Contiguous residue cropping of AlphaFold training features

Supplement section 1.2.8 of Jumper et al. (2021), pp. 7-8. Training crops the residue dimension of every feature to one contiguous window. Because the window is contiguous and shared by all features, pair features must be cropped on both axes with the same index set, which is what keeps residue i of the cropped target aligned with row i of the cropped pair tensor.

Usage

Alfcrop(
  seqlen,
  cropsize,
  start = 1,
  target = NULL,
  pair = NULL,
  msa = NULL,
  mode = "clamped"
)

Arguments

seqlen

Number of residues before cropping.

cropsize

Number of residues to keep.

start

One-based index of the first residue kept.

target

Optional per-residue feature matrix to crop by row.

pair

Optional pair feature matrix to crop on both axes.

msa

Optional MSA feature matrix, cropped along its residue axis.

mode

Either "clamped" or "unclamped".

Value

A list with the kept indices idx, the cropped target, pair and msa, startmax, estimate and method.

Details

The crop start is an argument, not sampled here, so the function is deterministic. The valid range the spec would sample from is returned alongside the crop.

References

Jumper et al (2021) Nature 596:583-589, Suppl. section 1.2.8

Examples

Alfcrop(seqlen = 5L, cropsize = 5L)
#> $idx
#> [1] 1 2 3 4 5
#> 
#> $target
#> NULL
#> 
#> $pair
#> NULL
#> 
#> $msa
#> NULL
#> 
#> $startmax
#> [1] 1
#> 
#> $estimate
#> [1] 5
#> 
#> $mode
#> [1] "clamped"
#> 
#> $method
#> [1] "AlphaFold contiguous residue cropping"
#>