Comparison of two skempi-foldx stores, record by record.

== How much changed ==

arm             paired  identical  changed  changed %
single-point      4232       3970      262       6.2%
multi-point       1763        175     1588      90.1%
all               5995       4145     1850      30.9%

'identical' means all twelve terms equal, not merely Interaction Energy.
Unpaired: 8 ambiguous (a name reaching two interface definitions), 0 unmatched.
    3SE4        KA134A                      single-point
    3SE4        KA134R                      single-point
    3SE4        LA31A                       single-point
    3SE4        MA130A                      single-point
    3SE4        PA27A                       single-point
    3SE4        PA27L                       single-point
    3SE3        YB43H,NB44E,SB47Q           multi-point
    3SE3        YB43A,NB44A,SB47A           multi-point

== How large the changes are ==

Shift in Interaction Energy over the records that changed, in kcal/mol.

arm                 n   median     mean      p90      p99      max    <0.5  sign flip
single-point      262    0.052    0.251    0.738    1.875    6.854     84%         17
multi-point      1588    0.242    0.521    1.343    3.691    9.960     67%        112

A sign flip is a record whose Interaction Energy changes sign -- stabilising to
destabilising or the reverse -- which is the change most likely to alter a conclusion.

Largest individual shifts:

    4K71_A_BC   GA503R,AA545V                    11.0524 ->    1.0929   (-9.9595)
    3SE4_B_A    RA226A                            0.0000 ->    6.8535   (+6.8535)
    2C5D_AB_CD  RA32E,RB32E                       0.0977 ->    6.9109   (+6.8132)
    2B2X_HL_A   VH50T,EH64K,QL28S,YL52N,TH33S,GH53W   18.6186 ->   12.3615   (-6.2571)
    3SGB_E_I    KI7R,PI8F,AI9R,TI11K,LI12R        4.5000 ->   -1.7555   (-6.2555)
    1CHO_EFG_I  EI7D,AI12V,TI14S,EI16D,RI18M,KI26T,GI29N,NI33I,EI40D,LI45I    2.1967 ->    7.6777   (+5.4810)
    2C5D_AB_CD  LC105E,LD105E                     0.0000 ->    5.1289   (+5.1289)
    4MNQ_ABC_DE LC6A,HE50W,PE51E,EE52F,YE53E     -0.6166 ->    3.8215   (+4.4381)
    1CHO_EFG_I  EI7D,LI15V,RI18M,EI40D            2.4846 ->    6.6671   (+4.1825)
    3L5X_A_HL   HL91Q,NL92D,EL93Y,YL94G          -0.0102 ->   -4.1759   (-4.1657)
    4MNQ_ABC_DE KC4A,HE50W,PE51E,EE52F,YE53E      0.7801 ->    4.8538   (+4.0737)
    1MHP_HL_A   TH33E,GH53Q                       7.3537 ->    3.4379   (-3.9158)
    1PPF_E_I    LI18M,TI30S,GI32D                 9.7735 ->   13.6190   (+3.8455)
    2C5D_AB_CD  TA26D,TB26D                       2.9207 ->    6.7259   (+3.8052)
    1A22_A_B    NA12R,MA14V,LA15V,RA16L,RA19Y,FA25S,DA26E,QA29S,EA30Q,EA33K   -3.0260 ->    0.7343   (+3.7603)

Complexes where the most records moved:

    identifier     records  changed  median |d|   max |d|
    1JTG_A_B           194      136       0.263     2.421
    2B2X_HL_A           94       91       0.214     6.257
    3SGB_E_I           279       88       0.370     6.255
    1CHO_EFG_I         275       84       0.431     5.481
    1DAN_HL_UT         114       81       0.024     1.804
    1R0R_E_I           267       75       0.379     2.189
    3S9D_A_B           142       59       0.147     2.115
    1PPF_E_I           246       56       0.458     3.845
    1AO7_ABC_DE        151       56       0.132     1.427
    3HG1_ABC_DE         53       48       0.560     2.645
    1BRS_A_D            75       45       0.350     1.495
    4RS1_A_B            49       41       0.063     1.132
    4G0N_A_B            72       40       0.217     2.220
    3HFM_HL_Y           91       38       0.183     1.315
    3L5X_A_HL           35       34       0.932     4.166

Movement by term, over the records that changed:

    term                         moved  median |d|   max |d|
    Interaction Energy            1837      0.2113    9.9595
    Backbone Hbond                1613      0.0272    2.5088
    Sidechain Hbond               1501      0.1415    7.1323
    Van der Waals                 1765      0.0781    2.3492
    Electrostatics                1735      0.0525    3.0395
    Solvation Polar               1758      0.1528    4.3162
    Solvation Hydrophobic         1760      0.0915    3.4731
    Van der Waals clashes         1693      0.0552    9.8492
    entropy sidechain             1791      0.0828    3.6507
    entropy mainchain             1633      0.0396    2.0414
    torsional clash               1642      0.0094    2.0085
    backbone clash                1622      0.0321    0.6414

== What characterises the records that moved ==

Whether a record changed is a property of its interface definition: a record moves exactly
when the list it was computed in changed composition, and a list belongs to a definition.

    single-point: 29 of 323 paired definitions hold at least one changed record, and between them hold all of them (3 changed entirely)
        unchanged definitions hold at most 240 mutations (106 hold exactly one -- a one-entry list is
        the same list in both releases, by construction)
    multi-point: 87 of 153 paired definitions hold at least one changed record, and between them hold all of them (54 changed entirely)
        unchanged definitions hold at most 4 mutations (46 hold exactly one -- a one-entry list is
        the same list in both releases, by construction)

    list length vs the fraction of a definition that moved:

        arm           length     definitions   mean fraction changed
        single-point  1-5                184                    0.04
        single-point  6-20                88                    0.07
        single-point  21-60               40                    0.06
        single-point  61+                 11                    0.08
        multi-point   1-5                 86                    0.19
        multi-point   6-20                45                    0.89
        multi-point   21-60               16                    0.96
        multi-point   61+                  6                    0.99

How far a record moved tracks side-chain repacking freedom. Quartiles of the clash
energy the record already carried in the older release:

        arm           quartile                  n  median |shift|
        single-point  Q1 (least clash)         65           0.010
        single-point  Q2                       65           0.039
        single-point  Q3                       65           0.055
        single-point  Q4 (most clash)          67           0.183
        multi-point   Q1 (least clash)        397           0.095
        multi-point   Q2                      397           0.188
        multi-point   Q3                      397           0.317
        multi-point   Q4 (most clash)         397           0.478

    multi-point: shift by number of substitutions

        substitutions         n  median |shift|
        2                   730           0.164
        3                   219           0.261
        4                   211           0.263
        5                   153           0.331
        6                    66           0.373
        7                    33           0.355
        8+                  176           0.537

    single-point: shift by the rotamer freedom of the residue introduced

        chi angles            n  median |shift|
        0                   170           0.028
        1                    11           0.055
        2                    47           0.058
        3                    19           0.359
        4                    15           0.198

== Whether the change matters ==

Each release against SKEMPI's measured ddG, on the records the two share. A model
trained on the older values is affected by the difference only to the extent that
these disagree.

Split three ways, because the records identical in both releases score identically by
construction and dilute the comparison: the middle block is where any real difference
has to show up, and the last is the control that must show none.

arm           subset                release       n  Pearson r  Spearman rho     MAE
single-point  all shared records    0.1.0      4053      0.419         0.435   1.138
single-point  all shared records    0.2.0      4053      0.418         0.436   1.138
single-point  changed in 0.2.0      0.1.0       257      0.416         0.374   0.904
single-point  changed in 0.2.0      0.2.0       257      0.393         0.392   0.909
single-point  identical in both     0.1.0      3796      0.420         0.438   1.154
single-point  identical in both     0.2.0      3796      0.420         0.438   1.154
multi-point   all shared records    0.1.0      1576      0.515         0.574   1.885
multi-point   all shared records    0.2.0      1576      0.518         0.587   1.871
multi-point   changed in 0.2.0      0.1.0      1418      0.499         0.557   1.946
multi-point   changed in 0.2.0      0.2.0      1418      0.501         0.569   1.931
multi-point   identical in both     0.1.0       158      0.585         0.430   1.332
multi-point   identical in both     0.2.0       158      0.585         0.430   1.332

Paired bootstrap on the change in Spearman, 1000 resamples, records drawn
together so each keeps its pair:

arm           subset                  delta rho          95% interval
single-point  all shared records         +0.001      [-0.002, +0.004]
single-point  changed in 0.2.0           +0.018      [-0.036, +0.075]
multi-point   all shared records         +0.012      [-0.005, +0.029]
multi-point   changed in 0.2.0           +0.012      [-0.006, +0.031]

An interval spanning zero says these records do not separate the two releases.

And the two releases against each other, which is how much a feature column moves:

arm                 n   Pearson r   Spearman rho      MAE
single-point     4232       0.996          0.994    0.016
multi-point      1763       0.945          0.944    0.469

