The core problem the team set out to solve is a mismatch of geometry and chemistry. NMNH is a truncated analog of natural cofactors: it keeps the nicotinamide ring that delivers hydride transfer but sheds most of the AMP moiety that natural enzymes use to recognize and bind their cofactors. Most wild-type enzymes either cannot use NMNH at all or do so only weakly. In P450BM3, the reductase domain normally passes electrons along the route NADPH to FAD to FMN to heme, and productive transfer requires the nicotinamide ring to sit close to the FAD cofactor. When the researchers modeled the wild-type enzyme bound to NMNH, they found the distance between the FAD-N5 atom and the NMNH-C4 hydrogen was 5.2 angstroms, too long for efficient electron transfer, which helps explain the very low activity of the unmodified enzyme with the mimic.

Because no crystal structure of P450BM3 bound to NADPH existed, the team first had to reconstruct that complex. They complemented an incomplete NADPH pose in the deposited structure 4DQL using the homologous structure 1J9Z, relaxed the model with molecular dynamics sampling, and then truncated NADPH to NMNH. Notably, AlphaFold3 predicted the protein backbone accurately but failed to reproduce the ligand positioning and conformational state of the crystalline enzyme, a limitation the authors cite as justification for their kinetically sampled modeling approach. The reconstructed complex revealed a network of residues, including Ser831, Thr846, and Cys1000 around the flavin ring, plus Thr846, Ser848, Trp1047, and Asp943 contacting the cofactor, that together underpin the enzyme’s natural NADPH dependence.

From these analyses the researchers distilled two design principles: the biomimetic cofactor must be locked into an interaction network that reproduces the electron transfer pattern of the natural cofactor, and geometric constraints should shorten the donor-acceptor distance between the nicotinamide moiety and the flavin ring. The challenge was turning those principles into a repeatable, quantitative workflow. Earlier rational design efforts had shown promise, including a glucose dehydrogenase engineered for 107-fold better NMNH recognition and variants identified by imposing distance constraints between NMNH and flavin cofactors, but the field lacked a unified, process-oriented pipeline that integrated cofactor positioning, conformational adaptation, and electron-transfer requirements.

HelixWings fills that gap with a four-step procedure. First, molecular dynamics trajectories of the enzyme in NMNH-bound and unbound states are fed into ICNNMD, an interpretable convolutional neural network framework that identifies the ten residues contributing most to conformational change, with deliberate emphasis on positions distal to the binding pocket to avoid local optima traps. Second, amino acids within four angstroms of the cofactor are screened for interaction analysis. Third, the two lists are cross-referenced to nominate top hotspot residues. Fourth, saturation mutagenesis of those hotspots is evaluated computationally under conformational constraints, with calculated binding energy changes used to rank a mutant library for experimental testing. As a validation, applying the scheme to the wild-type enzyme correctly flagged the previously reported beneficial positions W1047 and S848 as top hotspots.

Starting from the known beneficial mutations S848R and W1047S, which together improved NMNH-supported activity 3.2-fold over wild type, the pipeline singled out R680 and T906 as top hotspots in P450BM3. Experimental screening of the top twenty ranked variants, using cytochrome c as a substrate to quantify NMNH-supported electron transfer, identified S848R/W1047S/T906A as the standout, with a specific activity of 2.19 times ten to the third units per milligram. Its apparent catalytic efficiency of 11.3 per minute per micromolar represents a 404 percent increase over the wild-type value of 2.8. The T906S variant reached only 6.7, a marginal gain over the single S848R mutant, showing that not every computationally ranked mutation pays off, an outcome the authors attribute to the fact that calculated binding energies reflect interaction stability in a modeled state while catalytic efficiency also depends on dynamics and turnover.

Crucially, the team confirmed that the best variant supports full physiological catalysis, not just cofactor oxidation. With the substrate 12-pNCA, wild-type P450BM3 oxidized efficiently under NADPH conditions at a catalytic efficiency of 2.2 per minute per micromolar but showed no detectable oxidation under NMNH conditions. The triple mutant, by contrast, reached 2.0 under NMNH conditions, essentially matching the natural cofactor performance. Molecular dynamics simulations explained why T906A worked so well: alanine at position 906 maintained a hydrogen bond to NMNH with 79 percent occupancy, effectively locking the cofactor in an optimal binding position, while allowing neighboring residues to orient the nicotinamide C4 atom toward the flavin ring. The engineered arginine at position 848 appears to electrostatically compensate for the AMP recognition handle that NMNH lacks compared with NADPH.

To test generalizability, the researchers applied the same workflow to CYP116B46, a self-sufficient P450 whose reductase domain uses a different electron transfer route, FMN to a 2Fe-2S cluster to heme. The pipeline nominated residues D534, C653, and five top-ranked candidates from the AI model, and experimental screening yielded two improved variants, H501M and Q514F, with catalytic efficiency gains of 125 and 217 percent respectively over wild type for cytochrome c reduction. In an oxidation assay with 2-hydroxyphenylacetic acid, the Q514F variant generated a detectable product signal under NMNH conditions where the wild type produced none, although a semi-quantitative comparison indicated the mutant still falls short of the wild type’s NADPH-supported output. Interestingly, the beneficial CYP116B46 mutations sat farther from the immediate cofactor-contact environment than those in P450BM3, suggesting that in this system conformational regulation and longer-range structural effects dominate, and that the platform can identify which mechanistic layer matters for a given enzyme architecture.

The authors also quantified the computational savings. Exhaustive saturation mutagenesis across the 229 residues of the CYP116B46 reductase-domain model would take an estimated 916 hours, whereas ICNNMD-based prioritization reduced the candidate set to ten residues and roughly 40 hours of computation. A web interface at lujialab.org.cn/server makes the workflow accessible to other laboratories, and the code has been deposited publicly on GitHub.

The researchers are careful about the boundaries of their results. Experimental validation covered only self-sufficient P450 systems, and it remains unclear how well the quantitative design principles transfer to other cofactor-dependent enzyme families. The workflow’s reliability depends on the quality of the initial structural model and the adequacy of molecular dynamics sampling, and the conformational analysis relies on a single AI framework that has yet to be benchmarked broadly against alternatives. The substrate oxidation experiments confirm productive NMNH-supported catalysis but do not establish that the mimic is universally equivalent to NADPH. The authors likewise stress that AI-predicted residues should be treated as enriched candidates for mechanistic screening rather than deterministic winners, since experimental validation remains indispensable. Within those limits, HelixWings offers a mechanistically interpretable route to cheaper cofactor-dependent biocatalysis, one that pairs the pattern-recognition power of deep learning with the physical reasoning of classical enzyme engineering.

Subject of Research: AI-assisted computational enzyme engineering for biomimetic nicotinamide cofactor utilization in P450 systems

Article Title: HelixWings enables AI-driven enzyme engineering for biomimetic cofactor utilization

Article References: Zhang, C., Feng, Y., Ji, Y., Rong, M., Zhang, R., Yuan, Z., Shen, H., Tong, T., Miao, T., Fang, B., Tong, L., Fang, W., Wang, Z., Wei, H., Gao, B., & Zhang, L. (2026). HelixWings enables AI-driven enzyme engineering for biomimetic cofactor utilization. iScience, 29(11), Article 117873. https://doi.org/10.1016/j.isci.2026.117873

Image Credits: AI Generated

DOI: 10.1016/j.isci.2026.117873

Keywords: HelixWings, enzyme engineering, NMNH, biomimetic cofactors, cytochrome P450, ICNNMD, molecular dynamics, directed evolution, rational design, biocatalysis, P450BM3, CYP116B46