Rewrite config syntax

[jameshadfield]

When I implemented the configs in #72 I considered using the config structure that this PR introduces but felt it was too different from the previous format (and other configs in the nextstrain ecosystem). After thinking it over in the intervening months I found the time to implement it here and think it massively simplifies things and makes the configs and snakefile much easier to reason with.

Background

This PR adds two large sections to the README which are good background to the changes here.

This google doc discusses a number of different config formats. Of those detailed in that doc I think this one is the only one that scales to workflows such as avian-flu.

Seasonal flu uses a somewhat related syntax for certain builds where we join wildcards together into a single string, but I don't think there's much else similar about the approaches.

The main idea

Avian-flu builds use a three-tiered wildcard setup of subtype, segment and time (for instance subtype=h5n1, segment=ha, time=2y is responsible for nextstrain.org/avian-flu/h5n1/ha/2y). For each config parameter we represent the values as a dict where the keys are a slash-separated string of the wildcards, with "*" as a match-anything symbol. For a given wildcard combination we search the config parameters and pick the most specific match.

So for a parameter which never changed across any wildcard combination you could use "*" for each wildcard:

filter: 
  target_sequences_per_tree:
    "*/*/*": 3000

And if you wanted to have the h5n1 2y builds to use 5,000 sequences, the h5n1 all-time builds to use 10,000 and all the other builds to use 3000 you could write:

filter: 
  target_sequences_per_tree:
    "h5n1/*/all-time": 10000
    "h5n1/*/2y": 5000 # or 'h5n1/*/*'
    "*/*/*": 3000

This is very powerful when different parameters vary across different wildcards, as they do frequently within avian-flu. (I don't know of any other format which can handle this nicely.) For instance the minimum length varies over segment but not subtype or time, so we can write

filter:
  min_length:
    "*/pb2/*": 2100
    "*/pb1/*": 2100
    "*/pa/*": 2000
    ...

whereas the columns for DTA vary over subtype but not segment or time, and we can use the exact same syntax to convey this very different combination:

traits:
  columns:
    "h5nx/*/*": region
    "h5n1/*/*": region country
    "h7n9/*/*": country division
    "h9n2/*/*": region country

Simplifications

This format is flexible enough to express every "special case" in avian-flu. For instance, we no longer have to special-case genome-specific parameters for the cattle-outbreak build (both in the configs & the snakefiles) such as the columns to use for DTA:

# config YAML
- genome_columns: division
- columns: region country
+ columns:
+   "*/genome/*": division
+   "*/*/*": region country

# Snakemake
- columns = get_config(['traits', 'genome_columns'], wildcards) \
-        if wildcards.segment=='genome' \
-        else get_config(['traits', 'columns'], wildcards)
+ columns = resolve_config_value(['traits', 'columns'], wildcards)

Being able to use one syntax everywhere makes it much easier to explain and use config overlays (see #103, which this PR sits on top of).

Refactored config declaration of which builds to run

See these added docs for a summary of the new format. The new structure is both more flexible and more amenable to config overlays as it is a list not a dict and will therefore be overwritten by an overlay. This makes it trivial to run a build for a subset of builds without touching the snakefile or main configs, which is a common occurrence for me given that the GISAID workflow produces 48 trees by default.

Misc

I also ended up allowing the workflow to only target certain intermediate files (e.g. metadata & sequences) not run the full analysis. @lmoncla this was inspired by seeing your commented out additions in 1895865 which is something I've done commonly as well!

[jameshadfield]

If you'd like to test this (please do!) here's a simple example which covers at a conceptual level most of the changes introduced by this PR and #103:

From within the avian-flu repo checkout this PR:

git pull && git checkout 'james/update-config-syntax'
AVIAN_FLU=$( pwd )

And then in a separate analysis directory (outside of the repo) create a YAML file config.yaml:

filter:
  target_sequences_per_tree: 1000 # Note: this is the same as "*/*/*": 1000

builds:
  - subtype: h5n1
    segment:
      - ha
      - na
    time: 2y

traits:
  confidence:
    "*/ha/*": true # infer confidence for HA builds
    "*/*/*": false # not for everything else (the current default)

Then, within this analysis directory, run

snakemake --cores 2 --snakefile ${AVIAN_FLU}/gisaid/Snakefile -pf

This toy example allows us to easily compare the effect of inferring confidence for HA (LHS) vs not for NA (RHS):