Prospective Case Study
July 2026 · bioRxiv preprint
AI Designed a More Developable FGF21
Two FGF21 constructs, side by side — one from the literature, one from the Orbion platform. The AI design gave 2.4× less purified protein — yet concentrated to twice the level, stayed more monodisperse, and had been scored higher before the bench.
Çağlar Bozkurt, Evangelia Nathanail, Aniruddh Goteti · Orbion GmbH, Berlin
Wet-Lab Execution by Data Powered Therapeutics GmbH
The Real Bottleneck
For Hard Proteins, the Biology Isn't the Bottleneck
Getting well-behaved material is. Programs stall at construct design, expression, and purification — before the science even starts. Where to truncate, which tag, which host, how to purify so the protein survives concentration. These calls are still made from precedent and intuition, and they fail for most difficult targets.
Four coupled decisions. Improving one in isolation doesn't give you a molecule you can carry forward — which is why the platform scores and builds them together.
The Target · FGF21
A Metabolic Target That Fights Back
Fibroblast growth factor 21 — a secreted hormone behind a wave of late-stage MASH, type-2 diabetes, and obesity therapeutics. A compact β-trefoil core with a long, disordered, proteolysis-prone C-terminal tail, and a melting temperature near 46.8 °C. Exactly the kind of stability-challenged protein where the construct decides the outcome.
The whole story lives in that highlighted sliver. It's a hydrophilic, proline/serine-rich disordered segment — the piece the field routinely trims for expression and crystallography. The Orbion design kept it. Residue numbering: UniProt Q9NSA1, 209 aa. The sequence is 100% native — the novelty is the choice, not a mutation.
Two Designs, One Workflow
Literature Intuition Versus the Platform
Identical FGF21 core. Both were produced under the same expression and purification protocols — generated by Orbion's Bench module — so the comparison isolates the design itself: truncation boundary, signal peptide, and tag architecture.
Scientist Construct
The Published Folded-Core Boundary (PDB 6M6E)
- Boundary42–169 · Folded Core
- SignalBM40 (SPARC)
- TagN-Terminal Twin-Strep · 3C
Orbion AI Construct
Core + 23 Native C-Terminal Residues — a Boundary BLAST Finds Nowhere
- Boundary43–192 · Core + Tail
- SignalIg-κ
- TagC-Terminal 3C · Twin-Strep
The Orbion construct was generated by the Design module, and the shared expression and purification protocols by the Bench module — both constructs ran on those same protocols, executed scientist-in-the-loop. Every score above was locked before the constructs were synthesised.
Round One · Raw Yield
First, the Result That Went Against Us
We report it first, because it's the honest order. On raw output, the scientist's construct won — cleanly. A faster bio-layer-interferometry on-rate at every time point (day-3 response 3.5 vs 2.0 nm), and 2.4× more purified protein from the same culture. Both eluted as single, symmetric peaks; the human design simply made more.
Round One → Scientist≈ 24 mg/L vs 10 mg/L. The platform had predicted 6–16 mg/L for its own construct — the observed ~10 mg/L landed inside that window.
The Reversal · Developability
Then Came the Concentration Step
Developability is what happens next: does the protein survive being concentrated and handled? Here the two molecules diverged. The Orbion construct concentrated cleanly to 1.4 mg/mL. The scientist construct stalled at 0.7 mg/mL, shedding material to aggregation.
Twice the usable concentration — without aggregating.
More monodisperse, at 8× the DLS signal-to-noise (3074 vs 388).
Both are single, symmetric peaks — no aggregate in the void. The scientist peak is ~3× taller (more protein); the Orbion peak elutes earlier (larger apparent size, from its retained C-terminus — consistent with its larger hydrodynamic radius by DLS, 2.34 vs 1.78 nm). Traces are the real instrument data.
So the 2.4× yield lead didn't survive.The scientist construct entered concentration with far more protein and left with less usable material — ~70 µg against the Orbion construct's ~84 µg. Both were single clean bands by SDS-PAGE, so this isn't “dirty but soluble.” The prediction-guided design simply held together.
Called in Advance
What the Platform Predicted, Before the Bench
The ranking was right — and so were most of the property calls it rested on. We show the one it missed, too.
| Axis | Prediction (Mammalian) | Wet-Lab Outcome | Verdict |
|---|---|---|---|
| Composite Ranking | Orbion > Scientist · 68.7 vs 59.0 | Orbion More Developable | Matched |
| Yield (Orbion) | 6–16 mg/L | ~10 mg/L | Matched |
| Solubility | Both Soluble · ~99% | Both Soluble | Matched |
| Disorder | Orbion More Extended · 27.6% vs 13.5% | Larger Apparent Size, Longer C-Term | Matched |
| Aggregation | ~Equal · 50.2 vs 49.3 | Scientist More Aggregation-Prone | the Honest Miss |
| Thermostability | Δ ~0.3 °C | Unassignable (Low Trp) | Untestable |
Solubility matched but didn't discriminate — both constructs were soluble. The bounded claim: the platform's ranking and several property predictions were right in advance. The aggregation sub-score was the honest miss — it didn't separate the two, and at face value slightly favoured the wrong one.
A Biophysical Hypothesis
The Value Was in the Piece Everyone Trims
We are cautious about mechanism at n = 1, but the most likely explanation is biophysical. The retained segment — residues 170–192 — is hydrophilic and proline/serine/alanine-rich: the signature of a disordered, low-complexity region. Segments like this can act as entropic bristles — enlarging the hydrodynamic radius and shielding aggregation-prone surfaces, the same trick disordered solubility tags exploit.
“Removes the long low-confidence, PTM-rich C-terminal tail beyond the high-confidence folded region… retaining the folded core expected to drive binding-related function.”— verbatim platform design rationale
Consistent with what we measured: a larger hydrodynamic radius, earlier SEC elution, and clean concentration to higher levels — retaining a boundary the field trims improved, rather than harmed, developability.
What This Is — and Isn't
One Clean Data Point. We Won't Stretch It
This is a single, deliberately scoped pair — a demonstration of capability, not a population benchmark. Stated plainly, because it bounds what may be concluded.
- SAMPLE SIZESingle pair, single run. A case study, not a statistically powered benchmark — no significance testing is possible or claimed.
- CONFOUNDINGPartly confounded. The constructs differ in boundary, signal, and tag position — but signal and tag are removed before the developability measurements, so the mature proteins differ principally in the retained C-terminus.
- BUFFERBuffer-specific. Solution behaviour was measured in one formulation; a different buffer could shift the comparison.
- FUNCTIONActivity not assayed. "Developable" here means solution behaviour, not function — both constructs trim FGF21’s receptor-engaging regions.
- MEASUREMENTMeasurement caveats. The DLS comparison was noisier for the scientist construct, and apparent melting temperatures could not be confidently assigned for either — FGF21’s low tryptophan content makes the nanoDSF signal shallow.
- MASS SPECIdentity not confirmed by mass spec. Purity was assessed by SDS-PAGE; intact-mass confirmation would be standard before downstream use.
The honest statement is that the design choice paid off and is most consistent with the retained C-terminus — not that a single causal mechanism was proven. Quantifying the effect across many targets is the natural next step.
The Same Platform Runs on Your Targets
The highest-leverage decision is rarely “make the most protein” — it's “make the molecule that survives the bench.” Bring your hardest target and see how the predictions land against your own data.