2026-08-29
ALDE plus cell-free expression screens 120 miRFP670nano3 variants; the best is 4.35× brighter in bacteria, then 3.6–4.1× dimmer than parent in HEK293T cells.
GFP family proteins emit blue-green light that tissue absorbs and scatters. Far-red proteins in the miRFPnano line, engineered from bacteriophytochromes, sit in the tissue window and are still much dimmer. Classic directed evolution screens huge random libraries; when the assay is slow, that campaign never starts.
Active-learning directed evolution (ALDE) closes the loop. Each round trains a surrogate on sequence–function data and nominates the next variants. This study wires that loop to nanolitre cell-free expression and aims it at miRFP670nano3.
The surrogate is a Gaussian process on ESM3 embeddings, mean-pooled, column-standardized, then reduced to three principal components. Across nine DMS datasets, model class did not change final-best fitness (Friedman p=0.32). The GP stayed because it supplies posterior uncertainty for Bayesian acquisition.
The wet lab uses PURExpress from linear DNA, 5 μM biliverdin, and an N-terminal HiBiT tag for expression. The training label is weighted activity, 75% brightness and 25% expression, both normalized to parent. The design space is single-site saturation. Round zero has 24 variants: 12 near k-means centroids in embedding space, 12 HotSpot Wizard-tolerant sites. Five rounds total 120 variants. Rounds 1–3 measure single mutants; those beating parent are combined as triples in rounds 4–5.
Retrospective simulations on nine ProteinGym DMS sets hold the budget at 384 assays and sweep batch size from one batch of 384 down to 47 batches of 8. Cold-start priors are ESM-2 masked-marginal scores (pMMS) and ESM-IF1 inverse-folding likelihoods, plus scrambled controls.
Of 120 variants, 74 differ significantly from parent in weighted activity; 12 more than double brightness. Top variant K6G-L50F-A143T is 4.35× brighter and 25% lower in expression. Mutations near the biliverdin site, including C86A which breaks the covalent thioether, are consistently harmful. pMMS barely tracks measured brightness (Spearman ρ=0.098, p=0.286).
Mammalian cells reverse the ranking. In HEK293T transient transfection the parent far-red geoMFI is 3.6–4.1× higher than three engineered variants (ANOVA p=2.0×10^{-6}). Lentiviral stable lines and dose titrations agree. Purified bacterial protein still favors K6G-L50F-A143T and I130C-V138I-Q145K (ANOVA p=0.0011). Extinction coefficient and quantum yield do not move; thermostability drops.
On the simulations, smaller batches raise cumulative-mean fitness faster (mean within-dataset Spearman ρ=−0.927, p=0.0039). The effect on final-best fitness is weaker (ρ=−0.524, p=0.0195). The authors treat 48–96 samples per batch as a practical compromise. Zero-shot priors help only when they correlate with true fitness; that correlation tracks regret reduction (positive in 8/9 datasets, ρ=0.652). For miRFP670nano3, pMMS is uninformative.
A model-in-the-loop campaign can surface mutations that structure and conservation would not pick, in 120 measurements. It also produces a blunt warning: a variant selected under saturating biliverdin in cell-free lysate can lose completely in mammalian cells. PURExpress supplies 5 μM cofactor; mammalian media get 25 μM that crosses membranes poorly and is inhibited in serum. Spectra are unchanged, so the mutations likely raise the fraction of folded, cofactor-loaded protein. Remove the excess cofactor and that advantage disappears.
For AI-for-protein work, batch size and a confidence-weighted prior matter more than swapping the surrogate. Model class barely separated on nine DMS sets.
The trained GP does not generalize to held-out positions; mean pooling plus three PCs likely discarded functional signal. The search stayed at single sites, so combinatorial pocket mutants were never tried. Mammalian checks came after all five rounds; an in-cell assay each round would have caught the mismatch earlier. Thermostability fell, spectra did not, so the engineering win is mostly bacterial and cell-free. The manuscript is a bioRxiv preprint and has not been peer reviewed.