Repository navigation
-
EVidenceModeler is pinned to
v2.5.2_rust(hyphaltip/EVidenceModeler; wasv2.5.0_rust), set ininstall_scripts/pixi_install_evm.sh, which the pixi
environments and the container base image both use. The tag adds, over 2.5.0: parallelconvert_EVM_outputs_to_GFF3andrecombine_evm_outputs(--threads),
--skip_emptyand--concat_outfor the converter, buffered output ofevidence_modeler,evm-cliand the converters (the per-write cost was large on
shared filesystems for genomes with hundreds of thousands of contigs), progress logging of the conversion, and a fix that makes the results independent of
hash iteration order (ties between equal-scoring evidence and a floating-point intergenic sum were resolved differently from run to run).funannotate-runEVM.py
passes the new options when the installed build lists them in--help(see 1.9.1-beta.2). v2.5.1_rust holds these changes and v2.5.2_rust only changes the version string. -
funannotate setup: dbCAN isV14(dbCAN HMMdb release 14.0, 08/26/2025, 875 HMMs: 514 families + 5 cellulosome + 356
subfamilies; wasV11, 699 HMMs) and MIBiG is4.0(was1.4).downloads.jsononmasteris what 1.9.x reads at run
time. A database set up from the V11 link was recorded as dbCAN "14.0" while it held the 699 V11 HMMs (the readme next to the
files describes 14.0); with the V14 link the recorded version, 14.0 with date 2025-08-26, now matches the models. Subfamily hits
(for exampleGH2_1) are written as their family (CAZy:GH2), as before. MIBiG 4.0 headers have the same 7|fields that
annotatereads. Re-runfunannotate setup -i dbCAN mibig -f(or-u) to get them. Test:tests/test_downloads_versions.py. -
PASA is pinned to
v2.6.2(hyphaltip/PASApipeline; wasv2.6.1-rc.4). v2.6.2 fixes serious bugs in the PASA annotation
update thatfunannotate updateuses. In every v2.6.1-rc.N,Gene_obj.pmcached the exon order before the strand was set, so
minus-strand multi-exon models got CDS phases counted from the 3' end, and PASA compared scrambled proteins when it decided which
updates to accept. On five Mucorales genomes 6-22% of the updated mRNAs had internal stops (Absidia caerulea RSA 138: 3,535 tbl2asn
InternalStoperrors). v2.6.2 also fixes CDS phases 1 and 2 swapped for every GFF3 model PASA reads and writes back unchanged
(the# ORIGINALmodels; also in upstream PASA), the first-segment phase of 5'-partial models, and.gz/.bz2genomes in
Fasta_retriever.funannotate trainwas not affected. A PASA database made by v2.6.1-rc.N must not be reused byupdate. -
funannotate updatestops when more than 1% of the PASA mRNAs translate with an internal stop (funannotate predictoutput:
6 of 13,344), with a message that names the PASA phase bug and--table. Below that limit, a locus where any PASA mRNA has an
internal stop is replaced by its original model from the GFF3 given to PASA (no UTRs, no new isoforms); only that locus changes.
Merged loci revert to all their original models; new PASA loci with an internal stop are dropped. Before, these models were written
to the GBK/TBL while the log said they were dropped. Test:tests/test_update_pasa_phase.py. -
The training source recorded in the Augustus
parameters.jsonreadBUCSCO <lineage>; it now readsBUSCO <lineage>. -
funannotate predict, BUSCO training: Augustusstop_codonrows are dropped before the BUSCO models go through EVM's
augustus_GTF_to_EVM_GFF3.pl. BUSCO runs Augustus with--stopCodonExcludedFromCDS=False, so the CDS rows already end with the
stop codon, and upstream EvidenceModeler 2.1.0 counted the stop a second time. On the 370 BUSCO models of the test genome its converter wrote 871 CDS rows instead of 503 and
replaced the last internal exon of multi-exon genes with a copy of the stop codon; with thestop_codonrows removed it gives exactly the EVM 1.1.1 output (EVM 2.0.0 was not tested). That corrupted the
training models, which showed up as too few validated BUSCO models, SNAP finding 0 genes and failed Augustus training (issues #1078,
#1071, #1069). The pixi, Docker and conda installs use an EVM build that was not affected; a$EVM_HOMEthat is upstream EVM 2.x was.
getTrainResultsreturnsNoneand training stops with a message when Augustus printed no accuracy report, instead of an
UnboundLocalError(#1069).
BUSCO models that became Complete only in BUSCO's second, retrained Augustus pass have no per-gene GFF; they were never used for training, and are now counted and logged
(N of M BUSCO models were completed only by the retrained Augustus ...) instead of being skipped silently. -
A locus tag prefix that contains
-T(for example--name Al-Tr-26) no longer collapses every gene into one calledAlinannotate: only the
trailing-T<n>transcript suffix is stripped (#1089).funannotate predict --nameis rejected with the offending characters when it holds
anything but letters, digits,_,-and.; curly quotes pasted into it broke bedtools at the end of predict (#1074). -
funannotate annotate: SignalP 5/6 results use the first word of the ID, so secretome notes reach the transcripts when Phobius was not run
(#1110). A second transcript without annotations no longer receives the first transcript's PFAM/MEROPS/dbCAN/GO notes in the tbl file.which()
is False for anyOSError, so a missing program behind an unreadablePATHdirectory is no longer reported as installed (SignalP crash,
#1120). A BUSCO run that produced no result table stops with a message that names the lineage folder (#1108). antiSMASH contig names that
match no contig of the annotation are listed in a warning, and the MIBiG cross reference is skipped when no gene is in a cluster instead of
running diamond on an empty file (#1147).-d/--databaseis also used to findgo.obo(#1059). -
funannotate predict: a failingtrainGlimmerHMM(too few multi-exon training models) skips GlimmerHMM instead of ending the run (#1102).
The Augustus training directory is removed withSafeRemove(PR #1149);funannotate-p2g.py --contig_expandis an integer (PR #1023). -
funannotate setup: the dbCAN version is taken from theV<n>folder of the download link. Thereadme.txtthere now describes release 14.0, so
the V11 models were recorded as 14.0. -
funannotate check: the$EVM_HOMEmessage no longer says EVM 2.x is incompatible: the test only looks forEvmUtils/partition_EVM_inputs.pl, which EVM 2.1.0 also ships, so it never fired for 2.x. -
funannotate predictwith a very large genome (issue #1240, a 30 GB genome): minimap2 builds a multi-part index when the genome is longer than its
-Ipart size and then stops with an error that was hidden because its stderr went to/dev/null. The alignment now passes-Ilarge enough
for a genome over 4 GB, so the index is one part (--split-prefixis not an option: minimap2 refuses it together with--cs, which the BAM
parsers need). The index then needs memory in proportion to the genome. minimap2's stderr goes to<bam>.minimap2.log, which is kept and
reported when minimap2 fails. Smaller genomes run the same command as before. -
The training and evidence assessment (
docs/assessment_pasa2.6_fun1.9/: data, figures, scripts), thebenchmark_*.sbatchscripts and
plans/BENCHMARK_*.mdmoved to https://github.com/hyphaltip/funannotate-manuscript.docs/assessment_pasa2.6_fun1.9.rstis now a pointer. -
funannotate util gff2protandfunannotate util gff-renameno longer crash: they passedtable=togff2dict, whose argument is
transl_table, so every run raised aTypeError.--tablenow takes effect in both. -
funannotate annotate: version checks of emapper and diamond usepackaging.versionthrough a newparse_version()instead of
distutils.version.LooseVersion(distutils is removed in Python 3.12). A version string such as2.1.4-5-gabc1234orv2.0.11still
compares by its leading numbers, and a missing or unparseable string counts as 0. The check that refuses emapper versions above
2.0.0 and below 2.0.5 compared a string with a version object; it now compares two versions. -
funannotate.library.gb2outputis deprecated and emits aDeprecationWarning. Nothing in funannotate calls it, and it fails with a
TypeErrorfor GenBank files with mRNA features. It is removed in 1.10. -
funannotate predict: the choice between PASA models and BUSCO models for training, the minimum-training-models check and the
Augustus held-out test set size are now functions inlibrary.pywith unit tests; behaviour is unchanged. The held-out accuracy of the
trained Augustus parameters is written totraining_decisions.tsv(rowaugustus_accuracy); it is a report and does not switch the
training source. The accuracy table printed after training had the Sensitivity and Specificity headings swapped; they now match the values. -
funannotate predict --genemark_gtfwith contig names that are not in the genome no longer crashes in the recovery code: the models on
those contigs are dropped and the contig names are logged as an error. -
funannotate predict --tableand--mtable: an unassigned NCBI table id is reported before any work starts, and a table that reassigns
stop codons (4, 6, ...) logs a warning that SNAP, GlimmerHMM and CodingQuarry cannot use it.genetic_codes.is_valid_tablerejects
fractional ids such as1.5. -
funannotate check --strict --allow-missing a,baccepts a list of programs that may be missing (for examplegmes_petap.pl,GENEMARK_PATH
for images without GeneMark); other missing programs still fail. -
NCBI's native
table2asnis installed byinstall_scripts/install_table2asn.sh(macOS arm64 and Linux x86_64; it replaces
install_table2asn_macos.sh) and byDockerfile.base. The download's SHA-256 is logged, andTABLE2ASN_SHA256pins it.funannotate annotatealready preferstable2asnand falls back totbl2asn. -
The development environment (
pixidefault environment) includespytest, so the pytest-based tests run with
pixi run -e default python -m pytest tests. Thebaseenvironment used for the container image is unchanged. -
pairwiseAligninfunannotate updateandfunannotate util contrastusesBio.Align.PairwiseAlignerinstead of the
deprecatedBio.pairwise2, so it keeps working when Biopython removespairwise2. The scores are identical, and
both run on the pinned Biopython 1.79.