MAB skipping function: still broken in v2022.6?

[Hayden Scheiber]

Hello,

At the time of this posting, I am using the latest version of WESTPA, v2022.06. I am employing MAB bins in my WESTPA calculations. I was aware that prior to v2022.05, the skip function of the MAB scheme was broken, but apparently fixes were made in v2022.05 to alleviate this. However, I believe the MAB skip function is still not working as intended in version v2022.06.

I have run two nearly identical WESTPA calculations using 2D progress coordinates with and without the MAB skip function. I've pasted the binning scheme from my west.cfg file below. For context, this is a simulation of an antigen/antibody complex unbinding. The first pcoord is the total solvent accessible surface area (SASA) of the epitope + paratope regions, while the second pcoord is the protein-protein distance (i.e. the smallest distance between any two atoms of the two proteins).

Binning scheme:

    system_options:

      # Dimensionality of your progress coordinate

      pcoord_ndim: 2

      # Number of data points per iteration

      # Needs to be pcoord_len >= 2 (minimum of parent, last frame) to work with most analysis tools

      pcoord_len: 21

      # Data type for your progress coordinate

      pcoord_dtype: !!python/name:numpy.float32

      bins:

        type: RecursiveBinMapper

        base:

            type: RectilinearBinMapper

            boundaries:

                - [0, 15, inf] # Binding SASA (nm^2, typical values 5 - 20)

                - [0, 5, inf]  # Min P-P distance (Angstroms)

        mappers:

          - type: MABBinMapper # Prior to dissociation

            direction: [1, 0]

            skip: [0, 1] # Skip MAB binning along P-P distance direction

            nbins: [7, 1]

            at: [1.0, 1.0]

            mab_log: true

          - type: MABBinMapper # After dissociation

            direction: [0, 1]

            skip: [1, 0] # Skip MAB binning along SASA direction

            nbins: [1, 3]

            at: [20.0, 1.0]

            mab_log: true

      # Number walkers per bin

      bin_target_counts: 32

In my first set attempt, I used the above binning scheme with the skipping enabled as shown. As suggested by the comments, this was intended to skip MAB binning in the first MAB bin along the P-P distance coordinate, and then skip MAB binning along the SASA pcoord in the second MAB bin. In both cases, I initialize the system from 64 basis states with 8 segs-per-state. These bstates are all bound configurations, differing only slightly in configuration. 

With the above MAB scheme including MAB skipping:

1 target state(s) present

Calculating progress coordinate values for basis states.

64 basis state(s) present

Calculating progress coordinate values for start states.

0 start state(s) present

Preparing initial states

################ MAB stats ################

minima in each dimension:      [3.9172401, 1.5129446]

maxima in each dimension:      [6.704759, 2.070519]

direction in each dimension:   [1, 0]

skip in each dimension:        [0, 1]

###########################################

        Total bins:            19

        Initial replicas:      64 in 2 bins, total weight = 1

        Total target replicas: 608

       

1-prob: 0.0000e+00

Simulation prepared.

2 of 19 (10.526316%) active bins are populated

per-bin minimum non-zero probability:       0.015625

per-bin maximum probability:                0.984375

per-bin probability dynamic range (kT):     4.14313

per-segment minimum non-zero probability:   0.000488281

per-segment maximum non-zero probability:   0.0585938

per-segment probability dynamic range (kT): 4.78749

norm = 1, error in norm = 0 (0*epsilon)

Updating system with the options from the configuration file

The exact same calculation with both MAB bin skips turned off:

1 target state(s) present

Calculating progress coordinate values for basis states.

64 basis state(s) present

Calculating progress coordinate values for start states.

0 start state(s) present

Preparing initial states

################ MAB stats ################

minima in each dimension:      [3.9172401, 1.5129446]

maxima in each dimension:      [6.704759, 2.070519]

direction in each dimension:   [1, 0]

skip in each dimension:        [0, 0]

###########################################

        Total bins:            23

        Initial replicas:      256 in 8 bins, total weight = 1

        Total target replicas: 736

       

1-prob: 0.0000e+00

Simulation prepared.

8 of 23 (34.782609%) active bins are populated

per-bin minimum non-zero probability:       0.015625

per-bin maximum probability:                0.328125

per-bin probability dynamic range (kT):     3.04452

per-segment minimum non-zero probability:   0.000488281

per-segment maximum non-zero probability:   0.0195312

per-segment probability dynamic range (kT): 3.68888

norm = 1, error in norm = 0 (0*epsilon)

Updating system with the options from the configuration file

Why are so few bins filled when the skipping function is enabled? I would have expected 7 bins filled to begin with, spanning the first outer bin. Even after 50 WE iterations, once trajectories have diffused into both MAB bins, only 4 out of 19 bins are occupied!

Iteration 50 (50 requested)

Beginning iteration 50

128 segments remain in iteration 50 (128 total)

4 of 19 (21.052632%) active bins are populated

per-bin minimum non-zero probability:       2.595e-46

per-bin maximum probability:                1

per-bin probability dynamic range (kT):     104.965

per-segment minimum non-zero probability:   8.10937e-48

per-segment maximum non-zero probability:   0.0576782

per-segment probability dynamic range (kT): 105.578

norm = 1, error in norm = 0 (0*epsilon)

MAB binning in use

Bottleneck bin occupancy may not be accurately reported

Waiting for segments to complete...

################ MAB stats ################

minima in each dimension:      [3.9939365, 1.5525619]

maxima in each dimension:      [14.99696, 2.5262394]

direction in each dimension:   [1, 0]

skip in each dimension:        [0, 1]

###########################################

################ MAB stats ################

minima in each dimension:      [15.044079, 1.7822143]

maxima in each dimension:      [17.444405, 2.75611]

direction in each dimension:   [0, 1]

skip in each dimension:        [1, 0]

###########################################

Iteration completed successfully

Iteration wallclock: 0:11:53.871709, cputime: 12:09:49.575783

On the other hand, with the MAB skipping disabled, only 23 iterations later I find 19 of 23 bins filled:

Iteration 23 (50 requested)

Beginning iteration 23

608 segments remain in iteration 23 (608 total)

19 of 23 (82.608696%) active bins are populated

per-bin minimum non-zero probability:       2.81192e-23

per-bin maximum probability:                0.480124

per-bin probability dynamic range (kT):     51.1919

per-segment minimum non-zero probability:   8.78725e-25

per-segment maximum non-zero probability:   0.0265375

per-segment probability dynamic range (kT): 51.7621

norm = 1, error in norm = 0 (0*epsilon)

MAB binning in use

Bottleneck bin occupancy may not be accurately reported

Waiting for segments to complete...

################ MAB stats ################

minima in each dimension:      [4.177698, 1.5353204]

maxima in each dimension:      [14.98888, 2.5825822]

direction in each dimension:   [1, 0]

skip in each dimension:        [0, 0]

###########################################

################ MAB stats ################

minima in each dimension:      [15.004696, 1.6985209]

maxima in each dimension:      [19.229214, 3.457224]

direction in each dimension:   [0, 1]

skip in each dimension:        [0, 0]

###########################################

Iteration completed successfully

Have I made a mistake somewhere, or is it that the MAB skipping code is not functioning as intended?

Cheers,
Hayden

[Anthony Bogetti]

Hello Hayden,

A while back we found a bug in the MAB skip code for binning schemes with >1 dimensions and thought we have fixed it. It appears (based on your thorough post and some other internal tests) that this bug has not actually been fixed. We will coordinate to try and have this part of the code fixed in a timely manner so that it generates the results you are expecting.

Best,

Anthony

[Hayden Scheiber]

Hi Anthony,

Glad to hear you are working on resolving this bug. 

It's not a major problem for my simulations as I can just disable MAB skipping, but definitely worth fixing as it can help improve efficiency when employing high dimensional pcoords!

Best,

Hayden