Algorithm 29 of the Supplementary Information to Jumper et al. (2021),
p. 37. A small MLP on the single representation produces a distribution
over lDDT bins; the reported confidence is that distribution's mean.
Usage
Alfplddt(s, w1, w2, w3, bins = NULL, rtrue = NULL)
Arguments
- s
Single representation, an n x cs matrix.
- w1, w2
The two projections of line 1, each followed by a relu.
- w3
Projection to the bin logits (line 2).
- bins
Bin centres; defaults to the spec's 1, 3, ..., 99.
- rtrue
Optional ground-truth lDDT per residue, giving the loss.
Value
A list with plddt, the bin distributions p, the
loss, estimate, n and method.
Details
Line 4 of the published pseudocode writes the confidence loss without a
leading minus sign, which would make it something to maximise while
equation (7) adds it to a total that is minimised. The reference
implementation uses the negative log likelihood, so that is what is
computed here; the sign in the supplement is a typo.
References
Jumper et al (2021) Nature 596:583-589, Suppl. Algorithm 29
Examples
set.seed(14)
s <- matrix(rnorm(6), 3, 2)
r <- Alfplddt(s, w1 = matrix(rnorm(4), 2, 2), w2 = matrix(rnorm(4), 2, 2), w3 = matrix(rnorm(100), 50, 2))
round(r$plddt, 1)
#> [1] 50 50 50
