Recycling training objective of AlphaFold

Algorithm 31 and equation (48) of the Supplementary Information to Jumper et al. (2021), pp. 42-43. The objective is the average loss over all recycling iterations. Training never evaluates that average; it samples one iteration \(N'\) uniformly and trains only that one, stopping gradients into the earlier iterations and skipping the later ones entirely.

Usage

Alfrcyl(losses, nprime = NULL)

Arguments

losses

Loss at each recycling iteration.

nprime

Optional one-based iteration selected by training.

Value

A list with estimate, the equation (48) average, expected, ncycle, nprime and method.

Details

The sampled estimate is unbiased, which is the whole justification for the scheme: averaging the single-iteration estimate over every possible \(N'\) returns equation (48) exactly. Nothing is sampled here – pass nprime for one iteration's estimate, or omit it for the average.

References

Jumper et al (2021) Nature 596:583-589, Suppl. Algorithm 31, equation (48)

Examples

V <- c(1, 2, 3, 4, 5, 6, 7, 8)
Alfrcyl(V)
#> $estimate
#> [1] 4.5
#> 
#> $average
#> [1] 4.5
#> 
#> $expected
#> [1] 4.5
#> 
#> $ncycle
#> [1] 8
#> 
#> $nprime
#> NULL
#> 
#> $method
#> [1] "AlphaFold recycling training objective"
#>