AI-redesigned starting points and outcomes enhance protein evolution

General methods

Antibiotics from Gold Biotechnology were prepared in 1,000× stock solutions in water unless otherwise indicated: chloramphenicol (25 mg ml^–1 in 70% ethanol), carbenicillin (50 mg ml⁻¹), spectinomycin (50 mg ml⁻¹), tetracycline (10 mg ml⁻¹ in 50% ethanol) and kanamycin (25 mg ml^–1). PCR amplifications were carried out using either Phusion U Green Multiplex PCR Master Mix (Thermo Fisher Scientific) or Q5 Hot Start High-Fidelity 2× Master Mix (New England BioLabs). DNA oligonucleotides were synthesized by Integrated DNA Technologies. Plasmids encoding synthetic, human codon-optimized ATXN2 cDNA sequences were obtained from GenScript. Plasmid constructs were assembled via Golden Gate or Gibson cloning protocols following previously described methods^21,37. Cloning was performed in chemically competent E. coli Mach1 cells (Thermo Fisher Scientific) or NEB 5α cells (New England Biolabs). Plasmids from single colonies were amplified using the Illustra Templiphi 100 Amplification Kit (Cytiva) before sequencing via Sanger (Quintara Biosciences) or Nanopore (Quintara Biosciences or Plasmidsaurus). For bacterial experiments, plasmid DNA was purified using the QIAprep Spin Miniprep Kit (Qiagen), whereas plasmids for use in mammalian cells was isolated with the Plasmid Plus Midiprep Kit (Qiagen). All plasmids were eluted in nuclease-free water and quantified using a NanoDrop ONE UV-Vis spectrophotometer (Thermo Fisher Scientific). The list of plasmids and selection phages used in this study is available in Supplementary Table 1.

ProteinMPNN sequence design

Code and procedure for sequence design are available on GitHub (https://github.com/Nicholas-Krasnow/sequence-design-guide). In summary, AF2-predicted structures of each redesigned BoNT protease were used as input for ProteinMPNN, and residues outside specified thresholds for distance to substrate and evolutionary conservation were constrained from redesign. AF2 predictions were confirmed to agree with existing experimental structures while offering a model of regions unresolved by crystallography^54,55,56. Sequences were generated in groups according to the constraint cut-offs used. Substrate distance constraints were 14 Å and 18 Å for BoNT/E, 10 Å, 14 Å and 18 Å for BoNT/F, and 18 Å for BoNT/X as measured in PyRosetta⁵⁷. Conservation of each residue was determined as the residue frequency in a multiple sequence alignment of homologues identified from a database search of Uniref50 with the starting BoNT protease as the search query. Conservation constraints of 30% and 60% were used for all proteases; each residue that was at least as conserved as the cut-off and was the plurality residue in the alignment position was constrained from design. For BoNT/X, an additional criterion was applied to constrain residues predicted to lie within 14 Å of the belt domain in the holotoxin based on alignment to the BoNT/A holotoxin structure (Protein Data Bank ID 3BTA)⁵⁸. During sequence generation, two sampling temperatures of 0.1 and 0.3 were tested. Cysteine was excluded from design to avoid oxidation. Structures of output sequences were predicted in AF2 and evaluated by calculating the predicted local distance difference test score and root mean square deviation to the input structure measured in PyMOL.

PROSS sequence design

PROSS variants of BoNT/E were generated using the PROSS webserver (https://pross.weizmann.ac.il/step/pross-terms/). The same AF2 input structure and the same distance constraints of 14 Å and 18 Å used for BoNT/E ProteinMPNN were used for PROSS. Because PROSS directly uses a multiple sequence alignment to sample putative stabilizing mutations, the conservation constraints used for ProteinMPNN were not applied. Folding energy was calculated with the ref2015 energy function. The webserver was run until 44 non-duplicate sequences were generated for each distance constraint group.

Small-scale protein expression and purification

Starter cultures were grown overnight from single colonies and used to inoculate 5 ml of medium supplemented with 50 µg ml⁻¹ kanamycin per well in a 24-well plate (50 μl inoculum per well). Cultures were incubated at 37 °C with shaking until reaching mid-log phase, then were subjected to a 1-h cold shock on ice. Protein expression was induced with 1 mM IPTG (Gold Biotechnology) and cultures were incubated overnight at 16 °C.

Overnight cultures were harvested by centrifugation. Cell pellets were resuspended in 400 μl of B-PER reagent (Thermo Fisher Scientific) supplemented with 0.8 μl each of lysozyme and DNaseI and 16 μl of cOmplete protease inhibitor tablet solution prepared in 2 ml nuclease-free water (Roche). Lysis proceeded by incubation at room temperature for 15 min without freeze–thaw cycles. Lysates were transferred to microcentrifuge tubes and clarified by centrifugation at 20,000g for 5 min at 4 °C. To prepare affinity resin plates for purification (HisPur Cobalt Spin Plates, Thermo Fisher Scientific), storage solution was removed from the resin by centrifugation at 500g for 3 min. Wells were washed once with 400 μl ultrapure water and centrifuged again at 500g for 3 min. Plates were equilibrated by two washes with 400 μl wash buffer (20 mM HEPES and 200 mM NaCl, pH 7.3), each followed by centrifugation. Plates were chilled on ice.

Following lysis, 300 µl of each supernatant (cleared lysate) was transferred to the pre-chilled 96-well cobalt resin plate and incubated for 15 min on ice. Plates were centrifuged for 3 min at 500g at 4 °C, and flow-through was collected. The resin was washed four times with 400 μl of wash buffer containing 10 mM imidazole, each wash followed by centrifugation at 500g for 3 min at 4 °C. Bound proteins were eluted by incubating the resin with 200 μl of elution buffer (wash buffer + 250 mM imidazole, pH 7.3) for 1 min, followed by centrifugation at 500g for 3 min at 4 °C. Eluates were sealed and stored at 4 °C.

To remove imidazole and adjust buffer conditions for downstream assay, buffer exchange was performed using Zeba spin desalting plates (Thermo Fisher Scientific). Desalting plates were equilibrated to room temperature, and storage buffer was removed by centrifugation at 1,000g for 2 min. Each well was washed four times with 250 μl exchange buffer (50 mM HEPES and 5 mM NaCl pH 7.3), each followed by centrifugation. Purified protein samples (100 μl) were applied to the resin bed, followed by 20 μl of exchange buffer to ensure full recovery. The processed protein was eluted by centrifugation at 1,000g for 2 min and collected in a new 96-well plate. SDS–PAGE analysis was used to assess protein expression and purification across various stages. Purified protein samples were stored at 4 °C for subsequent quantification and activity assays. Protein concentrations were measured using Pierce BCA assay (Thermo Fisher Scientific).

Protease cleavage kinetics assay

FRET assays were conducted as previously described⁵⁹. In brief, substrate proteins were diluted to the indicated concentrations in reaction buffer (50 mM HEPES pH 7.3, 5 mM NaCl, 2 mM dithiothreitol and 10 µM Zn(OAc)₂). Proteases were diluted in reaction buffer to 20×, the indicated final concentration. Substrates (47.5 µl) were aliquoted into 96w white half-area, clear-bottom plates (Costar), and optimal gain values for 50% signal at 470 nm (donor fluorophore) and 100% signal at 526 nm were measured on a Tecan Spark plate reader. Diluted protease samples (2.5 µl) were added to substrate wells, mixed, then fluorescence emission was measured on the plate reader at 470 nm and 526 nm over a 1–2-h time course at 37 °C. After incubation, 1 µl trypsin (New England Biolabs) from a resuspension in 500 µl reaction buffer was added to each well, and emission at the same wavelengths at 37 °C was measured again until values plateaued. Conversion was calculated at each timepoint as previously reported according to the equation:

where [S]₀ is the initial substrate concentration, X is the emission at 470 nm:emission at 526 nm (Em470:Em526) at a given timepoint, neg is the Em470:Em526 from protease-free substrate control at the same timepoint, and pos is the Em470:Em526 ratio upon complete conversion (trypsinolysis). Data were analysed using Microsoft Excel and subsequently plotted and fit using PRISM (GraphPad). Cleavage rate was calculated as the slope of linear fit to conversion over time in the linear range. When a statistically significant rate conversion (Pearson correlation P ≤ 0.05) was not observed, the maximum cleavage rate was calculated as the mean of three replicates plus three standard deviations.

Medium-scale protein expression and purification

E. coli BL21 Star (DE3) cells (Thermo Fisher Scientific) transformed with plasmids encoding 6×His-tagged protease or FRET substrate constructs or MBP–GST substrate constructs were streaked onto kanamycin-containing agar plates, and single colonies were used to inoculate 2×YT cultures containing kanamycin (50 µg ml⁻¹). Cultures were grown overnight at 37 °C with shaking at 220 rpm. The following day, overnight cultures were diluted 1:100 into fresh 2×YT medium (250–1,000 ml) containing kanamycin and incubated at 37 °C until reaching mid-log phase (optical density at 600 nm (OD₆₀₀) ≈ 0.4–0.7). Cultures were then subjected to a 1-h cold shock on ice before induction with 1 mM IPTG (Gold Biotechnology). Induced cultures were incubated overnight at 16 °C with shaking.

Cells were harvested by centrifugation at 6,000g for 5 min and resuspended in cold lysis buffer (10 ml per 250 ml culture volume; 20 mM HEPES pH 7.3 and 200 mM NaCl, supplemented with cOmplete EDTA-free protease inhibitor tablet, Roche). Cell lysis was performed by sonication (amplitude 10, 3 s pulse on, 6 s pulse off, 6-min processing). The resulting lysate was clarified by centrifugation at 19,000g for 20 min at 4 °C.

For His-tagged proteins, the supernatant was loaded onto Ni-NTA affinity resin (Thermo Fisher Scientific; 1 ml resin per 250 ml culture), pre-equilibrated with wash buffer. Binding was carried out by gentle end-over-end mixing at 4 °C for 1 h. The resin was then transferred to gravity-flow columns and washed twice with 20 ml of cold wash buffer supplemented with 10 mM imidazole. Bound proteins were eluted in 4 × 1.5 ml fractions using wash buffer + 250 ml imidazole, pH 7.3. For MBP–GST substrate proteins, supernatants were instead loaded onto glutathione agarose resin (Thermo Fisher Scientific), washed twice with wash buffer, then eluted with wash buffer + 50 mM glutathione. Eluted fractions were evaluated by SDS–PAGE, and those containing the target protease were pooled and subjected to buffer exchange with Amicon 15-kDa centrifugal filters (Millipore Sigma) into the original wash buffer. Final protein concentrations were determined using the BCA assay. For long-term storage, proteins were combined with an equal volume of 20% glycerol dissolved in wash buffer, flash-frozen with liquid nitrogen, then stored at −80 °C.

Thermal stability assay

Thermal stability was determined by measuring T_m values according to a published protocol⁶⁰. Proteins were diluted in reaction buffer to 1 µg µl⁻¹. In a PCR plate, 1 µl diluted protein samples or 1 µl buffer-only blank samples were then added to 19 µl dye solution comprising reaction buffer with 800×-diluted SYPRO orange (Thermo Fisher Scientific) and mixed. Melting was observed by measuring fluorescence on a Bio-Rad quantitative PCR machine with the following protocol: equilibrate to 25 °C for 2 min, increase temperature by 0.5 °C, then equilibrate for 1 min and measure fluorescence at 560 nm, repeating the cycle until reaching 95 °C. Fluorescence versus temperature and the first derivative versus temperature were plotted, and T_m values were identified as the temperature of the maximum derivative value.

Preparation and transformation of chemically competent cells

Strain S20607 was used for all luciferase assays, phage amplification, plaque formation assays and PACE experiments. Chemically competent cells were prepared by diluting overnight cultures 1:50 into 50 ml of 2×YT medium supplemented with tetracycline. Cultures were grown at 37 °C with shaking (220 rpm) until reaching an OD₆₀₀ of 0.4–0.6. Cells were harvested by centrifugation at 4,000g for 10 min at 4 °C and gently resuspended in 5 ml of TSS buffer (LB containing 5% v/v DMSO, 10% w/v PEG 3350 and 20 mM MgCl). Aliquots were frozen on dry ice and stored at −80 °C for later use.

For transformation, 100 µl of thawed competent cells was added to a chilled mixture of plasmid DNA (1–2 µl per plasmid, up to three plasmids total) and 95 µl of KCM buffer (100 mM KCl, 30 mM CaCl₂ and 50 mM MgCl₂ in water). The mixture was gently stirred and incubated on ice for up to 10 min, followed by a 75-s heat shock at 42 °C. SOC medium (500 µl; New England Biolabs) was then added, and cells were recovered at 37 °C with shaking for 1 h. Transformed cells were plated on 2×YT agar containing the appropriate antibiotics and incubated at 37 °C for 16–18 h.

Luciferase assay for PACE circuit activation

S2060 cells were transformed with the relevant substrate accessory plasmid and pBAD protease expression plasmid. Single colonies were cultured overnight in 2×YT medium supplemented with the appropriate maintenance antibiotics. The next day, cultures were diluted 1:50 into fresh DRM medium containing the same antibiotics and grown at 37 °C with shaking (220 rpm) until reaching an OD₆₀₀ of 0.4–0.6. Arabinose (1 mM) was added to each sample to induce protease expression, 100 µl of each sample was transferred into a black-walled, flat clear-bottom 96-well plate (Costar) for measurement. Wells containing DRM only were included to obtain background measurements. Absorbance at 600 nm and luminescence were recorded using either a Tecan Infinite M1000 Pro or Tecan Spark microplate reader over a time course of 2.5 h. Luminescence readings were normalized to cell density by dividing the background-subtracted luminescence signal by the background-subtracted OD₆₀₀ absorbance.

Toxin purification and luciferase complementation assay for toxin delivery

Plasmids encoding each toxin, tagged with a C-terminal 6×His sequence, were transformed into BL21 DE3 competent cells (New England Biolabs). The transformed cells were plated on LB agar containing 50 µg ml⁻¹ kanamycin and incubated overnight at 37 °C. A single colony was selected and cultured overnight in 2×YT medium supplemented with 50 µg ml⁻¹ kanamycin at 37 °C. The following day, this culture was diluted 1:100 into fresh 2×YT medium and incubated at 37 °C with shaking (200 rpm) until the OD₆₀₀ absorbance reached approximately 0.6. At this point, cultures were subjected to a 20-min cold shock, after which protein expression was induced by adding 1 mM IPTG. The cultures were then incubated overnight with shaking at 16 °C.

Protein purification was performed at 4 °C. Cells were harvested via centrifugation at 4,000 rpm for 5 min and resuspended in lysis buffer (20 mM Tris-HCl, 500 mM NaCl, 1 mM MgCl₂, 10 mM imidazole, 10% glycerol, 1 mM PMSF, protease inhibitor cocktail, 100 µM ZnCl₂ and 100 mM arginine, pH 7.5). Lysis was achieved by sonication (3 s pulse on, 6 s pulse off, 3-min processing). The lysate was clarified by centrifugation at 18,000g, and the supernatant was applied to a Ni-NTA affinity column. The column was washed twice using a wash buffer containing 20 mM Tris-HCl, 500 mM NaCl, 50 mM imidazole, 1 mM MgCl₂, 100 µM ZnCl₂, 10% glycerol and 100 mM arginine (pH 7.4). Proteins were eluted with elution buffer (20 mM Tris-HCl, 500 mM NaCl, 1 mM MgCl₂, 250 mM imidazole, 10% glycerol, 100 µM ZnCl₂ and 100 mM arginine, pH 7.4). Eluted proteins were first concentrated using a 15-ml Amicon Ultra centrifugal filter (100,000 kDa cut-off) at 3,500g and buffer exchanged into 20 mM Tris-HCl, 150 mM NaCl, 100 µM ZnCl₂ and 10% glycerol, pH 7.5. A second concentration step was performed using a 0.5-ml Amicon Ultra filter (100 kDa cut-off) at 14,000 g. Protein concentrations were measured with the BCA assay, and samples were treated with 10 U mg⁻¹ thrombin at 4 °C overnight.

For HiBiT:LgBiT luciferase live-cell complementation assays, 15 K N2A cells were seeded in 96-well plates and transfected with 33 ng per well pLX304_CMV::mCherry-LgBiT-FLAG-NES (Addgene #199715) 24 h after seeding. After another 24 h, the media were replaced with 100 µl of complete medium containing 1 µM HiBiT-tagged toxin or HiBiT peptide. Following another 24-h incubation, cells were washed with PBS and bioluminescence was measured using the Nano-Glo Live Cell Assay System (Promega), according to the manufacturer’s instructions, using a Tecan Spark plate reader. In parallel, 50 µl of media was collected separately from the plate after treatment with Nano-Glo substrate and used to assess background signal. This background reading was subtracted from the luminescence values obtained from the treated cells.

PTEN cleavage assay in mammalian cells

Of HEK293T cells (American Type Culture Collection (ATCC), mycoplasma free) at 160,000 cells per millilitre, 2.7 ml was seeded in six-well culture plates. After 24 h, wells were co-transfected with 1.1 µg full-length FLAG-tagged PTEN in pCDNA3.1 vector (Addgene #22231) and 1.1 µg N-terminal OSBP pleckstrin-homology domain tagged (to promote colocalization with PTEN), C-terminal 6×His-tagged proteases in pCDNA3.1 using Lipofectamine 2000 transfection reagent (Thermo Fisher Scientific) following the manufacturer’s instructions. For no protease negative controls, 1.1 µg pUC19 plasmid was transfected instead of protease expression plasmid. Cell lysates were harvested 24 h later in lysis buffer (RIPA (Thermo Fisher Scientific) with 1 mM PMSF and cOmplete protease inhibitor (Roche)). The lysates were analysed by western blot for PTEN (using an anti-FLAG tag antibody) or protease (using an anti-6×His tag antibody) as well as GAPDH as a loading control. Anti-FLAG and anti-6×His were detected with horseradish peroxidase-conjugated secondary antibodies and chemiluminescence measurement, whereas anti-GAPDH was detected with IRDye 680RD-conjugated secondary antibodies and fluorescence measurement. Band intensity was quantified with Image Lab (Bio-Rad).

Substrate profiling assay

A single-stranded DNA oligo library of the SNAP25 sequence from D166 to D186 with every possible single-residue substitution was synthesized from Twist Biosciences. PCR was used to amplify the oligo library into double-stranded DNA, which was purified by PCR cleanup (Qiagen). A T7 substrate plasmid backbone was amplified by PCR, DpnI digested to remove template plasmids and purified by PCR cleanup. The double-stranded substrate library was cloned into the plasmid backbone by isothermal assembly (NEBuilder-HiFi, New England Biolabs), and the resulting plasmid library (t7s) was purified by ethanol precipitation using GlycoBlue coprecipitant (Thermo Fisher). Purified plasmid was electroporated into NEB-10β electrocompetent E. coli (New England Biolabs), recovered in 20 ml 2×YT medium supplemented with 25 mM glucose for 10 min, and transferred into 200 ml terrific broth supplemented with 50 µg ml⁻¹ carbenicillin for overnight recovery at 37 °C in a 220-rpm shaking incubator. The cloned plasmid library was harvested from NEB-10β outgrowth culture via midiprep (Qiagen), following the manufacturer’s protocols. The plasmid library was subsequently electroporated into S2060 containing pBAD-BoNT/E plasmid, recovered in 20 ml 2×YT medium supplemented with 25 mM glucose for 10 min, and transferred into 200 ml terrific broth supplemented with 50 µg ml⁻¹ carbenicillin and 50 µg ml⁻¹ spectinomycin for overnight recovery at 37 °C in a 220-rpm shaking incubator.

To perform selection, 1 ml of S2060 [pBAD-BoNT/E] [t7s] library culture was pelleted and resuspended in 1 ml of DRM to remove residual growth medium. The washed cell suspension was inoculated into 49 ml pre-warmed DRM supplemented with carbenicillin, spectinomycin and 20 µM arabinose in a disposable 125-ml bacterial culture flask. The flask was incubated at 37 °C with 220 rpm shaking to induce protease expression. Once the cells reached an OD₆₀₀ of 0.8, 1 ml of culture was inoculated into 49 ml DRM with carbenicillin, spectinomycin and arabinose supplemented with 25 µg ml⁻¹ chloramphenicol. Of induced culture, 1 ml was also inoculated into 49 ml DRM–carbenicillin–spectinomycin–arabinose without chloramphenicol. Selection cultures were incubated overnight at 37 °C with 220 rpm shaking. Following overnight selection, plasmids from both the plus and the minus chloramphenicol cultures were harvested by midiprep (Qiagen), following the manufacturer’s protocols. The substrate sequences from the plasmid pools of both cultures were amplified by PCR using primers rCFH0600 and rCFH0601 and barcoded using the Illumina Truseq adaptor system. The substrate sequences of the plasmid pools were sequenced on an Illumina MiSeq (Illumina). Sequencing data were analysed using custom scripts to count the occurrence of each substrate sequence within both the plus and the minus chloramphenicol conditions. For each substrate s, the log-enrichment score E_s was calculated using the following formula:

where C_s,+ is the read count of substrate s in the plus chloramphenicol condition and C_s,– is the read count of substrate s in the minus chloramphenicol condition. Raw read counts were normalized to the total number of reads across all substrates in each condition to account for differences in sequencing yield.

Plaque assay

S2060 cells were transformed with pJC175e (Addgene #79219) to generate S2208 (ref. ⁸). Overnight cultures of single S2208 colonies grown in 2×YT medium supplemented with carbenicillin and tetracycline were diluted 50-fold into fresh DRM with the same antibiotics and grown at 37 °C with shaking at 220 rpm to OD₆₀₀ ~ 0.6–0.8. Phage were serially diluted 100-fold (4 dilutions total) in DRM. Of cells, 100 µl was added to 10 µl of each phage dilution. Of top agar pre-warmed to 55 °C (LB medium + 0.33% LB agar), 500 µl was added and mixed by pipetting up and down once. These mixtures were then immediately pipetted onto 24-well plate wells containing solidified bottom agar (LB agar with 0.04% Bluo-Gal (Research Products International), no antibiotics). After solidification of the top agar, plates were incubated at 37 °C for 16–18 h. Phage titres were determined by counting plaques in dilution wells with approximately 10–100 plaques.

Overnight propagation assays

Host cells were prepared by transforming S2060 cells with accessory plasmid, and overnight cultures of single colonies were grown overnight in 2×YT medium supplemented with antibiotics. Cultures were diluted 50-fold into fresh DRM with antibiotics and grown at 37 °C with shaking at 220 rpm to OD₆₀₀ ~ 0.6–0.8. Host cells were infected with phage at a titre of 1 × 10⁵ PFU ml⁻¹ then grown at 37 °C with shaking at 220 rpm overnight. Phage were collected by centrifugation at 8,000g for 2 min. Titres of propagated phage were determined by plaque assay, and overnight propagation values were calculated as output phage titre/input phage titre.

General method for eVOLVER-enabled phage-assisted continuous evolution

PACE host strains, phage, turbidostats, lagoons and media were prepared as previously described^2,5,6,11. Hardware and software for autonomous eVOLVER and min-eVOLVER devices were used as reported for these systems^45,47. For comparison evolutions, to prevent differences in mutation profiles occurring due to differences in codon usage, phage-encoded proteases were designed so that the same codon was used between the WT and designed starting points at all synonymous positions.

Hosts were prepared by transforming chemically competent S2060s with AP(s) and MP6, plated on 2×YT agar with antibiotics and 100 mM glucose to suppress MP6 expression, then grown at 37 °C for 16–18 h. Colonies were picked into cultures of DRM with antibiotics, and a dilution series of six 10-fold dilutions of each culture was prepared and grown with shaking at 37 °C for 16–18 h. Cultures with OD₆₀₀ ~ 0.4–0.8 were used to inoculate 30 ml eVOLVER or min-eVOLVER reservoirs to serve as turbidostats. Turbidostats were set to maintain mid-log OD₆₀₀ of 0.6–0.8 at constant volume. Phage were subjected to evolutionary drift before PACE by preparing a drift host (S2208 transformed with MP6), infecting mid-log culture of drift host with phage, then collecting phage by centrifugation at 8,000g for 2 min after growing 12–16 h at 37 °C.

Lagoons were maintained at a volume of 5 ml with the indicated influx flow rate of host culture from the turbidostat. Mutagenesis was induced by pumping 250 mM arabinose into the lagoon to a constant final concentration of 10 mM. Lagoons were infected with phage by inoculating 500 µl phage at the desired titre into each lagoon. Samples (500 µl) of the phage population were taken at indicated timepoints from the lagoon waste needle. Timepoint samples were centrifuged at 8,000g for 2 min and phage-containing supernatants were stored at 4 °C. Lagoon titres were determined by plaque assay of phage samples. For sequencing of phage clones, eight single plaques were picked, and the transgene region was amplified with PCR and sequenced by Nanopore (Quintara Biosciences or Plasmidsaurus). Consensus clones were defined as the genotypes comprised each observed combination of recurring mutations.

General method for PANCE

Hosts and phage were prepared as described above for PACE. Cultures of host cells containing MP6 at OD₆₀₀ ~ 0.4–0.8 were supplemented with 20 mM arabinose to induce mutagenesis then aliquoted into 1 ml cultures in a 96 deep-well plate. Cultures were inoculated with selection phage at the indicated dilution, then grown 12–16 h at 37 °C and harvested the next day by centrifugation (4,000g for 10 min) and storage of supernatant at 4 °C. Phage were then used to infect the subsequent PANCE passage with hosts prepared the same way. The process was repeated for the number of passages indicated, with titres determined by plaque assay after each passage. Drift passages were performed as described for PACE.

LoopSeq

Phage pools from PACE were prepared by propagating phage with S2208 host cells overnight followed by PCR amplification of phage pools using phage transgene PCR primers used for plaque sequencing. A low cycle count of 6–14 cycles was used to reduce the frequency of recombination during PCR. Amplicon pools were purified by Ampure XP bead purification and size purity was confirmed by TapeStation. Sequencing of prepared samples and in silico assembly of synthetic long reads was performed at Element Biosciences. Long reads were analysed by aligning reads to the starting protease gene as the reference, determining the coding mutations compared with the reference, and counting the number of reads of each observed mutation combination. Abundant genotypes were defined as those with more than 100 assembled long reads mapped, which were plotted in histograms and included in mutation set sharing, entropy and Hamming distance analyses. Mutation combinations arising from a particular starting point were classified as shared if the same mutation combination occurred in at least one read of at least one replicate of the alternative starting point, and mutation combinations were classified as starting point specific if not.

Analysis of protein sequence embeddings in ESM-C

Sequence embeddings of evolved consensus genotypes in the WT BoNT/E background were generated according to the ESM repository embedding tutorial (https://github.com/evolutionaryscale/esm/tree/main/cookbook/tutorials). Principal component analysis was performed on each embedding layer, then the embedding layer with principal components 1 and 2 that clustered evolved protease sequences by substrate closest to the k-means clusters, according to Rand index, was selected for analysis (layer 5). The BoNT/E background mutants were plotted in the principal component space and the mutants were replotted with the same coordinates to show activity in the D3, D4 and PROSS1 backgrounds.

Gel shift proteolysis assays

Substrate proteins were diluted to the indicated concentration in reaction buffer. Proteases were diluted in reaction buffer to 20× the indicated final concentration. Of diluted protease samples or buffer-only negative controls, 2.5 µl was added to diluted substrates in PCR strip tubes, mixed, then incubated for the indicated reaction time at 37 °C. Reactions were quenched in LDS loading buffer supplemented with 2 mM dithiothreitol, then analysed by SDS–PAGE and Coomassie InstantBlue stain (Abcam).

Intact protein mass spectrometry

Proteolysis reactions were diluted to 100 ng μl⁻¹ in 25 mM HEPES pH 7.5, 10 mM MgCl₂ and 150 mM NaCl. Of diluted protein, 1 μl was injected onto a Waters BioAccord LC-TOF (composed of an ACQUITY I-Class UPLC and RDa detector with an ESI source) with an ACQUITY UPLC Protein BEH C4 column (300 Å, 1.7 μm, 2.1 × 50 mm) at 80 °C for the duration of analysis. Samples were desalted for 1 min before being eluted onto the mass spectrometer with a 5–85% gradient of acetonitrile in 2.5 min at a flow rate of 0.4 ml min⁻¹. Ionization was performed at a cone voltage of 55 V and desolvation temperature of 550 °C. The instrument scanned at a rate of 0.2 scans per second over the range 50–2,000 m/z. Protein m/z spectra were deconvoluted into intact mass using the MaxEnt1 function within UNIFI software (Waters).

Molecular dynamics simulations of protease–substrate complexes

WT BoNT/E and D3 protease–substrate complex structures from which simulations were initiated from were prepared by AF3 prediction of the protease with Zn²⁺ and the following SNAP25 cleavage site-flanking subsequence:

MDENLEQVSGIIGNLRHMALDMGNEIDTQNRQIDRIMEKADSNKTRIDEANQRATKMLG.

Because AF3 accurately placed the substrate in the correct register for the highly active SNAP25-cleaving proteases WT BoNT/E and D3, but not for the proteases evolved not to cleave SNAP25, evolved protease–substrate complexes were instead prepared by mutating the parent WT or D3 complex prediction to the evolved sequence.

All-atom molecular dynamics simulations were performed using OpenMM (v8.2)⁶¹. Protein systems were parameterized with the Amber ff14SB force field⁶², and the catalytic Zn²⁺ ion was treated using the standard Amber non-bonded 12-6 model⁶³. Existing hydrogen atoms were removed from input structures and re-added at pH 7.0 to ensure consistent protonation states. Each system was solvated in a rectangular box of TIP3P water⁶⁴ with a minimum padding of 10 Å between the solute and box edges. No counterions were added to preserve the native electrostatic environment of the peptide–Zn²⁺ complexes. Long-range electrostatic interactions were calculated using the particle mesh Ewald method⁶⁵, and all bonds involving hydrogen atoms were constrained using the SHAKE algorithm, permitting a 2-fs integration timestep.

Simulations were conducted in the canonical (NVT) ensemble at 298 K using a Langevin integrator with a friction coefficient of 1.0 ps⁻¹. Before production dynamics, each system underwent 500 steps of energy minimization to relieve steric clashes. Production trajectories were propagated for 3 μs per system (1.5 × 10⁹ integration steps), with coordinates saved every 100 ps for subsequent analysis. To reduce storage requirements, only protein and Zn²⁺ ion coordinates were retained in trajectory files, excluding bulk solvent. All simulations were performed on NVIDIA L40s and H100 GPUs, hosted on the MIT Engaging cluster administered by MIT Research Computing, using mixed-precision arithmetic. Checkpoint files were written every 50 ps to enable seamless restart upon computational resource preemption, ensuring continuous trajectory accumulation across multiple job submissions.

Ataxin-2 protease expression and cleavage assay in mammalian cells

Of HEK293T cells (ATCC) at 160,000 cells per millilitre, 2.7 ml was seeded in six-well culture plates. After 24 h, wells were co-transfected with 1.1 µg full-length FLAG-tagged ATXN2 cDNA in a pCDNA3.1 vector and C-terminal 6×His-tagged proteases in pCDNA3.1 using Lipofectamine 2000 transfection reagent (Thermo Fisher Scientific) following the manufacturer’s instructions. For protease expression blots, 1.1 µg protease plasmid was transfected, whereas for cleavage detection blots 0.11 µg was transfected. For no protease negative controls, an equivalent mass of pUC19 plasmid was transfected instead of protease expression plasmid. Cell lysates were harvested 24 h later in lysis buffer (RIPA (Thermo Fisher Scientific) with 1 mM PMSF and cOmplete protease inhibitor (Roche)). The lysates were analysed by western blot for ataxin-2 (using an anti-FLAG tag antibody) or protease (using an anti-6×His tag antibody) as well as GAPDH as a loading control. Anti-FLAG and anti-6×His were detected with horseradish peroxidase-conjugated secondary antibodies and chemiluminescence measurement, whereas anti-GAPDH was detected with IRDye 680RD-conjugated secondary antibodies and fluorescence measurement. Band intensity was quantified with Image Lab (Bio-Rad).

Mammalian cell viability assay to measure protease toxicity

Of HEK293T cells (ATCC) at 160,000 cells per millilitre, 100 µl was seeded in 96-well culture plates. After 24 h, wells were transfected with the indicated mass of indicated protease expression plasmid using Lipofectamine 2000 transfection reagent (Thermo Fisher Scientific) following the manufacturer’s instructions. Viability was measured 42 h post-transfection with the CellTiter-Glo kit (Promega) according to the manufacturer’s instructions. Cells were incubated at room temperature for 30 min before 100 µl CellTiter-Glo reagent was added to each well. Contents were mixed, signal was allowed to stabilize for 10 min, then luminescence dependent on cellular ATP was measured using a Tecan Spark plate reader.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.