Model-guided engineering of DNA sequences with predictable site-specific recombination rates

Site-specific recombination (SSR) enables precise DNA rearrangements but most applications emphasize final states (insertions, deletions, inversions) rather than controllable reaction kinetics. Large serine recombinases (LSRs) like Bxb1 are attractive due to irreversibility, simplicity, and efficiency, yet predictable control of SSR rates remains lacking. Prior work largely focused on DNA sequence specificity and directionality, or on protein engineering of recombinases, which is constrained by limited structural information and lack of high-throughput rate assays. The authors hypothesize that tuning the DNA attachment sequence (attP) can programmatically modulate Bxb1-mediated inversion rates while retaining specificity. They aim to develop a sensitive rate assay, screen attP sequence libraries, and train a machine-learning model to predict reaction rates from sequence, then validate predictions in vitro and in E. coli for applications such as kinetic control in gene circuits.

The paper situates LSRs among recombinases, highlighting their irreversibility, short recognition sites, and broad efficiency across organisms. Prior engineering approaches predominantly altered protein residues, but limited high-resolution recombinase–DNA structures (only one LSR–DNA complex with insufficient interface resolution) hinder rational design. Studies on Bxb1 attP/attB examined base conservation, specificity, and directionality; high-throughput profiling identified specificity determinants, revealing potential off-targets. However, these focused on specificity rather than quantitative kinetics. The authors note that DNA sequence determinants of SSR rates likely involve not only direct protein–DNA contacts but also shape/charge complementarity and water-mediated interactions, and that creating multiple DNA substrates per recombinase is more tractable for circuit design than engineering multiple recombinases.

• Library design: Analyzed Bxb1 attP/attB half-sites (attP-L/R, attB-L/R) and conserved positions likely critical for specificity. Performed saturation mutagenesis on attP-L to maximize tunability while maintaining recognition. Library 1 randomized 10 consecutive positions (9–18). Initial in vivo selection in E. coli indicated conservation at 9G, 10T, 18A; positions 11–17 tolerated substitutions. Library 2 targeted predominantly asymmetric positions (3G, 5G, 8G, 11T, 12G, 14C, 15C, 16A, 17G) with 19C fixed as a control.
• qPCR-based rate assay: Developed a selective qPCR assay that exponentially amplifies only inverted products using primers flanking the attP cleavage site and exploiting orientation change upon inversion. Identified 5 min at 30 °C as optimal to approximate initial rate conditions. Constructed a standard curve by mixing flipped/unflipped templates at fixed total copies (10^9 per 20 μl), achieving a linear response (r^2=0.998) with sensitivity down to ~10^3 flipped copies (~0.0001%).
• In vitro selection and NGS: Incubated DNA libraries with Bxb1 across a range of enzyme concentrations (1–300 nM) for 5 min (and a permissive 1 h condition at 300 nM), selectively amplified flipped products by qPCR, and subjected amplicons to Illumina sequencing. Calculated enrichment of WT bases per position and frequencies of individual attP-L variants.
• Converting frequencies to rates: Under low Bxb1 and DNA concentrations, treated each variant’s inversion as an independent first-order process. Converted NGS frequencies to sequence-specific flipping rate constants k_flip over the 5-min reaction using a derived relationship (Supplementary Method 2). Correlated NGS frequency with measured flipped percentage for selected sequences (r^2=0.83).
• Model construction: One-hot encoded sequences and fit a multiplicative position-specific model assuming independent contributions per position: k_flip = ∏_i Σ_j (W_ij X_ij). Linearized via logarithm and estimated weight scores W_ij by least squares on the top 3000 sequences selected at 1 nM Bxb1. Set W for WT bases to 1, normalizing k_flip,WT ≈1. Split 200 sequences into training (n=140) and validation (n=60) sets for evaluation.
• Model evaluation: Assessed prediction performance (r^2≈0.718 training; r^2≈0.714 validation). Performed sensitivity analysis by varying training set size (10–3000 sequences), observing stabilization of W values and convergence of prediction errors for ≥100–200 sequences. Compared W across selections at 1, 3, and 10 nM Bxb1; patterns were similar with attenuated magnitude at higher enzyme concentration.
• Experimental validation: Selected 12 attP-L variants spanning predicted rate range (including top mutants) and quantified in vitro flipping at 10 nM Bxb1, 1 nM DNA, 5 min via qPCR; predictions matched measurements. Tested selected variants in an intermolecular recombination assay, observing consistent relative rates. In E. coli, constructed a reporter system with Bxb1 donor plasmid and substrate plasmid where inversion drives mCherry expression; compared WT, a high-rate variant (e.g., S2), and a low-rate variant (C14A), confirming predicted dynamics.
• Gene circuit application and modeling: Designed a co-expression plasmid with two attP sites (attP1, attP2) governing GFP and mCherry expression through differential inversion events and incorporating product inhibition via enzyme sequestration on residual sites. Built a mechanistic model (Supplementary Method 3) to simulate relative and total expression versus k_flip(attP1/attP2). Constructed nine attP1/attP2 combinations (low= C14A, medium=WT, high=S2) and experimentally measured fluorescence after 12 h in M9 at 37 °C, validating model predictions.

• Developed a selective qPCR assay for SSR inversion with high sensitivity and quantitative accuracy: standard curve r^2=0.998; lower detection bound ~10^3 flipped copies (≈0.0001%).
• Library competence: Under most-permissive conditions (300 nM Bxb1, 1 h), 33% of library 1 and 51% of library 2 variants showed SSR competence, indicating substantial tolerance to attP-L substitutions.
• Enrichment and conservation: WT bases at conserved/homologous positions (e.g., 9G, 10T in library 1; 19C in library 2) were strongly enriched, while several positions (5, 8, 12, 17) favored alternative bases under stringent selection.
• Frequency-to-rate correlation: For selected variants, NGS frequency correlated with measured flipping percentage (r^2=0.83), enabling conversion to k_flip.
• Predictive model: A multiplicative position-specific weight model accurately predicted rates across variants. Cross-validation showed good agreement (training r^2≈0.718, validation r^2≈0.714). Weight scores identified substitutions increasing efficiency (e.g., 5A/5T, 8T, 12A/12C, 14T, 17A; W>1) and disruptive substitutions (e.g., 3A/3C, 11C, 12T, 14A/14G, 15A, 16G/16T; W<0.1).
• Tunability: Reaction rates programmable over four orders of magnitude. Combining beneficial substitutions yielded variants with up to ~10-fold higher initial rates than WT (e.g., S1, S2).
• Generality: Relative efficiencies preserved in both intra- and intermolecular recombination assays.
• In vivo validation: In E. coli, high-rate variants drove rapid mCherry expression; low-rate variants showed minimal fluorescence, matching in vitro predictions.
• Circuit-level control: A two-site attP1/attP2 system enabled predictable ratiometric and total expression control of GFP/mCherry. Experimental data across nine combinations aligned with model simulations, demonstrating effects of enzyme sequestration (product inhibition) on total output.

The study addresses the central challenge of predictably tuning SSR reaction kinetics by engineering the DNA attachment sequence rather than the enzyme. By creating a sensitive qPCR-based initial-rate assay and coupling it with NGS-based selections, the authors quantitatively linked attP-L sequence to inversion rate. A simple, interpretable position-specific model captured the intrinsic contributions of nucleotides and accurately predicted k_flip across diverse variants and conditions. These findings provide practical, composable parts (attP variants) with graded kinetics for synthetic biology and elucidate that independent base contributions, beyond strictly conserved contacts, can modulate recombinase activity via long-range or water-mediated interactions. The validated in vitro and in vivo demonstrations, including a co-expression circuit with mechanistically modeled enzyme sequestration, show that kinetic control can be harnessed to regulate timing, ratios, and magnitudes of gene expression, expanding recombinase applications beyond memory storage to dynamic circuit design.

This work introduces a high-throughput, model-guided framework to rationally engineer attP sequences that tune Bxb1-mediated recombination rates predictably. The qPCR assay enables accurate initial-rate measurements; NGS selections quantify sequence performance; and a position-specific weight model predicts k_flip from sequence, enabling rate modulation over four orders of magnitude. Engineered variants can outperform WT (up to ~10× increase) and function as kinetic tuning elements in living cells and synthetic circuits. Future directions include extending the approach to other LSRs and nucleic acid–modifying enzymes, exploring broader sequence spaces (e.g., additional positions or both half-sites) with subsaturation designs, integrating structural data as it becomes available to interpret weight scores mechanistically, and leveraging kinetic control for temporal ordering and multi-process coordination in complex genetic programs.

• Model assumptions: The predictive model assumes independent contributions of nucleotides across positions; cooperative or context-dependent interactions may be underrepresented, though DNA’s structural rigidity and observed tolerance support the approximation.
• Selection dependence: Weight score magnitudes vary with enzyme concentration during selection, potentially requiring condition-matched calibration for different contexts.
• Measurement limits: The qPCR assay detects down to ~0.0001% flipped products; extremely low-rate variants and sequences below top-enrichment thresholds have limited validation accuracy due to NGS depth and assay sensitivity.
• Scope of mutagenesis: The primary library focused on the attP-L half-site with nine randomized positions (library 2). Effects from attP-R or combined site changes were not comprehensively explored to avoid specificity loss.
• Structural interpretation: Lack of high-resolution Bxb1–DNA complex structures limits mechanistic attribution of specific weight scores to protein–DNA contacts.