Taleb coined the term “antifragility” to describe systems that benefit from stressors and volatility. While nature provides several examples of systems with antifragile behavior, manufacturing has so far only aimed to avoid or absorb stressors and volatility. This article surveys existing examples of antifragile system behavior in biology, biotechnology, software engineering, risk management, and manufacturing. From these examples, components of antifragile systems and principles to implement these components are derived and organized in a framework. The framework intends to serve as guidance for practitioners as well as starting point for future research on the design of antifragile systems in manufacturing.

Protein engineering through directed evolution and (semi)rational approaches is routinely applied to optimize protein properties for a broad range of applications in industry and academia. The multitude of possible variants, combined with limited screening throughput, hampers efficient protein engineering. Data-driven strategies have emerged as a powerful tool to model the protein fitness landscape that can be explored in silico, significantly accelerating protein engineering campaigns. However, such methods require a certain amount of data, which often cannot be provided, to generate a reliable model of the fitness landscape. Here, we introduce MERGE, a method that combines direct coupling analysis (DCA) and machine learning (ML). MERGE enables data-driven protein engineering when only limited data are available for training, typically ranging from 50 to 500 labeled sequences. Our method demonstrates remarkable performance in predicting a protein’s fitness value and rank based on its sequence across diverse proteins and properties. Notably, MERGE outperforms state-of-the-art methods when only small data sets are available for modeling, requiring fewer computational resources, and proving particularly promising for protein engineers who have access to limited amounts of data.

Cytochrome P450s are versatile catalysts for biosynthesis applications. In the P450 catalytic cycle, two electrons are required to reduce the heme iron and activate the subsequent reductions through proposed electron transfer pathways (eTPs), which often represent the rate-limiting step in reactions. Herein, the P450 BM3 from Bacillus megaterium was engineered for improved catalytic performance by redesigning proposed eTPs. By introducing aromatic amino acids on eTPs of P450 BM3, the “best” variant P2H02 (A399Y/Q403F) showed 13.9-fold improved catalytic efficiency (kcat/KM = 913.5 L mol−1 s−1) compared with P450 BM3 WT (kcat/KM = 65.8 L mol−1 s−1). Molecular dynamics simulations and electron hopping pathways analysis revealed that aromatic amino acid substitutions bridging the cofactor flavin mononucleotide and heme iron could increase electron transfer rates and improve catalytic performance. Moreover, the introduction of tyrosines showed positive effects on catalytic efficiency by potentially protecting P450 from oxidative damage. In essence, engineering of eTPs by aromatic amino acid substitutions represents a powerful approach to design catalytically efficient P450s (such as CYP116B3) and could be expanded to other oxidoreductases relying on long-range electron transfer pathways.

The combinatorial complexity of the protein sequence space presents a significant challenge for recombination experiments targeting beneficial positions. To overcome these difficulties, a machine learning (ML) approach was employed, which was trained on a limited literature dataset and combined with iterative generation and experimental data implementation. The PyPEF method was utilized to identify existing variants and predict recombinant variants targeting the substrate channel of P450 CYP116B3. Through molecular dynamics simulations, eight multiple-substituted improved variants were successfully validated. Specifically, the RMSF of variant A86T/T91H/M108S/A109M/T111P was decreased from 3.06 Å (wild type) to 1.07 Å. Additionally, the average RMSF of the variant A86T/T91P/M108V/A109M/T111P decreased to 1.41 Å, compared to the wild type’s 1.53 Å. Of particular significance was the prediction that the variant A86T/T91H/M108G/A109M/T111P exhibited an activity approximately 15 times higher than that of the wild type. Furthermore, during the selection of the regression model, PLS and MLP regressions were compared. The effect of data size and data relevance on the two regression approaches has been summarized. The aforementioned conclusions provide evidence for the feasibility of the strategy that combines ML with experimental approaches. This integrated strategy proves effective in exploring potential variations within the protein sequence space. Furthermore, this method facilitates a deeper understanding of the substrate channel in P450 CYP116B3.

Naturally occurring and engineered flavin-binding, blue-light-sensing, light, oxygen, voltage (LOV) photoreceptor domains have been used widely to design fluorescent reporters, optogenetic tools, and photosensitizers for the visualization and control of biological processes. In addition, natural LOV photoreceptors with engineered properties were recently employed for optimizing plant biomass production in the framework of a plant-based bioeconomy. Here, the understanding and fine-tuning of LOV photoreceptor (kinetic) properties is instrumental for application. In response to blue-light illumination, LOV domains undergo a cascade of photophysical and photochemical events that yield a transient covalent FMN-cysteine adduct, allowing for signaling. The rate-limiting step of the LOV photocycle is the darkrecovery process, which involves adduct scission and can take between seconds and days. Rational engineering of LOV domains with fine-tuned dark recovery has been challenging due to the lack of a mechanistic model, the long time scale of the process, which hampers atomistic simulations, and a gigantic protein sequence space covering known mutations (combinatorial challenge). To address these issues, we used machine learning (ML) trained on scarce literature data and iteratively generated and implemented experimental data to design LOV variants with faster and slower dark recovery. Over the three prediction−validation cycles, LOV domain variants were successfully predicted, whose adduct-state lifetimes spanned 7 orders of magnitude, yielding optimized tools for synthetic (opto)biology. In summary, our results demonstrate ML as a viable method to guide the design of proteins even with limited experimental data and when no mechanistic model of the underlying physical principles is available.

Secretion of proteins into the extracellular space has great advantages for the production of recombinant proteins. Type 1 secretion systems (T1SS) are attractive candidates to be optimized for biotechnological applications, as they have a relatively simple architecture compared to other classes of secretion systems. A paradigm of T1SS is the hemolysin A type 1 secretion system (HlyA T1SS) from Escherichia coli harboring only three membrane proteins, which makes the plasmid-based expression of the system easy. Although for decades the HlyA T1SS has been successfully applied for secretion of a long list of heterologous proteins from different origins as well as peptides, but its utility at commercial scales is still limited mainly due to low secretion titers of the system. To address this drawback, we engineered the inner membrane complex of the system, consisting of HlyB and HlyD proteins, following KnowVolution strategy. The applied KnowVolution campaign in this study provided a novel HlyB variant containing four substitutions (T36L/F216W/S290C/V421I) with up to 2.5-fold improved secretion for two hydrolases, a lipase and a cutinase. Key points • An improvement in protein secretion via the use of T1SS • Reaching almost 400 mg/L of soluble lipase into the supernatant • A step forward to making E. coli cells more competitive for applying as a secretion host Graphical Abstract

Protein engineering through directed evolution and (semi-)rational approaches has been applied successfully to optimize protein properties for broad applications in molecular biology, biotechnology, and biomedicine. The potential of protein engineering is not yet fully realized due to the limited screening throughput hampering the efficient exploration of the vast protein sequence space. Data-driven strategies have emerged as a powerful tool to leverage protein engineering by providing a model of the sequence-fitness landscape that can exhaustively be explored in silico and capitalize on the high diversity potential offered by nature However, as both the quality and quantity of the inputted data determine the success of such approaches, the applicability of data-driven strategies is often limited due to sparse data. Here, we present a hybrid model that combines direct coupling analysis and machine learning techniques to enable data-driven protein engineering when only few labeled sequences are available. Our method achieves high performance in predicting a protein’s fitness based on its sequence regardless of the number of sequences-fitness pairs in the training dataset. Besides reducing the computational effort compared to state-of-the-art methods, it outperforms them for sparse data situations, i.e., 50 − 250 labeled sequences available for training. In essence, the developed method is auspicious for data-driven protein engineering, especially for protein engineers who have only access to a limited amount of data for sequence-fitness landscape modeling.

In recent years, the engineering of flexible loops to improve enzyme properties has gained attention in biocatalysis. Herein, we report a loop engineering strategy to improve the stability of the substrate access tunnels, which reveals the molecular mechanism between loops and tunnels. Based on the dynamic tunnel analysis of CYP116B3, five positions (A86, T91, M108, A109, T111) in loops B-B′ and B′-C potentially affecting tunnel frequent occurrence were selected and subjected to simultaneous saturation mutagenesis. The best variant 8G8 (A86T/T91L/M108N/A109M/T111A) for the dealkylation of 7-ethoxycoumarin and the hydroxylation of naphthalene was identified with considerably increased activity (134-fold and 9-fold) through screening. Molecular dynamics simulations showed that the reduced flexibility of loops B-B′ and B′-C was responsible for increasing the stability of the studied tunnel. The redesign of loops B-B′ and B′-C surrounding the tunnel entrance provides loop engineering with a powerful and likely general method to kick on/off the substrate/product transportation.

The combination of diversity generation methods and ultrahigh-throughput screening (uHTS) technologies is key to efficiently explore nature\'s sequence space and elucidate structure-function relationships of enzymes. Beneficial substitutions often cluster in a few regions and simultaneous amino acid substitutions at multiple positions (e.g., by OmniChange) will likely lead to further improved enzyme variants. An extensive screening effort is required to identify such variants, as the simultaneous randomization of four codons can easily yield over 105 potential enzyme variants. The combination of flow cytometer-based uHTS with cell-free compartmentalization technology using (w/o/w) double emulsions (InVitroFlow), provides analysis capabilities of up to 107 events per hour, thus enabling efficient screening. InVitroFlow is an elegant solution since diversity loss through a transformation of host cells is omitted and emulsion compartments provide a genotype-phenotype linkage through a fluorescence readout. In this work, a multi-site saturation mutagenesis (mSSM) and an OmniChange library with four simultaneously saturated positions in the active site of CelA2 cellulase were screened using InVitroFlow. Screening of over 36 million events, yielded a significantly improved cellulase variant CelA2-M3 (H288F/H524Q) with an 8-fold increase in specific activity compared to the parent CelA2-H288F (83.9 U/mg) and a 41-fold increased specific activity (674.5 U/mg) compared to wildtype CelA2 (16.6 U/mg) for the substrate 4-MUC (4-methylumbelliferyl-β D-cellobioside).

Ionic liquids (IL) and aqueous ionic liquids (aIL) are attractive (co-)solvents for green industrial processes involving biocatalysts, but often reduce enzyme activity. Experimental and computational methods are applied to predict favorable substitution sites and, most often, subsequent site-directed surface charge modifications are introduced to enhance enzyme resistance towards aIL. However, almost no studies evaluate the prediction precision with random mutagenesis or the application of simple data-driven filtering processes. Here, we systematically and rigorously evaluated the performance of 22 previously described structure-based approaches to increase enzyme resistance to aIL based on an experimental complete site-saturation mutagenesis library of Bacillus subtilis Lipase A (BsLipA) screened against four aIL. We show that, surprisingly, most of the approaches yield low gain-in-precision (GiP) values, particularly for predicting relevant positions: 14 approaches perform worse than random mutagenesis. Encouragingly, exploiting experimental information on the thermostability of BsLipA or structural weak spots of BsLipA predicted by rigidity theory yields GiP = 3.03 and 2.39 for relevant variants and GiP = 1.61 and 1.41 for relevant positions. Combining five simple-to-compute physicochemical and evolutionary properties substantially increases the precision of predicting relevant variants and positions, yielding GiP = 3.35 and 1.29. Finally, combining these properties with predictions of structural weak spots identified by rigidity theory additionally improves GiP for relevant variants up to 4-fold to ∼10 and sustains or increases GiP for relevant positions, resulting in a prediction precision of ∼ 90% compared to ∼ 9% in random mutagenesis. This combination should be applicable to other enzyme systems for guiding protein engineering approaches towards improved aIL resistance.