Results
This page reports the headline closed-loop result: does the optimizer actually improve the design library, and does it beat a random-search baseline under an identical evaluation budget and the immunogenicity feasibility constraint?
All numbers come from the in-silico oracle stack (labelled proxies, not wet-lab assays); the contribution is the method and the closed-loop machinery, which a real experimental data stream drops into unchanged.
Optimizer benchmark
Mean final feasible-front hypervolume across random seeds (higher is better), each optimizer run through the same DBTL budget (6 cycles, library size 16, 16-sequence seed library). Reproduce with:
python benchmarks/benchmark_optimizers.py --n-seeds 5 --n-cycles 6
| optimizer | final feasible-front HV (mean ± std, 5 seeds) |
|---|---|
| qNEHVI (BoTorch) | 0.439 ± 0.016 |
| NSGA-II | 0.389 ± 0.019 |
| random (baseline) | 0.384 ± 0.014 |
Constrained Bayesian optimization (qNEHVI) wins decisively — its advantage over both NSGA-II and random search exceeds the across-seed standard deviation. This is the expected ordering when evaluations are the bottleneck: the GP surrogate spends the budget where expected hypervolume improvement is highest, subject to the immunogenicity constraint, instead of exploring blindly.
Reading it: every optimizer's hypervolume is non-decreasing across cycles (the archive only grows), so the meaningful comparison is the final front. NSGA-II and qNEHVI apply selection pressure toward the Pareto front and the immunogenicity constraint, so they dominate random search — which improves only by luck.
What the loop optimizes
Four maximization objectives under one hard constraint:
| signal | role | direction |
|---|---|---|
| contrast | ultrasound scattering | maximize |
| collapse pressure | mechanical set-point (multiplexing) | target closeness |
| expressibility | mammalian expression | maximize |
| solubility | aggregation resistance | maximize |
| immunogenicity | MHC-II epitope load | constraint (≤ ceiling) |
Caveats
- Oracles are in-silico proxies; absolute values are not experimental claims.
- The hypervolume is a deterministic Monte-Carlo estimate; small differences within a seed's noise band should not be over-interpreted.
- At this scale (cheap proxy oracle, modest budgets) a genetic algorithm competes with Bayesian optimization; BO's advantage grows with expensive evaluations and tight query budgets — the regime real wet-lab DBTL actually operates in.