henhao
Sign in
Evolutionary BiologyMolecular Biology and Evolution
Peer-reviewed journal article

MEGA12: Molecular Evolutionary Genetic Analysis version 12 for adaptive and green computing

Sudhir KumariD, Glen StecheriD, Michael SuleskiiD, Maxwell D. SanderfordiD, Sudip SharmaiD, Koichiro TamuraiD

Abstract

We introduce the 12th version of the Molecular Evolutionary Genetics Analysis (MEGA) software. This latest version brings many significant improvements by reducing the computational time needed for selecting optimal substitution models and conducting bootstrap tests on phylogenies using maximum likelihood (ML) methods. These improvements are achieved by implementing heuristics that minimize likely unnecessary computations. Analyses of empirical and simulated datasets show substantial time savings by using these heuristics without compromising the accuracy of results. MEGA12 also implements an evolutionary sparse learning approach to identify fragile clades and associated sequences in evolutionary trees inferred through phylogenomic analyses. In addition, this version includes fine-grained parallelization for ML analyses, support for high-resolution monitors, and an enhanced Tree Explorer. The MEGA12 beta version can be downloaded from https://www.megasoftware.net/beta_download.

PDFDOICodeData
Full-text digest ↓

02 /Specification & Classification

Click any tag to browse matching papers
Citations
2,745
FWCI
355.15
Field percentile
top 1%
OpenAlex Topics

In-depth summary

✦ Multi-agent verified full-text synthesis

Claude Opus 5.5 · max effort · 8 Nov 2026 · from the full text of version 1 (15 Dec 2024) · CC BY 4.0

MEGA12 adds filtered substitution-model selection, matching full-search choices in all 240 simulated nucleotide alignments, and adaptive bootstrapping with an 81% average speed-up, accelerating maximum-likelihood phylogenetics.

Maximum-likelihood selection among 24 substitution-model combinations and bootstrap tests often run with 500-2000 replicates are slow, and summed over MEGA's large user base they carry substantial energy and carbon costs. MEGA12 prunes likely suboptimal models early and sizes bootstrap replicate numbers to a target precision while preserving accuracy, and adds DrPhylo to flag concatenation-based clades that are fragile despite high support.

Key results

  1. ShownFiltered model selection recovered the exhaustive Full search's best-fit nucleotide substitution model in every simulated test, saving the most runtime when a complex model fit best, because nested simple best-fit models leave complex candidates unpruned.

    240 simulated DNA alignments (4 to 289 sequences, 186 to 18,171 sites; ten per each of 24 generating models), default BIC/AICc thresholds of 5: same best-fit model as Full in all 240 (100% concordance); runtime reduction reached 70% for complex best-fit models; Full-search time rose linearly with sequences x site configurations. · Adaptive computing in selecting the optimal…; Data sets analyzed; Options for analyses conducted; Figure legends

  2. ShownFor protein alignments, the Filtered option (adding an empirical-frequency +F screening step) sharply cut model-selection time and returned the Full-analysis best-fit model for most chloroplast protein domains and for their concatenation.

    45 chloroplast protein alignments (31-1388 amino acids, 10 species): ~87% time savings vs Full (Filtered-vs-Full time regression slope 0.18); identical model for 43 (95% concordance), a ΔBIC < 10 alternative for one, second-best model for the other. Concatenated alignment (11,039 amino acids): 3.67 vs 27.12 minutes, same model. · Adaptive computing in selecting the optimal…; Data sets analyzed; Figure legends

  3. ShownAdaptive bootstrapping (start at 25 replicates, add more until every clade's support standard error falls below a user-set threshold, default 5%) closely reproduced 500-replicate support values at a fraction of the runtime; the clade nearest 50% support set the replicate count.

    240 simulated datasets: adaptive vs 500-replicate bootstrap support (BS) values gave slope 0.99 through the origin (R^2 = 0.99); 25 to 124 replicates were needed, correlating -0.96 with minimum |BS - 50%|; speed-ups averaged 81% (range 61% to 95%; time-ratio slope ~0.20). Drosophila Adh example stopped after 87 replicates. · Adaptive Bootstrapping; Options for analyses conducted; Figure legends

  4. SuggestedDrPhylo, an evolutionary sparse-learning tool from MyESL now launchable within MEGA12, flagged a fungal clade with 100% bootstrap support as fragile and pinpointed the gene driving the contested placement of Ascoidea rubescens.

    Fungal maximum-likelihood tree (86 species, 1,232 concatenated genes; Shen et al. 2017): clade of A. rubescens plus 43 species had 100% bootstrap support but clade probability 0.27 owing to A. rubescens; BUSCOfEOG7W9S51 most strongly supported that placement, echoing earlier likelihood tests, while BUSCOfEOG7TN012 was strongly discordant, as its gene tree confirmed. · Integration of the DrPhylo application to assess…

  5. SuggestedFine-grained, site-level multithreading of likelihood calculations and removal of code bottlenecks (hash-table site-pattern mapping, redundant initializations) further accelerate maximum-likelihood computations, though thread scaling is sub-linear.

    Authors' tests, without reported datasets or timings: slightly under 50% efficiency with four threads vs one thread, limited by thread-distribution overhead and terminal nodes with few site configurations; a hash-table site-pattern map reportedly makes that initialization orders of magnitude faster. · Other improvements for phylogenetic analysis…; Elimination of computational bottlenecks

Methods, data and reproducibility

MEGA12's Filtered heuristic drops base substitution models whose BIC/AICc exceed the minimum by more than 5 points before testing +I/+G (and +F for proteins) variants; Adaptive bootstrapping adds replicates beyond 25 until every clade's support standard error is below 5%. Both were benchmarked against Full model search and 500-replicate bootstraps.

  • Simulated nucleotide alignments from a previously published study, ten drawn per each of 24 generating substitution models (4 to 289 sequences, 186 to 18,171 sites) · n = 240 datasets
  • Chloroplast protein alignments split at domain boundaries from the Chloroplast_Martin.meg concatenated example (31-1388 amino acids each; 11,039 amino acids concatenated) · 10 species (not named in text) · n = 45 protein alignments
  • Fungal maximum-likelihood phylogeny and concatenated gene alignment from Shen et al. 2017, used to demonstrate DrPhylo on the clade containing Ascoidea rubescens · fungi (86 species) · n = 86 species; 1,232 genes
  • Plant maximum-likelihood tree and per-gene alignments from Boachon et al., named in Methods as the DrPhylo dataset (differs from the fungal example in Results) · plants (52 species) · n = 520 protein-coding genes; 52 species
  • Drosophia_Adh.meg example alignment shipped with MEGA12, used to illustrate adaptive bootstrapping and support ranges in Tree Explorer · Drosophila (species not specified)
  • Anonymous opt-in usage reports (Windows GUI) and download records for MEGAX and MEGA11, January 2023 to September 2024
Methods & tools
MEGA12 (GUI and megacc command-line version), Filtered substitution model selection (BIC, AICc), Nucleotide models GTR, HKY, TN93, T92, K2P, JC with +I/+G, Adaptive and standard bootstrap, DrPhylo (evolutionary sparse learning), MyESL, Neighbor-joining and maximum parsimony (SPR) initial trees, Evolutionary Probabilities (EP) analysis, RelTime, Muscle, Tree Explorer and Tree Topology Editor, Caption Expert
Code
https://github.com/KumarMEGALab/MEGA-source-codehttps://www.megasoftware.net/beta_downloadhttps://www.megasoftware.net/
Data & accessions
https://figshare.com/s/0413dd262c2ed9df1bd2
Software & environment
MEGA12 (beta for MS Windows); model selection and bootstrap benchmarks used the megacc command-line version on a single thread, DrPhylo runs used the GUI; Windows 10 (64-bit); defaults of BIC/AICc threshold 5 and bootstrap standard-error threshold 5%.
Assay conditions & compute
Single Windows 10 desktop (64-bit) with 64 GB RAM and eight logical cores clocked at 3.36 GHz; e.g., Filtered model selection on the 11,039-amino-acid chloroplast concatenation took 3.67 min vs 27.12 min for Full.
Reproducibility gaps
GitHub repository currently holds MEGA11 source, with MEGA12 code to follow; the MEGA12 beta is MS Windows only, with macOS in early testing and Linux releases planned afterward; final packages promised at publication.
Limitations (5)
  • Filtered model selection saves less time when a simple model fits best, because simpler models nest within complex ones that then survive early screening; overall gains depend on the complexity of the best-fit model. · stated by the authors
  • Adaptive support values scatter around 500-replicate estimates because of the 5% standard-error stopping rule, which leaves values near 50% less precise; realized speed-ups depend on data properties and bootstrap stability. · stated by the authors
  • Multithreaded likelihood computation scales sub-linearly (slightly under 50% efficiency at four threads) owing to thread-distribution overhead and the few site configurations present at terminal nodes. · stated by the authors
  • Methods describe the DrPhylo test as a plant tree (520 genes, 52 species; Boachon et al.), whereas Results present a fungal tree (86 species, 1,232 genes; Shen et al. 2017), leaving the analyzed dataset ambiguous; some figure cross-references (Fig. 4 vs Fig. 5) also mismatch. · noted in this summary
  • Accuracy and speed were benchmarked on simulated DNA alignments of at most 289 sequences and one 10-species protein set, run single-threaded; no large phylogenomic benchmarks or direct energy or carbon measurements support the green-computing framing. · noted in this summary

Who should read:Read this if you run maximum-likelihood model selection or bootstrap tests in MEGA and want evidence on how the Filtered and Adaptive options trade runtime against accuracy, or want to probe fragile clades with DrPhylo.

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISTapo '26Czech '18Kumar '26Hoang '18Morel '18Prillo '23
59% semantic matchBioinformatics
Chris Tapo, Qiyun Zhu·bioRxiv·Jul 29, 2026