Every structure file below was confirmed real (multiple atoms per residue, not the 1-atom-per-residue
mock format).
01 Linker Optimization — PETase / Dockerin Fusion
Goal: find a linker that pushes the dockerin domain away from the catalytic triad, since the native
fusion crowds the active site and cuts activity roughly 6×. Larger distance is better.
outputs/results/ranked_linkers.csv — real ESM2 + ESMFold scoring
Candidate
Sequence
Len
ESM2 score
pLDDT
Triad→dockerin distance
Composite
GS_len25
GGGGSGGGGSGGGGSGGGGSGGGGS
25
-0.177
80
11.2 Å
68.9
GS_len20
GGGGSGGGGSGGGGSGGGGS
20
-0.271
80
9.3 Å
58.2
GS_len15
GGGGSGGGGSGGGGS
15
-0.613
70
16.0 Å
52.9
Native (baseline)
TSGGGDDGGSGG
12
-1.700
70
17.8 Å
40.0
Read this before trusting the ranking
Every GS-repeat candidate produced a shorter real distance than the native linker (9.3–16.0 Å
vs. 17.8 Å) — the opposite of the intended effect. GS_len25 still wins on composite score because its
ESM2 plausibility and pLDDT are strong, but by the distance metric this pipeline was actually built to
optimize, it's the weakest-performing option after GS_len20. ESMFold is known to predict flexible,
repetitive glycine-serine loops as artificially compact — this may be a modeling artifact rather than
true biology, and is worth confirming with a different structure predictor or in the wet lab before
trusting it.
fig — baseline (native linker) real ESMFold structurefig — GS_len25 (top composite score) real ESMFold structure
fig — composite score by candidate (outputs/results/linker_comparison.png)
02 MHETase Pocket Mutants — TfCa Enzyme
Goal: widen the substrate-binding pocket so MHET clears faster (currently only 32% converts to TPA
after 96h). Candidates are ranked by ESM2 plausibility, fold confidence, and pocket free volume.
outputs/results/ranked_mutants.csv — real ESM2 + ESMFold scoring
Construct
ESM2 score
pLDDT
Pocket free volume
Composite
WT (native)
-0.220
79
3551 ų
60.0
TfCaWA (control)
-0.750
78
3551 ų
23.5
mut_I69F
-2.592
79
3551 ų
23.2
mut_I69F_V376A
-2.656
79
3551 ų
21.1
This ranking is highly arbitrary and dependent on software design assumptions
Pocket free volume is identical — 3550.8 ų — for all four constructs, WT included. The metric
only counts backbone Cα atoms within a fixed sphere around the catalytic triad, treating every residue
as the same fixed-size sphere regardless of side chain. A single point mutation rarely shifts the
predicted backbone enough to change that count, so this metric cannot see the very effect it's meant
to measure. It contributes exactly 0 points to every row's composite score here — the ranking
above is effectively driven by pLDDT and ESM2 score alone. Separately, WT's ESM2 score (−0.22) and
TfCaWA's (−0.75) are hardcoded constants in the notebook's source code, not computed values — only
the two engineered mutants' scores are real ESM2 output. Net effect: WT "winning" the composite ranking
here reflects these gaps, not a genuine finding that the wild type is better than the engineered
mutants.
fig — WT TfCa-DoG real ESMFold structurefig — mut_I69F real ESMFold structure
fig — composite score by construct (outputs/results/pocket_comparison.png)
03 Bottom Line
The pipeline is mechanically sound — both notebooks produce genuine ESM2/ESMFold predictions,
verified end to end. But neither ranked table, taken at face value, supports a clear design
decision yet. The linker table's own headline metric (distance) argues against its own winner;
the mutant table's headline metric (pocket volume) doesn't discriminate at all. Before picking a
construct to synthesize, it's worth fixing pocket_free_volume() to account for actual side-chain
identity (not just Cα count), replacing the two hardcoded ESM2 placeholders with real computed scores,
and treating the GS-linker distance results as provisional pending a second structural check.
