# Choosing the Right Phylogenetic Inference Method ## Quick Answer - Choose maximum likelihood (ML) when you need a fast, well-established analysis of bacterial genomes with bootstrap support values and have limited computational time. - Choose Bayesian inference when your dataset is large, your model is complex, and you need posterior probabilities that directly quantify uncertainty in each clade. - Both methods assume your alignment and substitution model are correct, so poor data quality or model misspecification will produce confident but wrong trees. ## At a Glance | Decision Factor | Maximum Likelihood | Bayesian Inference | |---|---|---| | Computational demand | Lower, parallelizes well across cores | Higher, requires Markov chain Monte Carlo convergence checks | | Support measure | Bootstrap values from resampled data | Posterior probabilities from sampled trees | | Interpretation | Frequency a clade appears in resampled datasets | Probability a clade is correct given the model and data | | Best for bacterial genomics | Core genome SNP alignments, outbreak clusters, routine surveillance | Large pangenome datasets, complex models, when prior information matters | | Software examples | IQ-TREE, RAxML, FastTree | MrBayes, BEAST2, RevBayes | | Main limitation | Bootstrap can overestimate support with large alignments | Posterior probabilities can be overconfident with model misspecification | ## The Core Decision Problem in Bacterial Phylogenomics Bacterial phylogenomic analysis has become a standard step in outbreak investigations, evolutionary studies, and antimicrobial resistance surveillance. The World Health Organization identifies antimicrobial resistance as a global health threat that requires genomic surveillance to track resistant lineages and inform stewardship decisions. Researchers routinely generate whole-genome sequencing data and need to place isolates on a phylogenetic tree to understand transmission, emergence, and spread. The central problem is not whether to build a tree, but which inference framework to use. Maximum likelihood and Bayesian inference are the two dominant approaches, and each has distinct assumptions, computational costs, and statistical interpretations. A researcher who chooses the wrong method for their dataset can produce a tree with high support values that is nevertheless biologically misleading. The decision depends on three factors: dataset size, computational resources, and the type of support measurement needed. For bacterial genomes, the number of isolates can range from dozens in a clinical outbreak to thousands in a global surveillance project. The number of informative sites can be tens of thousands for core genome alignments or hundreds of thousands when recombination-filtered SNPs are included. This article provides a practical comparison of ML and Bayesian methods for bacterial genomes. It covers when to use each approach, how to interpret the support values, and what records and quality controls you need to maintain for reproducible research. ## Understanding the Two Inference Frameworks ### Maximum Likelihood Estimation Maximum likelihood inference searches for the tree and model parameters that maximize the probability of observing your sequence alignment. The method evaluates many candidate trees and selects the one with the highest likelihood score. For bacterial genomes, this typically means analyzing a core genome alignment or a SNP matrix derived from mapping reads to a reference genome. The likelihood score depends on the substitution model, which describes how nucleotides change over time. Common models for bacterial data include GTR with gamma-distributed rate variation and a proportion of invariant sites. The model parameters are estimated from the data, and the tree search explores topology space to find the best arrangement. ML methods are computationally intensive but have benefited from substantial algorithmic improvements. Modern implementations can analyze thousands of bacterial genomes in a few hours on a standard server. The method is deterministic in the sense that the same input data and settings will produce the same tree, which is valuable for reproducibility in public health laboratories. ### Bayesian Inference Bayesian inference takes a different approach. Instead of searching for a single best tree, it samples from the posterior distribution of trees using Markov chain Monte Carlo (MCMC). The posterior distribution combines the likelihood with a prior distribution that encodes your beliefs about the tree and model parameters before seeing the data. The output is a set of sampled trees, and the consensus tree summarizes the most common clades across the samples. The support for each clade is the posterior probability, which is the proportion of sampled trees that contain that clade. A posterior probability of 0.95 means that 95 percent of the sampled trees contain that clade. Bayesian methods are more flexible in some respects. They can accommodate complex models that are difficult to optimize with ML, such as relaxed molecular clocks for dating analyses. They also provide a natural framework for incorporating prior information, such as known sampling dates or known relationships between isolates. The main cost is computational. MCMC requires many iterations to converge, and you must verify that the chain has reached stationarity. Poor convergence produces posterior probabilities that are unreliable, and the method can be slow for large bacterial datasets. ## Dataset Size and Computational Resources ### Small to Medium Bacterial Datasets For datasets with fewer than 100 genomes, both ML and Bayesian methods are computationally feasible. The choice should be driven by the question you are asking. If you need a quick tree for a clinical outbreak investigation, ML with bootstrap support is usually the right choice. The analysis can be completed in minutes, and the bootstrap values provide a familiar measure of confidence. For a small dataset where you need to estimate divergence times or test alternative evolutionary scenarios, Bayesian inference is more appropriate. The posterior distribution provides a richer picture of uncertainty, and the MCMC samples can be used for downstream analyses such as ancestral state reconstruction. ### Large Bacterial Genomic Datasets When the dataset grows to thousands of genomes, ML methods become the practical default. The computational cost of Bayesian MCMC grows with the number of taxa, and convergence can be difficult to achieve for large alignments. ML implementations are designed to scale, and they can handle thousands of genomes with reasonable memory and time. The tradeoff is that bootstrap support values can become inflated with large datasets. When the alignment contains many informative sites, the bootstrap resampling produces very similar datasets, and the support values approach 100 percent even when the true tree is uncertain. This is a known limitation of the bootstrap, and it is especially pronounced in bacterial datasets with large core genomes. ### Computational Resource Planning Before starting an analysis, you should estimate the computational requirements. For ML, the main resources are CPU cores and memory. The alignment size in terms of taxa and sites determines the memory footprint, and the number of cores determines the speed. Most ML programs can run on a standard laboratory workstation for datasets up to a few hundred genomes. For Bayesian inference, the requirements are more demanding. The MCMC chain must run for a sufficient number of generations to converge, and you may need to run multiple chains to verify convergence. This can take days or weeks for large datasets. The memory requirements are also higher because the program must store the alignment and the sampled trees. A practical approach is to start with a small subset of your data to test the computational time and memory usage. This pilot run will tell you whether the full dataset is feasible with your available resources. If the full dataset is not feasible, you can consider ML with a reduced alignment or a Bayesian analysis with a smaller subset. ## Support Measures and Their Interpretation ### Bootstrap Support in Maximum Likelihood The bootstrap is a resampling technique that creates many replicate datasets by sampling alignment columns with replacement. Each replicate is analyzed with the same ML method, and the support for each clade is the proportion of replicates that recover that clade. Bootstrap values are interpreted as a measure of the robustness of the tree to the data. A value of 95 percent means that the clade appears in 95 percent of the resampled datasets. This is not the probability that the clade is correct, but it is a useful heuristic for identifying well-supported clades. For bacterial genomes, bootstrap values above 95 percent are generally considered strong support, while values below 70 percent are considered weak. The threshold depends on the context and the dataset. In outbreak investigations, a clade with high bootstrap support is used to define a transmission cluster. The main limitation of the bootstrap is that it does not account for model misspecification. If the substitution model is wrong, the bootstrap values will be confidently wrong. The bootstrap also becomes less informative with very large datasets, as the resampled datasets become very similar to the original. ### Posterior Probabilities in Bayesian Inference Posterior probabilities are the proportion of sampled trees that contain a given clade. They are interpreted as the probability that the clade is correct given the data, the model, and the prior. A posterior probability of 0.95 means that 95 percent of the sampled trees contain the clade. Posterior probabilities are generally higher than bootstrap values for the same clade. This is because the posterior probability integrates over the uncertainty in the model parameters, while the bootstrap only resamples the data. A clade with a posterior probability of 0.95 may have a bootstrap value of only 70 percent. The interpretation of posterior probabilities is more direct than bootstrap values, but they are also more sensitive to model misspecification. If the model is wrong, the posterior probabilities can be overconfident. This is a known issue in Bayesian phylogenetics, and it is the reason why model testing is essential before running a Bayesian analysis. ### Comparing Support Values Across Methods You cannot directly compare bootstrap values and posterior probabilities. A bootstrap value of 95 percent is not equivalent to a posterior probability of 0.95. The two measures have different statistical interpretations and different sensitivities to model misspecification. A practical approach is to run both methods on the same dataset and compare the resulting trees. If the trees are congruent and the support values are high, you can be confident in the result. If the trees differ, you need to investigate the cause, which is usually model misspecification or a problematic alignment. ## Model Selection and Its Impact ### The Role of the Substitution Model Both ML and Bayesian methods require a substitution model. The model describes the relative rates of different nucleotide changes and the distribution of rates across sites. The choice of model has a substantial impact on the resulting tree and the support values. The most common models for bacterial genomes are the GTR family, which allows different rates for each nucleotide change. The model is often extended with a gamma distribution to account for rate variation across sites and a proportion of invariant sites. Model selection is typically done using information criteria such as AIC or BIC, which balance the fit of the model against the number of parameters. For bacterial genomes, the model selection is often straightforward because the data are informative and the model is well established. However, the model should be tested for each dataset, because the optimal model can vary with the evolutionary distance and the base composition of the genomes. ### Model Misspecification and Its Consequences Model misspecification occurs when the chosen model does not adequately describe the evolutionary process. This can happen when the data has strong base composition bias, when there is recombination, or when the evolutionary rates vary substantially across sites. The consequences of model misspecification are different for ML and Bayesian methods. In ML, the bootstrap values can be inflated, and the tree can be biased toward the model assumptions. In Bayesian inference, the posterior probabilities can be overconfident, and the posterior distribution may not converge to the true tree. The best way to detect model misspecification is to compare the results from different models and different methods. If the trees are stable across models and methods, the result is likely robust. If the trees differ, you need to investigate the cause and consider more complex models. ### Practical Model Selection Workflow A practical workflow for model selection in bacterial phylogenomics is to start with a small subset of the data and test several models. The models should include GTR with and without gamma and invariant sites, and possibly a codon-based model if the analysis is on protein-coding genes. The model with the best AIC or BIC score is selected for the full analysis. The model should be tested again if the full dataset has different characteristics, such as a different number of taxa or a different alignment length. The model selection step is often overlooked, but it is essential for both ML and Bayesian methods. A poor model choice will produce a tree with high support values that are not reliable. ## Practical Workflow for ML Analysis ### Step 1: Prepare the Alignment The first step is to prepare a high-quality alignment. For bacterial genomes, this is usually a core genome alignment or a SNP matrix. The alignment should be checked for gaps, ambiguous characters, and recombination. The quality of the alignment is the most important factor in the analysis. A poor alignment will produce a poor tree, regardless of the inference method. The alignment should be reviewed for columns with many gaps and for sequences that are clearly misaligned. ### Step 2: Select the Model The model is selected using a model selection tool or by comparing likelihood scores. The model should be chosen based on the data, not on the default settings of the software. The GTR model with gamma and invariant sites is a reasonable default for bacterial genomes, but it should be tested. ### Step 3: Run the ML Search The ML search is run with the selected model. The search should be run with multiple starting trees to avoid local optima. The best tree is the one with the highest likelihood score. The search time depends on the dataset size and the number of cores. For a dataset of 100 genomes, the search can be completed in minutes. For a dataset of 1,000 genomes, the search may take several hours. ### Step 4: Run the Bootstrap The bootstrap is run with the same model and the same settings. The number of bootstrap replicates should be at least 100, and 1,000 is recommended for final analyses. The bootstrap replicates are summarized on the best tree. The bootstrap step is computationally expensive, but it is essential for the support values. The bootstrap can be run in parallel with the ML search, and the results are combined at the end. ### Step 5: Interpret and Record The final tree is interpreted with the bootstrap values. The clades with high bootstrap values are considered supported, and the clades with low values are considered uncertain. The tree is recorded with the model, the alignment, and the settings for reproducibility. The records should include the software version, the model, the number of bootstrap replicates, and the random seed. This information is essential for reproducing the analysis and for comparing results across studies. ## Practical Workflow for Bayesian Analysis ### Step 1: Prepare the Alignment and Model The alignment is prepared in the same way as for ML. The model is selected using the same criteria. The Bayesian analysis requires a prior distribution for the model parameters, which is usually set to the default values. The prior distribution should be checked for the data. The default priors are usually flat and uninformative, but they can be adjusted if you have prior knowledge about the evolutionary process. ### Step 2: Run the MCMC The MCMC is run with the selected model and prior. The chain is run for a sufficient number of iterations to converge. The number of iterations depends on the dataset size and the complexity of the model. The MCMC produces a set of sampled trees and parameter values. The samples are taken at regular intervals, and the first portion of the chain is discarded as burn-in. ### Step 3: Check Convergence The convergence of the MCMC is checked using the effective sample size (ESS) and the potential scale reduction factor. The ESS should be above 200 for the parameters of interest, and the PSRF should be close to 1.0. If the convergence is not achieved, the chain is run for more iterations. The convergence is also checked by running multiple chains and comparing the results. ### Step 4: Summarize the Posterior The posterior distribution is summarized as a consensus tree. The consensus tree is the tree with the highest posterior probability, and the clade support is the posterior probability. The posterior probabilities are recorded for each clade. The clades with posterior probabilities above 0.95 are considered supported, and the clades with lower values are considered uncertain. ### Step 5: Interpret and Record The final tree is interpreted with the posterior probabilities. The records include the model, the prior, the number of iterations, the convergence metrics, and the software version. The records are essential for reproducibility. The Bayesian analysis is more complex than ML, and the records must be detailed enough for another researcher to reproduce the analysis. ## Common Failure Patterns and How to Avoid Them ### Failure Pattern 1: Overconfident Support with Large Datasets The most common failure pattern is overconfident support with large bacterial datasets. The bootstrap values approach 100 percent and the posterior probabilities approach 1.0, even when the true tree is uncertain. This is caused by the large number of informative sites in bacterial genomes. The solution is to be cautious when interpreting support values from large datasets. The support values should be considered in the context of the dataset size and the biological question. A clade with 100 percent bootstrap support in a dataset of 1,000 genomes is not necessarily more reliable than a clade with 90 percent support in a dataset of 100 genomes. ### Failure Pattern 2: Ignoring Recombination Bacterial genomes are subject to recombination, which can create conflicting signals in the alignment. The recombination can cause the tree to be a mosaic of different evolutionary histories, and the support values can be misleading. The solution is to test for recombination before the analysis. The recombination can be detected with tools that scan the alignment for breakpoints. If recombination is detected, the analysis should be performed on the core genome or on the recombination-free regions. ### Failure Pattern 3: Using the Wrong Model The model is often chosen by default, without testing. The default model may not be appropriate for the data, and the resulting tree can be biased. The solution is to test the model for each dataset. The model selection should be performed with a model selection tool, and the selected model should be used for the analysis. ### Failure Pattern 4: Ignoring Convergence in Bayesian Analysis The Bayesian analysis is often run without checking the convergence. The MCMC may not have converged, and the posterior distribution is not reliable. The solution is to check the convergence for every analysis. The ESS and the PSRF should be checked, and the chain should be run for more iterations if the convergence is not achieved. ### Failure Pattern 5: Not Recording the Analysis The analysis is often run without recording the settings. The records are essential for reproducibility and for the interpretation of the results. The solution is to record the software version, the model, the settings, and the data for every analysis. The records should be stored with the results. ## Quality Controls and Reproducibility ### Authenticated Biological Materials The quality of the phylogenetic analysis depends on the quality of the input data. The bacterial isolates should be authenticated and the sequences should be checked for contamination. The American Type Culture Collection provides authenticated biological materials and strain information that can be used as references for quality control. The sequences should be checked for contamination with other bacterial species. The contamination can be detected by the presence of unexpected sequences in the alignment. The contaminated sequences should be removed or re-sequenced. ### Alignment Quality The alignment quality is the most important factor for the analysis. The alignment should be checked for gaps, ambiguous, and misaligned sequences. The alignment should be trimmed to remove the poorly aligned regions. The alignment quality can be assessed by the number of informative sites and the proportion of gaps. The alignment should be reviewed visually for the problematic regions. ### Reproducibility Records The analysis should be reproducible by another researcher. The records should include the software version, the model, the settings, and the data. The records should be stored in a version-controlled repository. The reproducibility is essential for the scientific process. The analysis should be run with a fixed random seed, and the results should be recorded. ### Quality Controls for the Analysis The quality controls for the analysis include the convergence checks for the Bayesian analysis and the bootstrap for the ML analysis. The quality controls should be performed for every analysis. The quality controls should be recorded in the analysis report. The report should include the support values, the convergence metrics, and the model selection results. ## Biosafety and Laboratory Context ### Risk Assessment for Bacterial Work The phylogenetic analysis is a computational task, but the input data is generated in the laboratory. The laboratory work with bacterial strains requires a risk assessment and appropriate containment. The CDC and NIH Biosafety in Microbiological and Biomedical Laboratories provides the principles for risk assessment, containment, and decontamination. The risk assessment should consider the bacterial species, the source of the isolates, and the procedures used. The containment should be appropriate for the risk group of the organism. The decontamination procedures should be followed for the waste and the equipment. ### Professional Laboratory Practice The laboratory work should follow the professional standards for microbiology. The American Society for Microbiology provides education and laboratory practice context for the microbiology workforce. The laboratory should follow the standard operating procedures for the bacterial culture and the DNA extraction. The laboratory quality should be monitored with the appropriate controls. The controls should include the reference strains and the negative controls. The quality control results should be recorded. ### The Public Health Context The phylogenetic analysis is often used for the public health surveillance of bacterial pathogens. The World Health Organization identifies antimicrobial resistance as a global health threat that requires surveillance and stewardship. The phylogenetic analysis is used to track the spread of resistant strains and to inform the public health response. The analysis should be performed in the context of the public health goals. The results should be interpreted with the epidemiological data and the clinical context. ## When to Escalate to a Specialist ### The Need for a Specialist The phylogenetic analysis is a specialized skill, and the interpretation of the results requires experience. The researcher should escalate to a specialist when the analysis is complex or when the results are unexpected. The specialist can be a bioinformatician, a phylogeneticist, or a public health microbiologist. The specialist can help with the model selection, the convergence, and the interpretation of the results. ### The Escalation Criteria The escalation criteria include the following: - The analysis produces a tree with conflicting support values across methods. - The convergence is not achieved after a reasonable number of iterations. - The model selection is not clear, and the results are sensitive to the model. - The recombination is detected, and the analysis is not straightforward. - The results are used for a public health decision, and the analysis needs to be reviewed. The escalation should be done before the results are used for the decision. The specialist should review the analysis and the results. ### The Documentation for Escalation The documentation for the escalation should include the alignment, the model, the settings, and the results. The documentation should be complete enough for the specialist to review the analysis. The documentation should be stored in a version-controlled repository. The specialist should have access to the data and the analysis. ## A Practical Decision Framework for Method Selection ### The Three-Question Screening Approach Before committing computational resources to either ML or Bayesian inference, researchers should run a structured screening process that aligns the analysis design with the biological question, dataset properties, and available infrastructure. This framework prevents the common error of defaulting to one method without considering the specific demands of the bacterial dataset at hand. The first question concerns the analytical goal. If the objective is to place isolates into clusters for outbreak investigation or routine surveillance, ML with bootstrap support is typically sufficient and delivers results quickly. If the objective includes estimating divergence times, testing explicit evolutionary hypotheses, or quantifying uncertainty in a way that informs downstream modeling, Bayesian inference provides the richer output. The World Health Organization emphasizes that genomic surveillance for antimicrobial resistance requires tracking resistant lineages, and the analytical goal should match the public health question being asked. The second question concerns dataset composition. Bacterial datasets vary in the number of genomes, the number of informative sites, and the degree of recombination. A dataset with fewer than 100 genomes and a core genome alignment of moderate length is tractable with either method. A dataset with thousands of genomes and a large pangenome alignment will strain Bayesian MCMC convergence and is better suited to ML. The number of informative sites matters because both bootstrap values and posterior probabilities become inflated when the alignment is large and the signal is strong. The third question concerns computational resources. The researcher must estimate the time and memory required for the analysis before committing to a method. For ML, the main cost is the tree search and the bootstrap replicates. For Bayesian inference, the cost is the MCMC chain length and the convergence checks. A pilot run on a subset of the data provides the empirical basis for this estimate. ### The Pilot Run Protocol A pilot run is a small-scale version of the intended analysis that uses a random subset of taxa or a reduced alignment. The purpose is to measure computational time, memory usage, and the behavior of the support values before running the full analysis. This protocol is especially important for Bayesian inference, where the MCMC chain length is not known in advance and convergence can be difficult to achieve. The pilot run should use at least 10 percent of the taxa or a random sample of 20 to 50 genomes, whichever is larger. The alignment should be the same as the full dataset, but the taxa are subsampled. The pilot run should use the same model and the same software settings as the intended full analysis. The time to completion and the peak memory usage are recorded. For ML, the pilot run provides an estimate of the time per bootstrap replicate and the total time for the full analysis. If the pilot run takes 10 minutes for 50 genomes, the full analysis with 500 genomes will take approximately 100 minutes, assuming linear scaling. The memory usage scales with the number of taxa and the alignment length, so the pilot run also provides a memory estimate. For Bayesian inference, the pilot run provides an estimate of the time per generation and the number of generations needed for convergence. The pilot run should be run for a fixed number of generations, and the convergence metrics should be checked. If the effective sample size is below 200 after the pilot run, the full analysis will require more generations. The pilot run also provides an opportunity to test the model selection. The model with the best AIC or BIC score in the pilot run is likely to be the best model for the full dataset, but the model should be tested again if the full dataset has different characteristics. ### Record Keeping for the Decision The decision framework requires a record system that captures the rationale for the method choice and the results of the pilot run. This record is essential for reproducibility and for the interpretation of the final tree. The record should include the following items: - The analytical question and the type of support measure needed - The dataset size, including the number of taxa and the number of informative sites - The computational resources available, including the number of cores and the memory - The pilot run results, including the time, the memory, and the convergence metrics - The selected method and the rationale for the selection - The software version and the model used for the analysis The record should be stored with the analysis results in a version-controlled repository. This record allows another researcher to understand why the method was chosen and to reproduce the analysis. The American Society for Microbiology provides professional laboratory practice context that emphasizes the importance of documentation and quality control in microbiology work. ### Interpreting the Decision Framework Output The decision framework produces a clear output: a method choice with a documented rationale. The output should be reviewed before the full analysis is run. If the decision is ML, the analysis proceeds with the bootstrap support. If the decision is Bayesian, the analysis proceeds with the MCMC and the posterior probabilities. The framework also identifies the risk points. If the dataset is large and the computational resources are limited, the framework flags the risk of overconfident support values. If the model selection is ambiguous, the framework flags the risk of model misspecification. These flags should be addressed before the full analysis. ### The Role of Authenticated Reference Strains The decision framework should include a quality control step that uses authenticated reference strains. The American Type Culture Collection provides authenticated biological materials and strain information that can be used as references for the analysis. The reference strains should be included in the alignment to verify that the tree recovers the expected relationships. The reference strains serve two purposes. First, they provide a positive control for the analysis. If the reference strains do not cluster as expected, the analysis has a problem. Second, they provide a calibration point for the tree. The reference strains can be used to root the tree or to estimate the divergence times. The reference strains should be included in the pilot run and in the full analysis. The placement of the reference strains should be recorded in the analysis report. If the reference strains are misplaced, the analysis should be investigated before the results are used. ### Escalation Criteria for the Decision Framework The decision framework includes escalation criteria for situations where the analysis is beyond the scope of the routine workflow. The escalation criteria are the following: - The pilot run does not complete within the expected time or memory budget - The convergence metrics are not achieved after a reasonable number of generations - The model selection is not clear, and the results are sensitive to the model - The reference strains are misplaced, and the cause is not obvious - The analysis is used for a public health decision, and the results need to be reviewed The escalation should be done before the full analysis is run. The specialist should review the pilot run results and the decision framework record. The specialist can help with the model selection, the convergence, and the interpretation of the results. ### The Role of the Decision Framework in the Laboratory Context The decision framework is part of the laboratory quality system. The laboratory work with bacterial strains requires a risk assessment and appropriate containment, as described in the CDC and NIH Biosafety in Microbiological and Biomedical Laboratories. The computational analysis is a downstream step, but the quality of the input data depends on the laboratory procedures. The laboratory procedures should follow the professional standards for microbiology. The American Society for Microbiology provides education and laboratory practice context for the microbiology workforce. The laboratory should follow the standard operating procedures for the bacterial culture and the DNA extraction. The quality control results should be recorded. The decision framework is also relevant to the public health context. The World Health Organization identifies antimicrobial resistance as a global health threat that requires surveillance and stewardship. The phylogenetic analysis is used to track the spread of resistant strains and to inform the public health response. The decision framework ensures that the analysis is appropriate for the question and the data. ### Common Failure Patterns in the Decision Process The decision process has its own failure patterns that are distinct from the failure patterns in the analysis itself. The first failure pattern is the use of the default method without the decision framework. The researcher chooses ML or Bayesian inference based on habit or the software default, without considering the dataset size, the computational resources, or the analytical question. This failure is avoided by running the decision framework for every analysis. The second failure pattern is the use of the pilot run results without the full analysis. The pilot run is a subset of the data, and the results may not scale linearly to the full dataset. The full analysis should be run with the same settings as the pilot run, and the results should be checked for consistency. The third failure pattern is the lack of documentation. The decision framework record is essential for the reproducibility and for the audit. The record should be stored with the analysis results, and the record should be complete enough for another researcher to reproduce the decision. The fourth failure pattern is the lack of escalation. The researcher may continue with the analysis even when the pilot run indicates a problem. The escalation criteria should be followed, and the specialist should be consulted when the criteria are met. ### The Decision Framework in Practice The decision framework is a practical tool that can be applied to any bacterial phylogenomic analysis. The framework is implemented in the following steps: 1. Define the analytical question and the support measure needed. 2. Assess the dataset size and the computational resources. 3. Run a pilot run on a subset of the data. 4. Record the pilot run results and the decision rationale. 5. Select the method and run the full analysis. 6. Check the results and the convergence metrics. 7. Record the analysis and the results. 8. Escalate to a specialist if the escalation criteria are met. The framework is designed to be flexible. The researcher can adjust the steps to the specific dataset and the available resources. The framework is not a substitute for the expertise of a phylogeneticist, but it provides a structured approach that reduces the risk of the common failure patterns. The decision framework is also a teaching tool. The framework can be used to train new researchers in the practical aspects of the phylogenetic analysis. The framework provides a clear structure for the decision and the documentation, which is essential for the professional practice of the microbiology workforce. ## Frequently Asked Questions ### What is the main difference between ML and Bayesian phylogenetic inference? The main difference is the statistical framework. ML searches for the single tree that maximizes the likelihood of the data, while Bayesian inference samples a distribution of trees and reports the posterior probability for each clade. The ML result is a point estimate with bootstrap support, and the Bayesian result is a distribution with posterior probabilities. ### Which method is faster for bacterial genomes? ML is generally faster for bacterial genomes. The ML search is optimized for large datasets, and the bootstrap can be parallelized. The Bayesian MCMC requires more iterations and is slower for large datasets. For datasets with more than a few hundred genomes, ML is the practical choice. ### Can I use both methods on the same dataset? Yes, you can use both methods on the same dataset. The comparison of the results can be informative. If the trees are similar and the support values are high, the result is robust. If the trees differ, you need to investigate the cause, which is usually model misspecification or recombination. ### How many bootstrap replicates do I need? The number of bootstrap replicates should be at least 100, and 1,000 is recommended for the final analysis. The bootstrap replicates are computationally expensive, but the support values are more stable with more replicates. The number of replicates should be recorded for the reproducibility. ### What is a good posterior probability threshold? A posterior probability of 0.95 is generally considered strong support. The posterior probabilities are higher than the bootstrap values for the same clade. The threshold should be interpreted in the context of the dataset and the model. ### How do I check the convergence of the MCMC? The convergence is checked using the effective sample size and the potential scale reduction factor. The ESS should be above 200 for the parameters of interest, and the PSRF should be close to 1.0. The convergence should be checked for every analysis. ### What should I do if my tree has low support? The low support indicates that the data does not strongly support the clade. The low support can be caused by the lack of informative sites, the recombination, or the model misspecification. The solution is to investigate the cause and to consider the additional data or the different model. ### Can I use the phylogenetic tree for the public health decisions? The phylogenetic tree can be used for the public health decisions, but the results should be interpreted with the epidemiological data. The tree is a hypothesis about the evolutionary relationships, and the support values should be considered. The public health decisions should be based on the combination of the phylogenetic, the epidemiological, and the clinical data. ## Using the Evidence | Source | Best use in this topic | Important limitation | |---|---|---| | [Biosafety in Microbiological and Biomedical Laboratories](https://www.cdc.gov/labs/bmbl/index.html) | official guidance | Check the linked page for current local requirements | | [Microbiology Resources](https://asm.org/) | official guidance | Check the linked page for current local requirements | | [Antimicrobial Resistance](https://www.who.int/health-topics/antimicrobial-resistance) | official guidance | Check the linked page for current local requirements | ## Related Bacteriology Guides - [Bumblefoot in Chickens: Bacterial Pododermatitis and Management](/knowledge/bacteria/avian-bacteria/bumblefoot-chicken-bacteria) - [Chicken Ka Bacteria: Understanding Bacterial Pathogens in Poultry](/knowledge/bacteria/avian-bacteria/chicken-ka-bacteria-poultry-pathogens) - [Bacterial Pathogens in Chickens: Salmonella, E. coli, and Other Avian Bacteria](/knowledge/bacteria/avian-bacteria/bacterial-pathogens-in-chickens-salmonella-e-coli) - [Food Safety in Poultry: Effective Methods to Kill Bacteria in Chicken](/knowledge/bacteria/avian-bacteria/killing-bacteria-in-chicken-food-safety) - [Bacterial Infections in Chickens: Clinical Signs and Diagnosis](/knowledge/bacteria/avian-bacteria/bacterial-infections-chickens-clinical-signs) ## References and Further Reading - [Biosafety in Microbiological and Biomedical Laboratories](https://www.cdc.gov/labs/bmbl/index.html). CDC and NIH. - [Microbiology Resources](https://asm.org/). American Society for Microbiology. - [Antimicrobial Resistance](https://www.who.int/health-topics/antimicrobial-resistance). World Health Organization. - [Bacteriological Analytical Manual](https://www.fda.gov/food/laboratory-methods-food/bacteriological-analytical-manual-bam). U.S. Food and Drug Administration. - [Antimicrobial Susceptibility Testing](https://clsi.org/standards/products/microbiology). Clinical and Laboratory Standards Institute. - [Culture Collections](https://www.atcc.org/). American Type Culture Collection. - [NCBI Pathogen Detection](https://www.ncbi.nlm.nih.gov/pathogens). National Center for Biotechnology Information. - [Evolutionary analysis of human respiratory syncytial virus collected in Myanmar between 2015 and 2018.](https://pubmed.ncbi.nlm.nih.gov/34020068). Infection, genetics and evolution : journal of molecular epidemiology and evolutionary genetics in infectious diseases, 2021. - [Evolutionary dynamics of influenza B strains detected from paediatric acute respiratory infections in central Vietnam.](https://pubmed.ncbi.nlm.nih.gov/32105864). Infection, genetics and evolution : journal of molecular epidemiology and evolutionary genetics in infectious diseases, 2020. - [Analysis of Acanthamoeba genotypes from public freshwater sources in Thailand reveals a new genotype, T23 Acanthamoeba bangkokensis sp. nov.](https://pubmed.ncbi.nlm.nih.gov/34453084). Scientific reports, 2021. - [Differences in the integrase and reverse transcriptase transmitted resistance patterns in Northern Poland.](https://pubmed.ncbi.nlm.nih.gov/28017912). Infection, genetics and evolution : journal of molecular epidemiology and evolutionary genetics in infectious diseases, 2017. > This article is educational and does not replace validated laboratory procedures, clinical diagnosis, institutional biosafety review, or public-health guidance.