AMAP

Docs / How it works

How it works

AMAP does not annotate genomes from scratch. It takes an annotation that already exists, compares it against one that is trusted, and adjudicates the differences.

The premise is that de novo annotators are good at finding genes and less good at their exact boundaries. Bakta will call a start codon a few codons late, or split one gene into two, or invent a short ORF that is not really there. If a curated reference for a close relative exists, those errors are visible: align the reference proteins onto the new genome, and where the two disagree, decide which is more likely right.

On MTBC genomes that adjudication changes roughly 30% of genes.

The three inputs to a decision#

The Bakta annotation. Every CDS Bakta called on your genome, with coordinates and a locus tag.

The Miniprot alignments. Every reference protein aligned onto your genome, with the coordinates where it landed. The reference proteins carry their own coordinates in their headers (Rv0001_1-1524), which lets a later step tell a real locus from a paralogous placement.

The false-positive set. Proteins that a de novo annotator calls on the reference genome itself but that are not in the curated annotation. Bakta's characteristic mistakes, in other words. For MTBC this is a curated file; with a custom reference AMAP derives it by annotating the reference and taking the difference.

The three inputs over the first 12.4 kb of CDC1551: Bakta's calls, the H37Rv proteins miniprot placed, and two of the calls Bakta makes on H37Rv itself that its curated annotation does not have, aligned here. Point at a gene for its coordinates.

The pipeline#

Ten steps, from genomes to report:

  1. eggNOG database setup, skipped if it is already built
  2. Lineage typing with pathotypr, MTBC only
  3. Annotation with Bakta
  4. Protein alignment with Miniprot
  5. Case classification: every CDS sorted by how it relates to the reference
  6. Correction: conflicts resolved, missing genes rescued, stop codons checked
  7. Functional annotation with eggNOG-mapper
  8. Duplication analysis: within and across genomes
  9. Pangenome with pangene
  10. Report

The workflow page walks through each with its inputs and outputs.

The correction itself#

Steps 5 and 6 are what makes AMAP more than a wrapper. Every CDS is sorted into one of five outcomes:

CaseMeaning
Case 1Bakta called it, the reference has nothing there. Potentially novel.
Case 2Coordinates match the reference exactly. Kept.
Case 3Coordinates match but alignment length differs. Kept.
Case 4Overlaps a reference gene, but the coordinates conflict. Must be resolved.
MissingThe reference has a gene there, Bakta missed it. Rescued.
Bakta's callreference protein, alignedwhat AMAP keeps
Case 2Perfect matchSame start, end, strand and length as an aligned protein. Kept, tied to it.Step through itCase 3Imperfect matchSame start, end and strand; the alignment's length differs. Kept, tied to the protein.The ruleCase 4Conflicting overlapOverlaps an aligned protein with other coordinates. The reference's win, unless the call is much longer.Step through itCase 1Potentially novelNothing aligns. Kept, unless Bakta is known to call it wrongly or the reference genome has its protein.Step through itMissingPut backA reference protein aligns where Bakta called nothing. It goes in at the alignment.Step through it
And the harder ones
The five outcomes, each on real genes of CDC1551: the same windows as the case pages, drawn small. Each card opens its story, step by step.

Cases 1 and 4 are where the work is: deciding which novel genes are real, and which set of coordinates wins a conflict. The correction algorithm gives the rules and the thresholds.

In this section#

Edit this page on GitHub