Arcadia tried to turn a green fluorescent protein orange; one of ten tested variants fluoresced, and it was green
Arcadia tried to turn a green fluorescent protein orange; one of ten tested variants fluoresced, and it was green
On September 4, Arcadia Science published a study on generating fluorescent proteins with specified colors and brightness levels. The algorithm guided amino acid substitutions toward a target spectrum; an independent computational evaluation and experiments in bacteria produced different results.
A fluorescent protein emits light when its chain folds into a functional structure and its chromophore matures. The chromophore is the part of the molecule that absorbs and emits light. The authors started with the green fluorescent protein EGFP and changed amino acids in the pocket surrounding the chromophore. This region determines which wavelengths the protein absorbs and emits, and therefore its color; the rest of the protein scaffold remained unchanged.
At each position selected for modification, the algorithm compared two measures: how closely the proposed amino acid substitution matched variants in the fluorescent protein family, and how close the predicted excitation and emission peaks were to the target. For five orange candidates, the team added a brightness check: a classifier assigned variants to bright or dim categories based on their signal in the green spectral range. Brightness was intended as an indirect indication that the protein would fold, mature, and function in a cell.
In the computational evaluation, a separate model that had no role in selecting substitutions assessed the results. Across 36 pairs of starting proteins and target spectra, the guided search came closer to the target in 27 cases, with an average advantage of 9 nanometers over selecting from previously generated sequences.
For the laboratory experiment, Arcadia generated 4 500 variants and selected ten. The team tested them in bacteria and measured their spectra. One variant produced a fluorescent signal. The algorithm had designed it to match the peaks of the orange protein mOrange, 548 and 562 nanometers, but the variant had excitation peaks near 390 and 500 nanometers and an emission peak at 520 nanometers, in the green spectral range.
The authors suggest that the orange chromophore matures poorly within the EGFP scaffold. The classifier had been trained on GFP variants with mutations throughout the chain, using signals measured in the green spectral range. In the new design campaign, mutations were concentrated around the chromophore and shifted the color away from green. Under these conditions, the classifier could have interpreted a spectral shift as a loss of protein function.
The authors propose starting the next campaign by checking whether the target color is achievable from the chosen starting protein. The next step would require measurements of variants with changes in the chromophore pocket to guide the model’s selection of new variants.