Overview of MODIFY, an ML algorithm to co-optimize library fitness and diversityWe developed MODIFY (Fig. 1), an ML-guided framework to design high-fitness, high-diversity enzyme libraries for the engineering of new enzyme functions. Given a set of specified residues in a parent enzyme, MODIFY affords combinatorial mutant libraries to strike a balance between maximal expected fitness and diversity, without requiring functionally characterized mutants as prior knowledge (Fig. 1a). MODIFY applies a novel ensemble ML model that leverages protein language models (PLMs) and sequence density models to make zero-shot fitness predictions and employs a Pareto optimization scheme to design libraries with both high expected fitness and high diversity. High levels of expected fitness ensure the effective sampling of functional enzyme variants, while the high diversity of designed enzyme libraries spanning a wide sequence space allows the exploration of new enzyme variants. Balancing fitness and diversity is achieved by solving the optimization problem: \(\max {{{\rm{fitness}}}}+\lambda \cdot {{{\rm{diversity}}}}\), with parameter λ balancing between prioritizing high-fitness variants (exploitation) and generating a more diverse sequence set (exploration). In this way, MODIFY traces out an optimal tradeoff curve known as the Pareto frontier, on which each point represents an optimal library where neither desiderata can be further improved without compromising the other (Fig. 1b). To refine the library, enzyme variants sampled from the library are further filtered based on protein foldability and stability (Fig. 1c). Applying MODIFY to design a high-quality starting library for both new-to-nature borylation and silylation, we identified a generalist biocatalyst with substantially altered loop dynamics from top-performing MODIFY variants (Fig. 1d).Fig. 1: MODIFY: an ML-guided framework for the design of enzyme engineering starting libraries with both high fitness and high diversity.a MODIFY leverages pre-trained protein language models and multiple sequence alignment (MSA)-based sequence density models to build an ensemble ML model for zero-shot fitness predictions, effectively eliminating evolutionarily unfavorable variants. b MODIFY co-optimizes the library’s diversity and predicted fitness, pinpointing the Pareto optimal balance between the two. MODIFY offers diversity control at a residue resolution, enabling researchers to either explore a diverse range of amino acids or focus on a subset of compatible amino acids based on biophysical and biochemical insights. c MODIFY further performs a quality control step to filter out problematic variants in the library based on protein foldability (ESMFold pLDDT) and stability (FoldX ΔΔG). d MODIFY-enabled discovery of effective generalist biocatalysts for enantioselective new-to-nature borylation and silylation.Accurate zero-shot fitness predictionWe first assessed the zero-shot fitness prediction ability of MODIFY using the ProteinGym benchmark dataset23, which comprises 87 DMS assays that provide experimental measurements of protein fitness, spanning different functions such as enzyme catalytic activity, binding affinity, stability, and growth rate. This benchmark thus represents a holistic evaluation of MODIFY’s zero-shot fitness prediction across various protein families and functions. We compared MODIFY with its constituent models: two state-of-the-art PLMs (ESM-1v24 and ESM-225), two MSA-based sequence density models (EVmutation26 and EVE27), and a hybrid PLM that incorporated MSA data (MSA Transformer28). These individual models were previously established as effective unsupervised predictors in protein fitness and disease variant effect prediction24,27. We found that no single baseline consistently outperformed the others. In contrast, MODIFY’s ensemble predictor stood out by delivering accurate and robust predictions (Fig. 2a) and achieving the best Spearman correlation for the largest number of times (34/87; Fig. 2b). Across ProteinGym, MODIFY consistently outperformed at least one of the baselines in all 87 DMS datasets and often ranked at or near the top (Fig. 2a). As ProteinGym covers a wide array of protein families, this result demonstrates the general utility of MODIFY in zero-shot fitness prediction across a wide range of proteins.Fig. 2: MODIFY achieves accurate and robust zero-shot protein fitness prediction.The ensemble ML model of MODIFY was compared with five state-of-the-art unsupervised protein fitness predictors (ESM-1v, ESM-2, EVmutation, EVE, and MSA Transformer) for zero-shot protein fitness predictions. a Comparison on the ProteinGym benchmark, which contains 87 Deep Mutational Scanning (DMS) assays across diverse protein families, using Spearman correlation as the evaluation metric. b The counts of each method achieving the best performance (including ties) on the 87 DMS datasets. c The average performances of all methods on proteins with low, medium, and high MSA depths.Additionally, stratifying results based on the MSA depth of parent sequences in ProteinGym indicated that MODIFY outperformed all baselines for proteins across low, medium, and high MSA depths (Fig. 2c; Supplementary Information A.1). In contrast, no single baseline consistently outperformed other baselines across the three categories. These results underscore MODIFY’s capacity to provide reliable fitness predictions for diverse protein families, including those lacking ample homologous sequences, highlighting its general applicability. We further compared MODIFY with the baseline methods on the latest release (v1.0) of the ProteinGym benchmark dataset29 with 217 DMS assays (Supplementary Information A.1). The results mirrored these findings (Supplementary Figs. 1 and 2), featuring the superior zero-shot protein fitness capability of MODIFY across a diverse array of proteins. It should also be noted that MODIFY achieved the highest zero-shot protein fitness prediction for DMS assays measuring catalytic or related biochemical activities (Supplementary Fig. 2b), further highlighting the suitability of MODIFY for enzyme engineering.Since the majority of DMS datasets in ProteinGym focused on single mutants, we further examined MODIFY’s fitness prediction ability for high-order mutants using the experimentally characterized fitness landscapes of three proteins, including GB130, ParD331,32, and CreiLOV33, covering combinatorial mutation spaces of 4, 3, and 15 residues, respectively. MODIFY achieved notable performance improvements over other baselines, suggesting its generalizability in predicting the fitness of high-order mutants (Supplementary Fig. 3). Taken together, these results demonstrate the superior accuracy and robustness of MODIFY in predicting variant fitness across a diverse range of protein families, which lays the groundwork for the library design algorithm detailed below.In silico evaluation of starting library design on GB1With these benchmark results in hand, we next applied MODIFY to optimize a starting library on a four-site combinatorial sequence space for the GB1 protein (Fig. 3a). The fitness landscape of these sites (V39, D40, G41, and V54) was previously mapped out experimentally30, where the fitness was defined by both stability (fraction of folded proteins) and function (binding affinity to IgG-Fc). This experimentally derived dataset allowed for a retrospective assessment of the quality of our MODIFY library.Fig. 3: MODIFY designs high-quality combinatorial starting libraries for GB1.a The 3D structure (PDB: 1PGA) of GB1. The four residues mutated to create combinatorial libraries are colored in blue. b The Pareto frontier of MODIFY library designs on GB1, with each point representing a library corresponding to a diversity strength λ. Blue points are MODIFY-informed designs by varying the residue-specific diversity weight α40 for residue 40 while fixing other weights. c, d The mean experimental fitness and diversity (average entropy) of the designed libraries, each with 500 GB1 variants. Random sampling, NNK, and Exploitation are included as the baseline methods. The bar plots represented the mean ± SD over 5 independent repetitions. e The mean experimental fitness of sequences sampled from the library distribution of NNK, MODIFY, and MODIFY-informed, respectively. The curves and error bands represent mean ± SD over 5 independent repetitions. f, g Experimentally measured fitness and MODIFY’s zero-shot fitness predictions of single-mutation GB1 variants. arb. unit, arbitrary unit. h, i The amino acid (AA) distribution of the MODIFY and MODIFY-informed libraries.A unique strength of MODIFY is the optimization of the composition diversity of amino acids at the residue-level resolution, controlled by a diversity hyperparameter αi for residue i (Methods), which generalizes previous methods that only optimize diversity at the sequence level34,35. Here, we first applied MODIFY’s default setting (denoted as MODIFY) to design the library, assigning equal diversity weights αi to all four sites (Fig. 1b and Methods). MODIFY afforded a library striking an optimal balance between library diversity and mean predicted fitness of the library (Fig. 3b). By contrast, the commonly used NNK library produced a library with high diversity but low mean predicted fitness. Upon assessing a 500-sequence library designed by MODIFY and NNK using ground-truth fitness data, the MODIFY library exhibited higher mean experimental fitness (Fig. 3c) and preserved library diversity as indicated by average entropy (Fig. 3d). In contrast, the NNK library—although marginally more diverse—was predominantly populated with nonfunctional variants (Fig. 3d), as indicated by its minimal mean experimental fitness, which was similar to that of a control library that samples sequences uniformly at random (Fig. 3c). Importantly, MODIFY’s improvements were consistently observed across varying library sizes (Fig. 3e). Additionally, Exploitation, a variation of MODIFY that only prioritized variants by zero-shot fitness prediction with no consideration of diversity, resulted in a less diverse library (Fig. 3c,d). We further compared MODIFY with DeCOIL35, a recent ML-assisted library design method, and HotSpot Wizard36, and observed that the MODIFY library had both higher mean experimental fitness and higher diversity (Supplementary Fig. 4; Supplementary Information A.5). Extending MODIFY to design a fifteen-site combinatorial library for the fluorescent protein CreiLOV33 mirrored these findings (Supplementary Fig. 5; Supplementary Information A.6), underpinning MODIFY’s effectiveness in striking an advantageous diversity-fitness balance across different protein families.Furthermore, we explored MODIFY’s informed setting (denoted as MODIFY-informed; Fig. 1b) and showed how prior knowledge of a protein’s fitness can be incorporated through MODIFY’s residue-level diversity control. In this experiment, we assumed that the experimentally determined fitness data of GB1 single mutants is available to inform the design of high-order mutants. Given the linear scaling of single-site substitutions with the number of mutation sites (20 × 4 = 80 mutants for GB1), obtaining such data is experimentally feasible and cost-effective. We observed a disparity when comparing MODIFY’s zero-shot predictions to empirical fitness at position D40—MODIFY predicted all 19 possible substitutions at this site to be disadvantageous, while experiment data indicated the opposite for most mutations (Fig. 3f, g). This discrepancy showcases a potential misalignment between broad evolutionary patterns captured by zero-shot predictive models such as PLMs and the specific fitness determinants for a given protein, which resulted in the overrepresentation of the wild-type amino acid (AA) over other possibly beneficial mutations (Fig. 3h, residue 40). To counteract this effect, MODIFY-informed strategically increased the diversity weight α40 (Supplementary Information A.5), promoting the diversity of AAs for residue 40. While beneficial single mutations may not always translate into high-order mutants with improved fitness due to negative epistasis37, promoting the diversity on site D40 may increase the chances of discovering diverse and functional four-site mutants. This approach was validated by our evaluation: the informed library (Fig. 3b) not only induced a higher diversity at both residue and sequence levels (Fig. 3d, i) but also achieved a higher mean experimental fitness compared to MODIFY (Fig. 3c, e).Together, this study highlights MODIFY’s strength in creating combinatorial libraries that effectively balance fitness with diversity. In contrast to many current library design methods34,35, MODIFY further introduces residue-level diversity control, allowing the integration of prior knowledge into the library design process. In addition to DMS fitness data, other forms of prior knowledge, such as active-site residue effects revealed by biocatalysis data, can also be incorporated to tailor MODIFY library design.MODIFY library improves downstream MLDEThe sequence composition of screening libraries plays a crucial role in MLDE, as the paired sequence-fitness data is used to train supervised ML models to guide further directed evolution experiments. To probe the impact of MODIFY library on MLDE, we simulated an in silico MLDE experiment using the GB1 landscape (Fig. 4a). We first designed a 500-variant library using five methods: MODIFY, Exploitation, NNK, FoldX38—a biophysical stability prediction model, and FuncLib39—an automated method for designing combinatorial mutations at enzyme active sites. For FoldX, the top 500 mutants predicted to be most stable (lowest ΔΔG) were selected. For FuncLib, all 209 designed mutants were included (Supplementary Information A.5). We then mapped the fitness landscape of GB1 onto a 2D t-distributed stochastic neighbor embedding (t-SNE) plot for an intuitive view of library composition (Supplementary Information A.5). We observed that the NNK library sampled sequences evenly scattered across the entire landscape, but most of them are low-fitness variants (Fig. 4b) with some sequences including stop codons (Fig. 4g). The FoldX library and the FuncLib library were enriched in a single high-fitness region with limited sequence diversity (Fig. 4c, d). In contrast, the MODIFY library contained variants enriched for multiple fitness peaks (Fig. 4f), suggesting a Pareto optimal library with higher mean fitness than the NNK library (Fig. 4b) and higher diversity than the FoldX, FuncLib, and Exploitation libraries (Fig. 4c–e).Fig. 4: MODIFY library improves the performances of machine learning-guided directed evolution (MLDE) on GB1.a An in silico MLDE experiment was simulated on the GB1 landscape, where an ML model was trained to predict the sequence-fitness relationships using the variant sequences in the MODIFY library and their associated experimentally characterized fitness as the training data. The trained ML model was then applied to prioritize novel fitness-enhanced variants. b–f t-SNE visualization of the library sequences in the GB1 fitness landscape. Variants from various libraries (NNK, FoldX, FuncLib, Exploitation, and MODIFY) were colored in red. arb. unit, arbitrary unit. g Stratified bar plots of library sequences based on their fitness ranges: (WT, Max]: better than the wildtype, (0, WT]: lower than wildtype but higher than 0, {0}: zero fitness, with stop codon: variants with stop codons. h–j The performance of ML models trained on fitness-labeled sequences from each library. The mean fitness (h), the max fitness (i), and the recall of the top 100 variants (j) as a function of the top K prediction were shown. The curves and error bands represent mean ± SEM over 25 independent repetitions.Next, we paired sequences from the five libraries with their ground-truth fitness30 and used this data to train an ML model to predict the sequence-fitness relationship. In this study, all five methods used the same encoding strategy (one-hot) and ML model architecture (random forest regressor), and the only difference was sequences in the training set defined by each library. Using these models, we predicted fitness for a set of withheld variants (Supplementary Information A.5) and ranked them accordingly. We then compared the five methods with respect to the true fitness values (mean and maximum) and the recall of the top 100 variants within their top K predictions. This provided a measure of an ML model’s hit rate in MLDE given a test budget of K sequences. The ML model trained on the MODIFY library outperformed all others, exhibiting the highest mean/max fitness and recall for high-fitness variants (Fig. 4h–j). Interestingly, although the FoldX library and the FuncLib library contained a higher fraction of better-than-wildtype variants (Fig. 4g), the ML model trained on the libraries performed consistently the worst with regard to mean/max fitness and recall (Fig. 4h–j). This result underscored the importance of maximizing diversity in library design for exploring the sequence space. Overall, this in silico MLDE experiment suggested that the high-fitness, high-diversity starting libraries designed by MODIFY readily translate to improved accuracy of ML models in MLDE, thus accelerating the protein engineering process.Experimental validation of MODIFY led to novel biocatalysts from wild-type cytochrome c with excellent activity and enantioselectivityWe next experimentally validated MODIFY in the design of starting functional enzyme libraries to enable valuable biocatalytic transformations that were not known in natural enzymology20. In particular, we sought to design functional enzyme libraries that could simultaneously promote two stereoselective new-to-nature biotransformations, including the carbon–boron (C–B) and the carbon–silicon (C–Si) bond formation reactions (Fig. 5). Although organoborane and organosilane compounds are of significant value to theranostics40 and synthetic chemistry41, enzymes that catalyze the formation of C-B and C-Si bonds are not known in nature. Previously, through laboratory-directed evolution via iterative site-saturation mutagenesis and screening, two variants derived from wild-type Rhodothermus marinus cytochrome c (Rma cyt c), a small thermophilic heme protein whose native function is electron transfer (Fig. 5a)42, were separately evolved to catalyze C-B43 and C-Si44 formation. Each triple mutant arose from three rounds of directed evolution, with Rma cyt c V75R M100D M103T (denoted as the RDT variant, indicating its mutant type of amino acids; similar abbreviations will be used hereafter)43 and Rma cyt c V75T M100D M103E (TDE variant)44 being an effective borylation and silylation biocatalyst, respectively.Fig. 5: MODIFY library of Rma cytochrome c led to new and distinct C-B and C-Si bond-forming biocatalysts.a Crystal structure (PDB: 6CUK) of Rma cytochrome c. The six targeted residues 75, 99, 100, 101, 102, and 103 to create combinatorial libraries are colored in blue. b Biocatalytic carbon-boron (C-B) bond formation. c Biocatalytic carbon-silicon (C-Si) bond formation. d The Pareto frontier of MODIFY library designs on Rma cytochrome c. e, f The AA distribution of MODIFY and MODIFY-informed libraries. g Top 10 variants for C–B bond formation reactions from the MODIFY-informed library. †Previously engineered borylation biocatalyst via three rounds of directed evolution. h Top 10 variants for C–Si bond formation reactions from the MODIFY-informed library. ‡Previously engineered silylation biocatalyst via three rounds of directed evolution. i, j Total activity (yield) and enantioselectivity (fraction of the major enantiomer) of the variants from MODIFY-informed and NNK libraries for the biocatalytic C–B and C–Si bond formation reactions, respectively. k, l The enantioselectivity and activity correlations between biocatalytic C–B bond formation and C–Si bond formation for the variants from the MODIFY-informed library. The error band indicates the 95% confidence interval of the regression line. TTN = total turnover number. e.r. = enantiomeric ratio. For formatting purpose, MODIFY-informed library was denoted as MODIFY in i–l.The MODIFY algorithm allowed us to sample high-fitness regions in the Rma cyt c sequence space not previously available from laboratory-directed evolution. In particular, we aim to engineer cytochrome c variants that can catalyze both C-B (Fig. 5b) and C-Si (Fig. 5c) bond formation with excellent efficiency and stereocontrol. The development of such generalist stereoselective enzymes to catalyze multiple biotransformations has remained a challenging task, as most evolved enzyme variants are reaction-specific. Guided by the crystal structure of Rma cyt c (Fig. 5a)42,45, we constructed a MODIFY library focusing on sequence optimizations for the α-helix residue 75 proximal to the heme cofactor and five flexible loop residues 99, 100, 101, 102, and 103 (Fig. 5d–e). In wild-type Rma cyt c, the M100 residue is bound to the Fe center to confer a hexacoordinate Fe42. To generate a catalytically active Fe center, we leveraged MODIFY’s residue-level diversity control and eliminated M100 from our designed library (Supplementary Information A.7). We further enhanced the residue-level diversity at site 75 in light of the proximity of this residue to the heme cofactor (Fig. 5f). This MODIFY-designed Rma cyt c library contained the top 1000 variants (Supplementary Data 1). The gene fragment library was synthesized using the oligo-pool technology46 and cloned into a pET–22b(+) vector with an N-terminal pelB sequence (see Supplementary Information A.7 for details). As a negative control, a randomized combinatorial library based on the NNK degenerate codon was also experimentally evaluated. In our experiments, 160 clones of the MODIFY library were randomly selected and screened in both the C-B and the C-Si bond-forming reactions in the form whole-cell biocatalysts or cell-free lysates (Fig. 5b–c; Supplementary Information A.7). Chiral HPLC analysis was performed to determine the yield and enantiomeric ratio (e.r.) of the organoborane and organosilane products.Biotransformation results in Fig. 5i, j showed that MODIFY offered markedly improved results relative to the NNK control in both the C-B and the C-Si bond-forming processes. Specifically, in biocatalytic C-B bond formation, MODIFY library afforded a 2.2-fold higher averaged yield and a fourfold higher averaged enantiomeric ratio (Fig. 5i). In C-Si bond formation, MODIFY library provided a 1.9-fold higher averaged yield and a 1.3-fold higher averaged enantiomeric ratio (Fig. 5j). Importantly, an array of C-B and C-Si bond-forming biocatalysts with excellent activity and enantioselectivity emerged from this MODIFY library (Fig. 5g, h; Supplementary Tables 3 and 4). Interestingly, these best-performing borylation and silylation enzyme variants are 6 mutations away from the previously experimentally evolved RDT43 and TDE variants44, showcasing MODIFY’s ability to identify novel functional variants not easily available by other means.Notably, among the best-performing MODIFY borylation biocatalysts, the MGAANQ variant displayed a TTN of 2880 and an e.r. of 95:5 (Fig. 5g, entry 2) outperforming the experimentally evolved RDT variant (Fig. 5g, entry 1). In addition to the MGAANQ variant, four other MODIFY variants, including MLYPPT (Fig. 5g, entry 3), MQVANQ (entry 4), MESANQ (entry 5) and MELQNQ (entry 6), outperformed the RDT variant with respect to their total turnover numbers. Similarly, among the best-performing silylation biocatalysts, the SFLTNQ variant displayed a TTN of 3,320 and an e.r. of 98:2 (Fig. 5h, entry 2), outperforming the experimentally evolved TDE variant (Fig. 5h, entry 1). In addition to the SFLTNQ variant, another variant VQFPPQ also provided better TTN (Fig. 5h, entry 3) relative to the TDE variant.Intriguingly, among these Rma cyt c MODIFY variants, many incorporated a proline (P) residue into the flexible loop, indicating a substantial change in loop conformation and dynamics among these MODIFY X→P variants47. Moreover, functional double-proline mutants with a proline at residues 99, 100, 101, and 102, including MLYPPT, MPQPNQ, VQFPPQ, KPWPNY, and SPIPAM, were uncovered from this MODIFY library. The altered loop conformation of these proline mutants with excellent catalytic activity and enantioselectivity represents a departure from the canonical structures of the experimentally evolved Rma cyt c RDT and TDE variants, further highlighting the power of MODIFY in revealing novel enzyme variants. Importantly, from this single-round screening, we identified a generalist Rma cyt c variant MLYPPT, which is highly active and enantioselective for both the borylation (2880 TTN, 94:6 e.r. (Fig. 5g, entry 3)) and the silylation (2,740 TTN, 97:3 e.r. (Fig. 5h, entry 6)) reactions, providing a rare example of promiscuous biocatalyst variant reminiscent of general small-molecule catalysts with broad utility.The availability of a library of functional Rma cyt c variants for both C–B and C–Si bond formation also allowed us to interrogate the enzyme activity and enantioselectivity correlations between borylation and silylation reactions, a task not previously achievable due to the limited availability of functional variants. In general, Rma cyt c variants that are highly enantioselective for C–B bond formation were also found to be highly enantioselective for C–Si bond formation, as revealed by a Pearson correlation coefficient of 0.72 between the percentage of the major borane enantiomer and that of the major silane enantiomer (Fig. 5k). Similarly, variants exhibiting a higher activity in C–B bond formation are also usually more active in C–Si bond formation, despite a slightly smaller Pearson correlation coefficient (0.47) (Fig. 5l). Together, the ability to profile enzyme variant activity and selectivity in two biocatalytic reactions offer a rare opportunity to shed light on mutational effects on multiple fitness landscapes.MD simulations offer insights into altered loop dynamics of MODIFY Rma cytochrome c variantsTo gain further insights into the flexible loop dynamics of the newly uncovered protein mutants, we carried out molecular dynamics simulations of the Fe carbene intermediates of these cytochrome c variants without and with NHC-BH3 and PhMe2SiH substrates in the active site for selected best-performing MODIFY variants and previously evolved RDT and TDE variants. Previous studies revealed the key role of this flexible loop as the dynamic lid flanking the active site of this compact heme protein and regulating catalysis43. Our MD simulations reveal significant changes in front loop (99–103) conformations and dynamics of MODIFY variants (Fig. 6). To quantify the flexibility of each variant, B-factor values48 (Bi, Å2) were calculated from root-mean-square fluctuation (\({\rho }_{i}^{rmsf}\)) of Cα atoms in MD simulations. For the Fe carbene intermediates of TDE and RDT variants (Fig. 6a), the front loops are moderately rigid, as indicated by the blue-colored region. Although the front loop dynamics of MPQPNQ remain similar to TDE and RDT variants, both the MLYPPT and SPIPAM mutants show enhanced flexibility of the front loop, as indicated by yellow to red colors. Interestingly, for the MLYPPT variant, a substantial flexibility increase is also observed for α-helix residues 91-98. The flexibility enhancement of the front loop could allow the enzyme to better accommodate the NHC-BH3 or PhMe2SiH substrate, leading to improved reaction efficiency. To model the substrate near-attack-conformations49 that promote the borylation and silylation, the distance between the carbene carbon and hydrogen atom of each substrate is restrained to be within 2.4–2.8 Å (Figs. 6c, d). Unlike RDT and TDE variants, upon the binding of either the NHC-BH3 or the PhMe2SiH substrates, the front loop (99–103) of MLYPPT is characterized by further enhanced flexibility, allowing for the loop to change conformation for better substrate bindings50. The rigidity of TDE, RDT, and MPQPNQ is due to the conserved water-mediated hydrogen bond network in the front loop50, which does not exist in MLYPPT and SPIPAM variants (Fig. 6b). The lack of hydrogen bond networks, thus, leads to the more flexible front loop of these variants, as well as the more flexible α-helix region in MLYPPT. The substantially improved loop flexibility of the MLYPPT variant to accommodate different types of substrates may contribute to its reaction generality for both the C–B and the C–Si bond-forming processes.Fig. 6: Classical molecular dynamics (MD) simulations.a Representative snapshots revealed contrasting loop dynamics of Fe carbene intermediates of MODIFY-designed and experimentally evolved Rma cyt c variants. b Overlay of 10 most-populated snapshots obtained from 1000 ns MD simulations and the presence and absence of hydrogen bond network in Rma cyt c TDE and MLYPPT variants, respectively. c Representative snapshots of enzyme-substrate complexes of Fe carbene with the NHC-BH3 substrate. d Representative snapshots of enzyme-substrate complexes of Fe carbene with the PhMe2SiH substrate. The backbone of residues 91–103 is colored based on B-factor value of the Cα atom of each residue.
