes of Health Genomic Data Sharing Policy establishes expectations for responsible sharing of genomic data generated through NIH-funded research [3]. Researchers must ensure that benchmark datasets comply with applicable privacy regulations and data sharing policies.

When using clinical data for benchmarking, researchers should verify that the data have been de-identified appropriately and that their use is consistent with the original consent and data use agreements. The emergency department prediction study used the MIMIC-IV-ED database, a public electronic health record resource, to create benchmark datasets [9]. Public data resources of this type provide a pathway for benchmarking while respecting patient privacy.

Reproducibility Standards

The microbiome cancer research literature advocates for extended training opportunities, community standards, and best practices for sharing data and code to advance transparency and reproducibility [21]. These principles apply broadly to bioinformatics benchmarking. Researchers should share code, data, and documentation to the extent permitted by privacy and intellectual property considerations.

Professional Escalation Criteria

When benchmark results will inform clinical decisions or regulatory submissions, researchers should involve appropriate expertise. If a benchmark reveals unexpected performance patterns, such as a model performing far better than expected on a difficult task, investigate potential data leakage before reporting results. If benchmark results conflict with established biological knowledge, consult domain experts before drawing conclusions.

At a Glance

Reporting Template for Benchmarking Results

A standardized reporting template improves the comparability of benchmark studies. The following sections should appear in any benchmark report.

Dataset Description

Describe each dataset used in the benchmark, including the source, the number of samples, the number of features, the class distribution, and any filtering or preprocessing applied. Provide access information so that other researchers can obtain the data. The European Bioinformatics Institute training resources and the National Center for Biotechnology Information data resources are appropriate sources for public biological data [1][2].

Preprocessing Pipeline

Document every preprocessing step, including normalization, imputation, feature selection, and dimensionality reduction. Specify which parameters were estimated from the training data and how they were applied to the test data. This documentation is essential for verifying that no data leakage occurred.

Model Configurations

Describe each model included in the benchmark, including the algorithm, the hyperparameter values, and the training procedure. Specify the hyperparameter search method and the ranges explored. Record the computational resources used for training.

Evaluation Protocol

Describe the data splitting strategy, the number of cross-validation folds, and the statistical tests applied. Specify the evaluation metrics and how they were computed. Describe any multiple testing corrections applied.

Results

Report performance metrics for each model on each task, including measures of variability such as standard deviations or confidence intervals. Report the results of statistical tests comparing models. Include failure analysis describing the types of errors each model makes.

Code and Data Availability

Provide access to the code used for the benchmark and the processed data. The FAIR Guiding Principles provide a framework for making research outputs findable, accessible, interoperable, and reusable [4]. Public availability of code and data enables other researchers to verify results and build upon the work.

Frequently Asked Questions

What is model benchmarking in machine learning?

Model benchmarking is the systematic comparison of machine learning algorithms on defined prediction tasks using standardized data splits, evaluation metrics, and statistical tests. In bioinformatics, benchmarking helps researchers select appropriate algorithms for problems such as drug sensitivity prediction, cell type annotation, and genomic prediction. The process involves preparing datasets, designing validation strategies, computing performance metrics, and applying statistical tests to determine whether observed differences are meaningful.

How do I choose the right data splitting strategy for my benchmark?

The data splitting strategy should reflect the structure of the data and the intended application. If the data contain multiple samples from the same patient or biological entity, use grouped splitting to keep related samples together. If the model will encounter new biological families or entities, include held-out groups in the test set. Nested cross-validation provides unbiased performance estimates when model selection is part of the pipeline. The synthetic lethality study demonstrated that performance varies across data splitting scenarios, so the choice of splitting strategy should be justified and documented [5].

What metrics should I use to evaluate my models?

The choice of metrics should match the prediction task and the intended use of the model. Classification tasks benefit from precision, recall, F1 score, and area under the ROC curve, particularly when classes are imbalanced. Regression tasks use mean squared error, mean absolute error, and correlation coefficients. Ranking tasks require ranking-specific metrics. The synthetic lethality study evaluated methods on both classification and ranking tasks because different applications require different performance criteria [5].

How can I prevent data leakage in my benchmark?

Prevent data leakage by fitting all preprocessing steps on the training data only and applying them to the test data. This includes scaling, imputation, feature selection, and dimensionality reduction. Use grouped cross-validation to prevent related samples from appearing in both training and test sets. Audit labels for computational derivation that could introduce leakage. The synthetic lethality study found that excluding computationally derived labels from training improved all methods [5].

Why do my benchmark results differ from published studies?

Benchmark results depend on many factors, including the dataset, the preprocessing pipeline, the hyperparameter values, and the evaluation protocol. Differences in any of these factors can produce different results. The RNA secondary structure study found that deep learning methods performed well when the test data resembled the training data but lost their advantage on new RNA families [12]. Compare your protocol with the published study to identify sources of discrepancy.

How many models should I compare in a benchmark?

The number of models should be sufficient to answer the research question without creating excessive computational burden or multiple testing problems. The synthetic lethality study compared 12 methods [5], the drug sensitivity study compared four algorithm families [6], and the single-cell annotation study compared eight models [14]. The appropriate number depends on the complexity of the task and the availability of computational resources.

When should I use nested cross-validation?

Use nested cross-validation when model selection is part of the benchmarking pipeline. Nested cross-validation provides an unbiased estimate of the performance of the model selection process by performing hyperparameter selection within each outer training fold. The lesion-symptom mapping study used nested cross-validation to evaluate models along multiple dimensions including parcellation schemes, neuroimaging modalities, and model types [13][15].

How do I report benchmark results for reproducibility?

Report the dataset description, the preprocessing pipeline, the model configurations, the evaluation protocol, and the results with measures of variability. Provide access to code and processed data. The FAIR Guiding Principles provide a framework for making research outputs findable, accessible, interoperable, and reusable [4]. Public benchmarks such as the critical care benchmark suite make source code available so that others can replicate results and build upon the work [16].

Related Bioinformatics Guides

• The Protein Data Bank (PDB): Archival Standards, Structural Validation Metrics, and Bioinformatics Integration Protocols
• Machine Learning for Predicting T-Cell Epitope Immunogenicity
• Machine Learning Classification of Antimicrobial Resistance Genes
• Machine Learning for Variant Effect Prediction on Protein Stability
• Machine Learning for Predicting Host-Virus Protein-Protein Interaction Networks

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Benchmarking machine learning methods for synthetic lethality prediction in cancer.. Nature communications, 2024.
• A comprehensive benchmarking of machine learning algorithms and dimensionality reduction methods for drug sensitivity prediction.. Briefings in bioinformatics, 2024.
• Benchmarking Parametric and Machine Learning Models for Genomic Prediction of Complex Traits.. G3 (Bethesda, Md.), 2019.
• Benchmarking machine learning models for late-onset alzheimer's disease prediction from genomic data.. BMC bioinformatics, 2019.
• Benchmarking emergency department prediction models with machine learning and public electronic health records.. Scientific data, 2022.
• Machine learning approaches for drug combination therapies.. Briefings in bioinformatics, 2021.
• Benchmarking MS/MS Featurization Strategies for Machine Learning-Driven Metabolite Structure Annotation.. Journal of the American Society for Mass Spectrometry, 2026.
• Machine learning for RNA 2D structure prediction benchmarked on experimental data.. Briefings in bioinformatics, 2023.
• Benchmarking machine learning models in lesion-symptom mapping for predicting language outcomes in stroke survivors.. 2025.
• Benchmarking Machine Learning Models for Cell Type Annotation in Single-Cell vs Single-Nucleus RNA-Seq Data. 2025.
• Benchmarking Machine Learning Models in Lesion-Symptom Mapping for Predicting Language Outcomes in Stroke Survivors. 2024.
• Benchmarking machine learning models on multi-centre eICU critical care dataset.. 2020.
• Benchmarking machine learning models for the analysis of genetic data using FRESA.CAD Binary Classification Benchmarking. 2019.
• Role of Artificial Intelligence and Machine Learning in Sustainable Drug Discovery. Brazilian Archives of Biology and Technology, 2024.
• Interpretable Machine Learning in Enzyme Classification: A SHAP-Guided Analysis of Global Structural Features. Journal of Artificial Intelligence Practice, 2026.
• Ensemble deep learning in bioinformatics. Nature Machine Intelligence, 2020.
• Bioinformatics challenges for profiling the microbiome in cancer: pitfalls and opportunities.. Trends in Microbiology, 2024.
• The R Language: An Engine for Bioinformatics and Data Science. Life, 2022.
• Machine Learning: aplicación a datos RNA-Seq. 2016.
• Machine learning: an indispensable tool in bioinformatics.. Methods in Molecular Biology Clifton N J, 2010.
• Integrating host-microbiome multi-omics with machine learning: methods, benchmarks, and translational applications. Science China Life Sciences, 2026.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.