iGEM · PETosome Engineering — Full Explainer

What This Pipeline Actually Found, and What To Do Next

What can we infer from the ESM2/ESMFold results?

1. Background — what is a PETosome?

This project includes a PETosome — a synthetic enzyme complex that does this breakdown in two steps, like two workers on an assembly line, both physically attached to a shared scaffold protein so they work side-by-side:

PETase (the enzyme "ICCG") chops long PET polymer chains into smaller fragments called MHET and BHET.
MHETase (the enzyme "TfCa") finishes the job, converting MHET into TPA and ethylene glycol — the two raw materials PET is originally made from.

Both enzymes are attached to the scaffold through a small "hook" domain called a dockerin, connected to the enzyme by a short, flexible chain of amino acids called a linker.

Two real problems, from real wet-lab data

Problem 1 (this report's Part 1): attaching the dockerin to PETase cut its activity 6× on a simple test substrate, and reduced TPA production by 44% on real PET film — much worse than the 7% reduction seen with a smaller substrate (BHET). The team's hypothesis was that the dockerin had been physically crowding the enzyme's active site, and bigger substrates might suffer more because they need more room.

Problem 2 (this report's Part 2): after 96 hours, the PETase alone produced 67% MHET but only 32% TPA — meaning MHETase is the slower, bottleneck enzyme. The team designed one hand-picked mutant (called TfCaWA) to try to fix this, but hadn't tested it thoroughly or explored alternatives systematically.

This pipeline uses AI tools to generate and rank candidate fixes for both problems before spending wet-lab time and money synthesizing and testing them — a computational triage step.

2. The tools — what ESM2 and ESMFold actually do

Both tools come from the same family of AI models, originally developed by Meta AI, trained on huge databases of real protein sequences.

ESM2 — "does this sequence look like a real protein?"

ESM2 is a protein language model — the same basic idea as a large language model for text, except trained on protein sequences instead of English. It learned, from millions of real proteins, which amino acid sequences tend to occur, and in what contexts. Given a sequence with one region blanked out, it can estimate how "natural" different fill-ins would be — similar to how a text model can guess a missing word from context. This pipeline uses that ability to score candidate linkers and mutations: a higher ESM2 score means the candidate looks more like a real, evolutionarily plausible protein sequence in that position — it does not mean the enzyme will actually work better.

ESMFold — "what 3D shape will this sequence fold into?"

ESMFold takes a full amino acid sequence and predicts its 3D structure — where every atom sits in space. It's built on top of ESM2's learned representations, which lets it skip a slow step that older tools like AlphaFold2 require (searching large databases for similar sequences). That makes ESMFold much faster, at some cost in accuracy — a trade-off that is expected to make these findings conjectural rather than definitive.

Every ESMFold prediction comes with a per-residue confidence score called pLDDT (0–100 scale; above 70 is generally considered reliable, below 50 is low-confidence). This lets you tell not just what shape the model predicts, but how much to trust each part of that shape.

Integrity check on this run

Both notebooks were run end to end with real ESM2/ESMFold (not the placeholder/mock mode built into the code for quick testing). Every structure file was independently verified — checked for multiple atoms per residue (a real fold) rather than the single-atom-per-residue placeholder format.

3. Part 1 — The linker problem (PETase crowding)

If the dockerin is crowding the active site, maybe a longer or differently-composed linker can push it further away, giving PET polymer chains room to reach the enzyme's catalytic residues. The pipeline tested several linkers built from repeats of glycine and serine (GGGGS) — a very common, maximally flexible "spacer" sequence used throughout protein engineering — at five different lengths, plus a few AI-suggested alternatives, against the native 12-residue linker currently in use.

outputs/results/ranked_linkers.csv — real ESM2 + ESMFold scoring, this run
Candidate	Sequence	Length	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

Composite score blends all three columns (30% ESM2 plausibility, 30% fold confidence, 40% distance), each rescaled to a 0–100 range so they're comparable, then combined with those weights.

Baseline structure — fig — baseline (native linker), real ESMFold prediction. Red = catalytic triad, gold = linker, blue = dockerin.

Top candidate structure — fig — GS_len25 (top composite score), real ESMFold prediction

Linker composite score comparison — fig — composite score by candidate

4. What does "distance" actually mean here?

The active site is a pocket-like cleft that a PET polymer chain has to physically fit into for the enzyme to cut it. The dockerin domain, sitting nearby, can partly block that entrance — like furniture pushed up against a doorway. "Distance" is measured as the closest 3D gap between any atom of the three catalytic residues and any atom of the dockerin domain, in the single 3D structure ESMFold predicts. A bigger number means the model predicts the dockerin sits further from the doorway; a smaller number means it's closer, and more likely to obstruct larger substrates trying to get in.

This connects to the real wet-lab numbers that motivated the whole notebook: activity dropped much more on real PET film (44%) than on BHET (7%). We might infer that small molecules can still squeeze past a partial obstruction, while long polymer chains need much more clearance — so distance is being used as a stand-in (a "proxy") for that clearance, not as a direct measurement of enzyme speed.

The surprising real result

Every real GS-linker candidate actually came out closer to the active site than the native linker (9.3–16.0 Å vs. 17.8 Å) — the opposite of what we had expected a longer flexible spacer to do. GS_len25 still wins the overall composite score because its ESM2 and pLDDT numbers are strong, but by the notebook's own headline metric — the one built specifically to address Problem 1 — it's actually one of the weaker options.

There are two very different explanations for this, and the data alone can't tell us which is true:

It could be a real biological finding — the charged residues in the native linker (two aspartates) might genuinely behave differently than a pure glycine-serine repeat in this specific protein context, in a way that happens to extend it further than expected.
It could be an ESMFold artifact. Long, low-complexity, repetitive sequences like GGGGSGGGGSGGGGSGGGGSGGGGS are exactly the kind of region ESMFold is known to struggle with. In reality, a flexible linker doesn't have one fixed shape — it's constantly wiggling through many possible conformations (this is called being "intrinsically disordered"). ESMFold, like most structure predictors, is built to output a single best-guess structure, and for genuinely flexible regions it tends to default to an artificially compact, folded-back shape rather than capturing the true floppy, extended ensemble. That could easily make a longer linker look shorter/closer than it would really behave in solution.

5. Part 2 — The pocket problem (MHETase bottleneck)

Here the goal is different: instead of moving a domain further away, the aim is to widen the substrate-binding pocket of the MHETase enzyme itself, so the intermediate product (MHET) can enter and be converted to the final product (TPA) more easily. The team already had one hand-designed mutant (TfCaWA, combining two mutations: I69W and V376A) as a starting guess, and the pipeline explores whether AI-guided alternatives might do better.

outputs/results/ranked_mutants.csv — real ESM2 + ESMFold scoring, this run
Construct	ESM2 score	pLDDT	Pocket free volume	Composite
WT (native)	-0.220	79	3551 Å³	60.0
TfCaWA (control)	-0.750	78	3551 Å³	23.5
mut_I69F	-2.592	79	3551 Å³	23.2
mut_I69F_V376A	-2.656	79	3551 Å³	21.1

WT structure — fig — wild-type TfCa-DoG, real ESMFold prediction

Mutant structure — fig — mut_I69F, real ESMFold prediction

Mutant composite score comparison — fig — composite score by construct

Why this ranking can't be trusted yet

Some limitations can be found in the "Pocket free volume" column: it's identical — 3550.8 Å³ — for all four rows, including the wild type. That's not a coincidence and not a sign the pocket is genuinely unchanged; it's a limitation in how the code measures pocket size. The measurement function only counts how many backbone "anchor" atoms (called Cα atoms — one per amino acid, roughly its structural midpoint) fall within a fixed-size sphere around the catalytic residues, and treats every amino acid as if it were the same fixed-size ball regardless of whether it's actually a small residue like alanine or a bulky one like tryptophan. A single point mutation rarely moves the protein's backbone enough to change that atom count — even though it very much changes the actual physical bulk sitting in the pocket (that's the whole point of the mutation). So this metric is structurally blind to the exact effect it's supposed to detect, and it contributes zero useful information to every row's score here.

On top of that: the wild type's ESM2 score (−0.22) and TfCaWA's (−0.75) are hardcoded numbers written directly into the notebook's code, not values ESM2 actually computed this run. Only the two engineered mutants' ESM2 scores are real. Combined with the pocket-volume issue, the wild type "winning" this ranking is mostly an artifact of these two gaps — not evidence that the engineered mutants are worse.

6. What we might do next

ASoftware fixes and computational double-checks

We might refactor the pocket-volume metric to actually use side-chain information. The current function (pocket_free_volume()) only looks at Cα positions. A meaningfully better version would use each residue's real side-chain atoms and their actual van der Waals radii (which differ by element — carbon, nitrogen, oxygen, sulfur all take up different space), or better yet, use a dedicated cavity-detection tool built for exactly this purpose (e.g. fpocket, POVME, or CASTp) instead of a hand-rolled sphere count.
We might replace the two hardcoded ESM2 scores for the wild type and TfCaWA with real computed values, using the same esm2_mut_score() function already used for the other mutants, so all four rows are scored on equal footing.
We might get a second opinion on the linker structures from a different structure-prediction method — running the same sequences through AlphaFold2 (or ColabFold, a fast web-hosted version of it) would show whether the surprising distance results hold up under a model that works differently, or whether they're specific to ESMFold's known weakness on repetitive sequences.
We might try to run each fold multiple times with different random seeds or numbers of "recycles" (ESMFold's internal refinement passes) and look at the spread of results, not just one prediction. If the triad-to-dockerin distance varies a lot between runs, that's itself useful evidence the linker region isn't being predicted with real confidence — regardless of what pLDDT alone says.
We might code an entirely new molecular dynamics (MD) simulation on the top candidates instead of relying on ESMFold's single static structure. MD can show how the linker moves over time, giving a distribution of distances instead of one snapshot — directly addressing the "flexible regions don't have one true shape" problem discussed above.
We might re-run the composite ranking with different weightings once the pocket-volume metric is fixed, and check how sensitive the final ranking is to the exact weights chosen — with only 3-4 candidates being compared, small changes in weighting can flip the order.

BWet-lab experiments that might confirm or refute this

Direct distance measurement: a technique called single-molecule FRET (Förster resonance energy transfer) can measure the real, dynamic distance between two points on a protein in solution — this might give some actual experimental ground truth for the triad-to-dockerin distance, rather than relying only on a computational prediction.
Express and test the top real candidates. We might synthesize the top-ranked linker variants, express them in E. coli, and re-run the existing activity assays — a fast small-molecule test (PNPB) first, then the slower, more realistic test on actual PET film, measuring how much MHET/BHET/TPA is produced by HPLC. This could directly test whether distance (predicted or real) actually correlates with restored activity.
Circular dichroism (CD) spectroscopy can check whether a linker is behaving as an unstructured, flexible region (as intended) versus accidentally folding into some secondary structure; this might provide another perspective on the "is this really floppy" question raised above.
For the mutants, we might run real enzyme kinetics (measuring K_m and k_cat, the standard measures of how well an enzyme binds and processes its substrate) on the wild type, TfCaWA, and the AI-suggested mutants side by side. This could be a way to actually know whether the pocket changed in a functionally meaningful way — computational pocket-volume metrics, even a well-built one, are still just a proxy.

This pipeline has produced genuine AI predictions, but the two metrics each problem was specifically designed to test — active-site distance and pocket volume — are surprising and unexpected. If this research is to proceed, experts will help the student team discuss and deliberate regarding which physical experiments ought to follow these computational experiences.

Glossary

PETase / MHETase: Enzymes that break down PET plastic in two sequential steps.
Dockerin: A small "hook" domain that lets an enzyme clip onto a shared scaffold protein.
Linker: A short, usually flexible stretch of amino acids connecting two protein domains.
ESM2: An AI language model trained on protein sequences; scores how "natural" a sequence looks.
ESMFold: An AI model that predicts a protein's 3D structure from its sequence alone.
pLDDT: ESMFold's per-residue confidence score, 0–100; above 70 is generally considered reliable.
Cα (alpha carbon): A single backbone atom in each amino acid, often used as a simplified stand-in for the residue's position.
Composite score: A single 0–100 ranking number built by combining several metrics with chosen weights.
Intrinsically disordered region: A part of a protein with no single fixed 3D shape — it naturally moves between many conformations.