Model

AnTigone targets hyperandrogenism driven by the gut-androgen axis, a pathway in which colonic bacteria convert steroid precursors into bioactive androgens, particularly testosterone, which are then reabsorbed into circulation. Our dry-lab work asks which engineered enzyme variants are worth building to reduce testosterone, and how those variants would be expected to behave once placed in a living bacterial system.

Two enzymes are central to this axis. AKR1D1 is the human hepaticOf the liver. Δ4-3-ketosteroid 5β-reductaseAn enzyme that adds hydrogen to a steroid that has a double bond at position 4 and a ketone (C=O) at position 3., a member of the aldo-keto reductase superfamilyA large family of enzymes that reduce aldehydes and ketones using NADPH., which inactivates androgens and glucocorticoidsA class of steroid hormones, such as cortisol, that regulate metabolism and the immune response. in the liver and initiates bile acid synthesisThe liver's production of bile acids, which help digest fats. (Chen & Penning, 2014; Faucher et al., 2008).

ci2350 is a bacterial Δ4-3-ketosteroid 5β-reductase from Clostridium innocuum, identified by Arp et al. (2025) as a member of the Old Yellow Enzyme family (COG1902)Flavin-dependent enzymes that reduce carbon-carbon double bonds., a flavin-dependent ene-reductase that couples steroid reduction to NADH oxidation under the anaerobic conditionsConditions without oxygen. of the gut. It has been shown to reduce progesterone, cortisone and hydrocortisone directly, and C. innocuum extracts have been reported to reduce testosterone (Arp et al., 2025; Stokes & Hylemon, 1985).

The two enzymes are not homologousSharing a common evolutionary origin.: they belong to unrelated families and have converged independently on the same 5β-reduction chemistry, with AKR1D1 acting on Δ4-3-ketosteroids spanning C18 to C21 classesSteroids grouped by how many carbon atoms they have; C21 steroids include progesterone and cortisol. (Chen et al., 2011), while ci2350 has been shown to reduce C21 steroids (Arp et al., 2025).

The selectivity problem

The capacity of both enzymes to accept testosterone as a substrate makes them candidates for engineering, but wild-typeThe natural, unmodified version of an enzyme. activity alone is insufficient. Effective therapeutic application requires that the enzyme preferentially reduces testosterone over structurally similar steroids, and progesterone is the most demanding comparison. The two molecules share the same core steroid scaffold and differ critically at a single position: testosterone carries a 17β-hydroxyl group, whereas progesterone bears a C17 acetyl substituent (Chen & Penning, 2014). Any engineered selectivity must therefore be achieved by discriminating between these two chemically subtle substituents on an otherwise identical molecule.

This problem cannot be addressed experimentally without first knowing which variants are structurally worth testing. The sequence space of possible active siteThe pocket of an enzyme where the substrate binds and is converted. mutations is far larger than can be screened at the bench, so a computational filter is needed to narrow it to a tractable set.

Binding affinity alone, however, does not predict how a variant would behave in a bacterial system. Cell growth, enzyme expression levels, and competition between substrates all shape the outcome, and a variant that binds testosterone well in isolation may still perform poorly when those factors are considered.

Our approach

Our dry lab work addresses both challenges through two integrated strands.

  1. Enzyme specificity: Perform rational mutagenesisChoosing specific mutations based on the protein's structure, rather than mutating at random. on ci2350 and AKR1D1 that make the enzymes prefer testosterone over progesterone, using molecular dockingA computer simulation that predicts how well a molecule fits into a protein's binding site. to guide which residues to change.
  2. Mathematical modeling: Build a mechanistic model of bacterial growth, enzyme expression and competitive kinetics that uses the docking results to predict how each variant would perform in practice and which parameters most affect the prediction.

Together, these form a connected predictive layer: the first narrows a large design space to a small set of structurally plausible candidates, and the second takes those candidates and asks whether their effect would actually show up in a real bacterial system.

Enzyme specificity engineering

Figure 1. Computational pipeline for engineering steroid substrate specificity. Structures from the PDB (AKR1D1, 3CAV) or AlphaFold (ci2350) were prepared for docking; the functional pocket was confirmed with fpocket v4.0 and CD Search; testosterone and progesterone were docked in AutoDock Vina 1.2 to identify residues within 4 Å of C17; mutants there were screened in PyRosetta (ΔΔG > +5 REU excluded) and re-docked for selectivity. Picture made with BioRender

Structural Analysis and Active-Site Mapping

Protein structure acquisition

Three-dimensional structures for the two candidate enzymes were obtained through different routes, reflecting the availability of experimental data for each. The human AKR1D1 structure was retrieved from the RCSB Protein Data BankA public database of experimentally determined 3D structures of proteins. (PDB entry 3CAV) (Berman et al., 2000), resolved at 1.90 ÅÅngström, a unit of length equal to one ten-billionth of a meter. A smaller number means a sharper structure. in complex with NADPH (Faucher et al., 2008).

  • Drag rotate
  • ScrollPinch zoom
  • Right-dragTwo fingers pan
Model 1. Crystal structure of human AKR1D1 (PDB 3CAV, chain A) in complex with NADPH. NADPH is highlighted in cornflower blue

No experimentally determined structure was available for ci2350. A three-dimensional model was generated using AlphaFoldAn AI program that predicts a protein's 3D structure from its amino acid sequence. v2.2.0 (Jumper et al., 2021), with the input sequence taken from Arp et al., 2025. Per-residue confidence scores (pLDDTAlphaFold's per-residue confidence score from 0 to 100. Above 90 is very high.) exceeded 90 (pLDDT > 90, very high) across the model, confirming that the predicted structure met the threshold for structure-based analysis.

  • Drag rotate
  • ScrollPinch zoom
  • Right-dragTwo fingers pan
Model 2. AlphaFold v2.2.0 model of ci2350 from Clostridium innocuum. Per-residue confidence exceeded 90 (very high) across the model.

Structure preparation

All structures were prepared following the standard AutoDockFree docking software that predicts how a small molecule fits into a protein. receptor preparation protocol (Eberhardt et al., 2021; Morris et al., 2009). Water molecules and non-essential heteroatomsAtoms other than carbon and hydrogen, such as water, ions or bound molecules in a structure file. were removed. NADPH was present in the PDB structure of AKR1D1 and was retained to preserve the biologically relevant binding geometry. For ci2350, no cofactor is present, since it is a predicted structure.

Hydrogen atoms were added, and protonation statesWhether each chemical group carries an extra hydrogen ion at a given pH. were assigned at physiological pH (7.4), consistent with standard receptor preparation for molecular docking (Morris et al., 2009). Structures were then converted to PDBQTThe file format AutoDock uses for docking inputs, storing atom positions, charges and rotatable bonds. format using MGLTools (prepare_receptor4.py) (Morris et al., 2009), which assigns Gasteiger partial chargesA fast method for estimating the electric charge on each atom. and defines rotatable bonds for the docking software.

Ligand preparation

Testosterone and progesterone share the four-ring steroid scaffold, with rings labeled A to D from the end bearing the C3 ketone (IUPAC-IUB Joint Commission on Biochemical Nomenclature, 1989). They are identical across rings A to C, including the Δ4-3-keto motif on the A ring that is the site of reduction, and differ only at the C17 substituent on the D ring, with 17β-hydroxyl in testosterone and an acetyl group in progesterone (Model 3 and Model 4).

  • Drag rotate
  • ScrollPinch zoom
  • Right-dragTwo fingers pan
Model 3. 3D structure of testosterone retrieved from PubChem CID 6013
  • Drag rotate
  • ScrollPinch zoom
  • Right-dragTwo fingers pan
Model 4. 3D structure of progesterone retrieved from PubChem CID 5994

Both ligandsThe small molecules that bind to a protein; here the steroids. were retrieved from PubChemA public database of chemical molecules and their structures. (CID 6013 and CID 5994, respectively) (Kim et al., 2025) as three-dimensional SDF structures. Each ligand was processed using MGLTools (prepare_ligand4.py) (Morris et al., 2009) to assign partial charges and define rotatable bonds, generating PDBQT files for docking input. No additional energy minimizationAdjusting a molecule's shape to the arrangement with the lowest energy. was performed prior to docking, as AutoDock Vina performs internal conformational samplingTrying out many shapes (conformations) of the molecule. through multiple independent search runs during the docking process (Eberhardt et al., 2021).

Conserved domain analysis

The sequence of each candidate enzyme was submitted to the NCBI Conserved Domain Search (CD Search) (Marchler-Bauer et al., 2017), which compares a query sequence against the Conserved Domain Database (CDD) to identify evolutionarily conserved functional domainsStretches of protein found across a family of related proteins, usually because they matter for function. and their annotated residue positions. These annotations cover substrate binding, cofactor coordination and catalytic function, and are inferred from structurally and biochemically characterized members of the same protein family. They provided a sequence-level map of the expected functional site and served as the reference for interpreting and validating the predicted active site pockets.

For AKR1D1:

CD Search assigned the query sequence (NP_005980.1) to the AKR_AKR1D1-3 family within the aldo-keto reductase superfamily, returning 32 annotated active site positions and a separately annotated catalytic tetradThe four residues that together carry out the chemical reaction. (Figure 2 and Table 1).

The enzyme uses NADPH and a conserved catalytic tetrad of D53, Y58, K87 and E120 to deliver a hydrideA hydrogen atom carrying two electrons, which the enzyme passes to the steroid. to the Δ4 double bond on the steroid A-ring. Y58 acts as the general acid, with K87 and D53 tuning its acidity. E120 is what distinguishes AKR1D1 from the rest of the superfamily: it replaces the histidine conserved in all other AKRs and acts as a superacid, polarizing the C3 ketone (Di Costanzo et al., 2008).

Table 1. Conserved domain features of AKR1D1 (NP_005980.1) from NCBI CD Search. The four tetrad residues are a subset of the active site set.
Active siteCatalytic tetrad
G24, T25, Y26, D53, Y58, K87, W89, E120, Y132, S169, N170, Q193, Y219, S220, P221, L222, G223, T224, S225, W230, L239, A256, I271, P272, K273, S274, F275, N276, R279, E282, N283, W314D53, Y58, K87, E120

For ci2350:

For ci2350, CD Search assigned the query sequence to the OYE_like_FMN_family within the TIM barrelA common protein fold in which eight helices and eight strands form a barrel shape. superfamily (Table 2), returning annotated tracks for the active site, substrate binding site, FMNFlavin mononucleotide, a helper molecule that many enzymes use to move electrons. binding site and a putative catalytic residue (Figure 3).

The domain covers the N-terminal half of the 642-residue protein, and all annotated functional positions fall within it. These active site residues were combined with those of Arp et al. (2025), who identified Y167 as the structural counterpart of the catalytic tyrosine in E. coli 2,4-dienoyl-CoA reductase (PDB 1PS9), with N166, S171 and N253 in the surrounding region. This combined set served as the reference for validating the fpocket predictions.

Table 2. Conserved domain features of ci2350 from the NCBI CD Search with the active site residues reported by Arp et al. (2025).
Active siteCatalytic / substrate-binding residues
A24, G26, V59, Q100, H165, N166, Y167, S171, R215, N253, G315, R316H165, N166, Y167, S171, N253

CD Search Visualization

Active site identification

Binding pockets were identified computationally using fpocket v4.0 (Le Guilloux et al., 2009), which detects cavities through Voronoi tessellation and alpha-sphere clusteringGeometric methods: the first divides space into regions around each atom, the second fits spheres into the gaps between atoms. Groups of neighboring spheres mark a cavity. and returns a ranked list of candidate pockets, each described by geometric and physicochemical descriptors. These predicted pockets were then cross-referenced against residue annotations from NCBI CD Search to identify which pocket best corresponded to the known catalytic and substrate-binding residues of each enzyme family. The pocket showing the greatest agreement with the annotated functional site was selected.

For AKR1D1:

fpocket identified 22 candidate pockets in the AKR1D1 structure. Cross-referencing against the CD Search annotations showed a single dominant match: Pocket 1 contained 29 of the 32 annotated active site residues and all four members of the catalytic tetrad (Table 3). The remaining pockets carried five or fewer annotated residues each, and twelve contained none. Pocket 1 was therefore selected as the functional active site.

Table 3. fpocket results for AKR1D1 cross-referenced against CD Search annotations.
PocketAnnotated residues foundCatalytic tetrad
Pocket 129/32 – G24, T25, Y26, D53, Y58, K87, W89, E120, Y132, S169, N170, Q193, Y219, S220, P221, L222, T224, S225, W230, L239, A256, I271, P272, K273, S274, R279, E282, N283, W3144/4 – D53, Y58, K87, E120
Pocket 205/32 – L222, G223, T224, S225, L2390/4
Pocket 134/32 – Y26, S225, W230, K2730/4
Pocket 152/32 – Y58, F2751/4 – Y58
Pocket 212/32 – Y26, Y1320/4
Pocket 21/32 – N1700/4
Pocket 181/32 – W2300/4

For ci2350:

Similarly, fpocket identified 38 candidate pockets in the predicted ci2350 model. Pocket 1 was the clearest match, containing 9 of the 12 annotated residues, including H165 and Y167. Two further pockets carried catalytic residues: Pocket 5 contained N166 and S171, and Pocket 37 contained Y167 (Table 4).

Table 4. fpocket results for ci2350 cross-referenced against CD Search and Arp et al. (2025).
PocketAnnotated residues foundCatalytic residues
Pocket 19/12 – A24, G26, V59, Q100, H165, Y167, R215, G315, R3162/4 – H165, Y167
Pocket 52/12 – N166, S1712/4 – N166, S171
Pocket 371/12 – Y1671/4 – Y167
Pocket 31/12 – R3160/4

In both enzymes Pocket 1 covered the majority of the annotated active site residues and was therefore used to define the docking search space. The box center was taken as the mean x, y and z coordinates of the pocket's alpha spheres, and each dimension was set from the span of those coordinates along that axis (maximum – minimum) plus a 4 Å margin to ensure sufficient room for the ligand to sample orientations within a tightly enclosed pocket.

Because Pocket 1 did not capture every annotated residue in either enzyme, the box was then expanded as required so that all CD Search positions fell within the search space. After expanding, the final search box parameters for both AKR1D1 and ci2350 are given in Table 5.

Table 5. Docking search box parameters.
ParameterAKR1D1ci2350
Center x (Å)5.1823.793
Center y (Å)27.7984.007
Center z (Å)−35.5273.619
Size x (Å)22.030.0
Size y (Å)29.020.0
Size z (Å)20.022.0

Establishing the Selectivity Baseline

Molecular docking

Docking was performed using AutoDock Vina 1.2 (Eberhardt et al., 2021) with an exhaustivenessA docking setting: higher values search more thoroughly. setting of 32 and an energy range of 4 kcal/mol. Vina combines a stochasticInvolving randomness, so results can differ slightly from run to run. global search for scoring, which samples ligand position, orientation and torsion anglesThe angles at which parts of a molecule can rotate around its bonds., with local gradient-based optimization that refines promising candidates into nearby energy minima.

Each ligand was docked independently into each prepared enzyme structure, and the top-ranked pose was recorded along with its predicted binding affinity in kcal/mol. To establish a selectivity baseline, we defined a selectivity score as the difference in predicted affinity between testosterone and progesterone, ΔG(testosterone)–ΔG(progesterone)ΔG(testosterone) – ΔG(progesterone). A negative value indicates that testosterone is preferred by the enzyme. The inherent stochastic variability of the Vina scoring function, approximately ±0.5 kcal/mol, was treated as the noise threshold (Eberhardt et al., 2021), and selectivity differences below this value were not considered meaningful.

After docking, the two enzymes gave comparable absolute affinities but opposite directions of preference (Table 6).

Table 6. Docking affinities and selectivity scores.
EnzymeProgesterone (kcal/mol)Testosterone (kcal/mol)Selectivity (ΔGT − ΔGP)Preference
AKR1D1−9.8−9.3+0.5Progesterone
ci2350−9.1−9.2−0.1Testosterone

For AKR1D1:

AKR1D1 bound progesterone 0.5 kcal/mol more tightly than testosterone. Converting this through ΔΔG(∣ΔGT–ΔGP∣)=RTln(ratio)ΔΔG (|ΔGT – ΔGP|) = RT ln(ratio), with RT = 0.616 kcal/mol at 310 K, gives approximately 2.3-fold selectivity for progesterone. This is consistent with AKR1D1's biological role, where progesterone intermediates are the primary physiological substrates. The docking visualization is shown below (Figure 4 and Figure 5).

Figure 4. Top-ranked docking pose of testosterone in the AKR1D1 active site (−9.3 kcal/mol), with NADPH and the catalytic tetrad Y58 labeled.

Figure 5. Top-ranked docking pose of progesterone in the AKR1D1 active site (−9.8 kcal/mol), with NADPH and the catalytic tetrad Y58 labeled.

For ci2350:

ci2350 showed a marginal preference for testosterone (−0.1 kcal/mol, ~1.2-fold), well within the noise floor and cannot be read as meaningful selectivity (Table 6). The useful conclusion is negative, which makes ci2350 a workable starting point for engineering a preference in the required direction, since selectivity must be increased rather than reversed. The docking visualization is shown below (Figure 6 and Figure 7).

Figure 6. Top-ranked docking pose of testosterone in the ci2350 active site (−9.2 kcal/mol), with catalytic residue Y167 labeled

Figure 7. Top-ranked docking pose of progesterone in the ci2350 active site (−9.1 kcal/mol), with catalytic residue Y167 labeled.

Neither wild-type meets the selectivity requirement, and both were taken forward to contact analysis for rational mutagenesis.

Contact analysis and residue identification

From the top-ranked docking pose, all enzyme residues within 4 Å of each docked ligand were identified. This is a standard cut-off for direct protein–ligand contacts and captures residues involved in van der WaalsWeak attractions between atoms that are very close to each other. interactions and short-range electrostaticsAttraction or repulsion between electrically charged groups. (Le Guilloux et al., 2009). Residues in proximity to the C17 position were given particular attention, as this is the only position at which the two substrates differ, as discussed previously in the ligand preparation (Chen & Penning, 2014).

For AKR1D1:

Both substrates share an identical Δ4-3-keto motif on the A ring, so the residues recognizing that end cannot distinguish between them. Hence, the six shared contacts, two of which (Y58, E120) are the catalytic tetrad (Table 7), confirm that both ligands dock in a catalytically productive orientation.

Progesterone contacts two hydrophobic residues, V309 and L311, which can form a nonpolar surface that accommodates its acetyl group, while testosterone contacts I57, which could offer a hydrogen bondA weak attraction between a hydrogen atom on one molecule and an oxygen or nitrogen atom on another. for the 17β-hydroxyl. V309 and L311 were targeted against progesterone, while I57 favors testosterone to increase affinity. All four were mutated singly and in combination.

Table 7. Contact residues within 4 Å of the docked ligands in AKR1D1.
Contact categoryResiduesRole
Shared (both steroids)Y26, Y58, E120, W230, M313, W314A/B/C ring contacts – conserved Δ4-3-ketosteroid recognition, including Y58, E120; exclude from mutations
Progesterone onlyV309, L311C17 acetyl contacts – primary mutagenesis targets
Testosterone onlyI5717β-hydroxyl contacts – targeted to favor testosterone

For ci2350:

As in AKR1D1, the shared contacts cluster around the conserved Δ4-3-keto end of the molecule and include the catalytic residue, indicating that both ligands adopt a productive orientation. Y167 was excluded from mutagenesis, since altering it would be expected to compromise catalysis rather than selectivity.

A107, R215 and V263 appear to line the space progesterone's acetyl group occupies (Table 8). R215 is the most interesting of the three: the other two are small hydrophobic residues, but arginine can donate hydrogen bonds, so its contact may be a specific polar interaction rather than a shape fit, which would make it the cleanest target, along with V59, F71, Q100, M254 and L259 to increase the specificity towards testosterone.

Table 8. Contact residues within 4 Å of the docked ligands in ci2350.
CategoryResiduesInterpretation
Shared (both steroids)A24, M25, G26, M110, Y167, E262, Y347, F350, T351A/B/C ring contacts – conserved Δ4-3-ketosteroid recognition, including the catalytic Y167; exclude from mutations
Progesterone onlyA107, R215, V263C17 acetyl contacts – targeted to weaken progesterone binding
Testosterone onlyV59, F71, Q100, M254, L259Possible 17β-hydroxyl contacts – targeted to favor testosterone

Rational mutagenesis

Rational mutagenesis is a protein engineering approach in which specific amino acid substitutions are chosen in advance from structural and mechanistic knowledge of the enzyme, rather than generating them at random and screening (Lutz, 2010).

The residues in Table 7 and Table 8 were substituted on physicochemical grounds, altering the size, polarity or hydrogen-bonding capacity of the pocket near C17 in a direction consistent with a shift towards testosterone (Alford et al., 2017). Both single-point variants and multi-residue combinations were designed for both enzymes.

For AKR1D1:

At V309 and L311, the substitutions cover three different mechanisms: serine (S) and threonine (T) shrink the side chainThe part of an amino acid that differs between the 20 types and gives it its properties. and add polarity to the pocket; tyrosine (Y), phenylalanine (F) and tryptophan (W) fill the cavity so the acetyl group has less room; and glutamate (E) and lysine (K) place a charge against a neutral substituent (Table 9). Comparing them should indicate which property actually governs discrimination towards progesterone.

At I57, all five substitutions replace an aliphaticMade of plain carbon chains, without charges or rings. side chain with one capable of hydrogen bonding, any of which could pair with testosterone's 17β-hydroxyl. Combinations test whether the two mutation effects are independent. If they are, the selectivity shifts should roughly sum.

Table 9. Mutagenesis design for AKR1D1
StrategyTargetMutations designedRationale
Disrupt progesterone C17 contactsV309, L311V309T/S/Y/F/W/E/K, L311S/T/Y/F/ERemove the hydrophobic surface stabilizing the acetyl group
Enhance testosterone 17β-OH recognitionI57I57S/T/N/Q/YIntroduce a hydrogen-bond donor or acceptor near the 17β-hydroxyl
CombinationsV309 + L311, V309 + L311 + I57Various combinationTest additive effect of multiple substitutions

For ci2350:

At A107 and V263, the substitutions add volume, narrowing the space progesterone's acetyl group occupies. For R215, the approach is removal rather than bulk. R215A and R215G strip the side chain back, which would eliminate any hydrogen bond it forms with the acetyl carbonyl (Table 10).

For testosterone, F71Y adds a hydroxyl to an aromatic ring already in contact with the 17β-hydroxyl, and V59S/T and Q100S/T introduce or shorten a polar side chain so the hydrogen bond geometry can be optimized. Similarly, combination tests were performed.

Table 10. Mutagenesis design for ci2350.
StrategyTargetMutations designedRationale
Disrupt progesterone C17 contactsA107, R215, V263A107V/L/F, R215A/G/E, V263E/K/D/I/L/FAdd bulk to crowd the acetyl group (A107, V263), or remove the R side chain and its possible hydrogen bond (R215)
Enhance testosterone 17β-OH recognitionV59, F71, Q100V59S/T, F71Y, Q100S/TIntroduce or reposition a hydrogen-bonding group for the 17β-hydroxyl
CombinationsMultiple positionsVarious combinationTest additive effects across both strategies

Screening and mutant analysis

Stability screening

Structural stability was evaluated using PyRosetta (Chaudhury et al., 2010) with the ref2015 scoring function (Alford et al., 2017). The ref2015 function returns a single score summing van der Waals packing, hydrogen bonding, electrostatics, implicit solvationTreating the surrounding water as a continuous medium instead of modeling individual water molecules. and backbone and sidechain torsional strain, with lower, more negative values corresponding to a better-packed and more stable structure.

Local sidechain repackingRe-optimizing the positions of nearby amino acid side chains after a mutation. was performed within a 5.0 Å radius of the mutated position, capturing the first layer of residues most likely to be directly affected by the substitution (Chaudhury et al., 2010). Stability was expressed as the difference between mutant and wild-type (ΔΔG=E(mutant)−E(wild−type))(ΔΔG = E(mutant) − E(wild-type)), canceling the common baseline and isolating the effect of the substitution. A high positive ΔΔG indicates that the mutation destabilizes the structure (Alford et al., 2017).

Variants with ΔΔG above +5 REURosetta Energy Units, the score PyRosetta uses. Lower means more stable. were excluded; those below this threshold were considered structurally acceptable and advanced to docking (Alford et al., 2017; Kellogg et al., 2011). The threshold is deliberately permissive. It is set to reject severely destabilized variants, such as unresolvable steric clashesAtoms placed so close together that they overlap and push each other apart. or lost buried hydrogen bonds.

For AKR1D1:

24 designed variants were evaluated for stability checks, where 13 passed the +5 REU threshold and were carried forward; 11 were excluded (Table 11).

Table 11. Stability screening results for AKR1D1 calculated using ΔΔG = E(mutant) − E(wild-type), with a threshold of +5 REU.
VariantΔΔG (REU)Outcome
V309T−2.29Pass
V309S−0.80Pass
V309E−3.29Pass
V309W+42.80Excluded
V309K+8.32Excluded
V309Y
+793.17
Excluded
V309F
+777.77
Excluded
L311S
−3.19
Pass
L311E
−2.42
Pass
L311T
−1.91
Pass
L311Y
−1.68
Pass
L311F
−0.95
Pass
I57Y
0.00
Pass
I57N
+2.69
Pass
I57S
+4.23
Pass
I57Q
+7.87
Excluded
I57T
+9.69
Excluded
Combo 1: V309T + L311S
−3.44
Pass
Combo 6: V309T + L311S + I57S
+0.80
Pass
Combo 5: V309W + L311Y
+30.34
Excluded
Combo 2: V309Y + L311S
+34.29
Excluded
Combo 4: V309Y + L311E + I57N
+36.74
Excluded
Combo 7: V309Y + L311S + I57T
+43.97
Excluded
Combo 3: I57S + V309Y
+797.40
Excluded

V309 is tightly packed and rejected every aromatic substitution, with tyrosine and phenylalanine returning values above +770 REU, resulting in clashes. L311 tolerated every substitution, including tyrosine and phenylalanine, making it the better position for adding bulkier ones. At I57, only short polar side chains passed, threonine failing where serine succeeded. Combinations containing V309Y or V309W were excluded regardless of their partner.

For ci2350:

31 designed variants were evaluated, where 17 passed the +5 REU threshold and were carried forward; 14 were excluded (Table 12).

Table 12. Stability screening results for ci2350 calculated using ΔΔG = E(mutant) − E(wild-type), with a threshold of +5 REU.
VariantΔΔG (REU)Outcome
R215A−3.54Pass
R215E−0.13Pass
R215G+0.54Pass
A107L+18.14Excluded
A107F+53.92Excluded
A107V
+56.56
Excluded
V263F
−2.45
Pass
V263D
+1.36
Pass
V263K
+2.21
Pass
V263E
+2.36
Pass
V263I
+4.96
Pass (borderline)
V263L
+9.46
Excluded
Q100T
−9.43
Pass
Q100S
−7.66
Pass
V59S
−2.36
Pass
F71Y
+6.83
Excluded
V59T
+9.27
Excluded
Combo 13: R215A + V263I + Q100T
−7.19
Pass
Combo 4: F71Y + Q100T
−2.61
Pass
Combo 8: R215A + V263I
−2.42
Pass
Combo 3: R215A + V263E
−1.13
Pass
Combo 7: R215E + V263E
+2.28
Pass
Combo 9: R215E + V263I
+4.84
Pass (borderline)
Combo 12: R215A + V263I + F71Y
+8.25
Excluded
Combo 1: A107V + R215A
+53.01
Excluded
Combo 2: A107F + R215E
+53.79
Excluded
Combo 14: A107V + R215A + V263I
+57.84
Excluded
Combo 11: A107F + V263I
+58.88
Excluded
Combo 6: A107V + R215A + F71Y
+59.86
Excluded
Combo 10: A107V + V263I
+61.52
Excluded
Combo 5: A107V + F71Y
+63.41
Excluded

A107 failed at every substitution, with values from +18 to +57 REU, so the alanine cannot absorb any added bulk, and even every combination containing A107 failed, regardless of its partner. R215 was the opposite: all three substitutions passed, and R215A was mildly stabilizing, meaning the arginine can be removed without structural cost. Among the testosterone-directed designs, Q100S, Q100T and V59S were clearly stabilizing, while F71Y and V59T fell just outside the threshold. The combinations that survived are those built on R215 and V263.

Mutant docking and candidate prioritization

All variants passing the stability threshold were subjected to the same docking protocol as the wild-type enzyme, with testosterone and progesterone docked independently and selectivity scores calculated as before. Variants demonstrating a consistent selectivity shift beyond the noise threshold were identified as priority candidates for experimental characterization.

For AKR1D1:

Table 13. Mutant docking results for AKR1D1 (top-ranked variants).
VariantProgesterone (kcal/mol)Testosterone (kcal/mol)Selectivity (ΔGT − ΔGP)Outcome
Wild-type−9.8−9.3+0.5Reference
V309S−9.8−9.4+0.4Favored progesterone
V309E−9.8−9.4+0.4Favored progesterone
I57Y−9.8−9.4+0.4Favored progesterone
Combo 1: V309T + L311S−9.9−9.4+0.5Favored progesterone
Combo 6: V309T + L311S + I57S−10.4−9.8+0.6Favored progesterone

No AKR1D1 variant shifted selectivity meaningfully. The three best mutants, V309S, V309E and I57Y, moved the score by just 0.1 kcal/mol (Table 13), well within the noise threshold, and every variant kept a positive score, so progesterone remained preferred throughout. None of the substitutions tested was able to redirect AKR1D1 towards testosterone.

For ci2350:

Table 14. Mutant docking results for ci2350 (top-ranked variants).
VariantProgesterone (kcal/mol)Testosterone (kcal/mol)Selectivity (ΔGT − ΔGP)Outcome
Wild-type−9.1−9.2−0.1Reference
Combo 8: R215A + V263I−8.5−9.4−0.9Highly favored testosterone
V263K−6.9−7.4−0.5Favored testosterone, but affinity reduced
Combo 13: R215A + V263I + Q100T−8.0−8.5−0.5Favored testosterone, but affinity reduced
Combo 9: R215E + V263I−7.4−7.7−0.3Favored testosterone, but affinity reduced
V263D−9.1−9.3−0.2Favored testosterone

The combination mutation of R215A + V263I is the clear leading candidate, with a selectivity score of −0.9 kcal/mol and a shift of −0.8 against the wild-type, and the only variant to exceed the noise threshold (Table 14). It is also the only one where testosterone binding improved in absolute terms while progesterone binding weakened, so the gain comes from both directions at once. The other candidates gained selectivity by losing affinity.

R215A + V263I was therefore selected as one of the candidates for experimental characterization and used for activity modeling. The detailed analysis of variant specificity towards testosterone is presented on the Results page.

The complete implementation of the pipeline for enzyme specificity analysis is available on GitLab.

Modeling enzyme production and activity

Biological systems combine processes that run at different speeds and affect one another, so the outcome of an experiment is rarely obvious from the individual parts. Mathematical models address this by describing each process with an equation whose terms correspond to defined biological events (Kraikivski, 2023). These equations are Ordinary Differential Equations (ODEs)Equations describing how quantities change continuously over time., solved numerically. Writing the system this way states every assumption openly and lets us change one parameter at a time, which is rarely practical at the bench (Alon, 2006).

The specificity analysis found no AKR1D1 variant that shifted preference towards testosterone, while ci2350 produced several promising candidates, so the model was built around ci2350 alone. These findings are described on the Results page.

Integrated growth and expression model

Figure 8. Workflow of the integrated growth and expression model. Growth data and literature values define the parameters of two coupled layers, which are solved numerically and then tested by parameter sweeps and Monte Carlo sampling. Picture made with BioRender

System definition

The system modeled is ci2350 produced from a Twist pET vectorA common plasmid for producing proteins in E. coli from a T7 promoter. in E. coli BL21(DE3)An E. coli strain made for high-level protein production., with the gene under the T7 promoterA strong promoter read by T7 RNA polymerase, which BL21(DE3) makes when IPTG is added. and induced with 0.8 mM IPTGA lab chemical that switches on the T7 system.. The quantity of interest is the amount of enzyme per cell because it determines how much enzyme is available to act on the steroids. It is described on the Parts page.

Growth model

A batch culture in a closed flask grows exponentially while nutrients are plentiful and then slows as they are consumed. The logistic equationA growth equation that starts exponential and levels off as the culture fills its container. is the simplest description of this behavior (Verhulst, 1838).

dX/dt=µmax⋅X⋅(1−X/K)dX/dt = µmax · X · (1 − X/K)

X is the culture density (OD600Light absorbance at 600 nm, used to measure how dense a bacterial culture is.), µmax the maximum specific growth rate (per hour) and K the carrying capacityThe maximum culture density the medium can support., the OD600 at which growth stops. The factor (1−X/K)(1 − X/K) falls from 1 to 0 as the culture fills, so it acts as the brake on growth. A lag phaseThe initial period before bacteria start growing. was not included, so the simulated culture starts to grow immediately from the starting density.

From our experimental data, the cells double every 28 minutes (µmax = 1.48 h⁻¹). We did not measure the plateau, so K was set to OD600 15 (Table 15), a plausible density for a rich medium, and its influence is examined in the sensitivity analysisChanging one input at a time to see how much the result depends on it.. From these parameters, the growth saturates near 7 h.

Table 15. Parameters of the growth model
ParameterValueSource
Maximum specific growth rate (µmax)1.48 h⁻¹Fitted from our experimental data
Carrying capacity (K)15 (OD600)Assumption
Starting density (X₀)0.05 (OD600)Starting value used in the simulation

Expression model

Since mRNAMessenger RNA, the copy of a gene that the cell reads to make a protein. and the enzyme behave differently inside the cell, we examined them separately. m is the mRNA level in the cell and E the enzyme level, both in relative units. mRNA is short-lived: messages are transcribed at rate α and degraded at rate δm, so the cell maintains a small working pool at a steady level, giving the ribosomesThe cell's protein-making machines. a constant supply to work from (Alon, 2006).

dm/dt=α−δm⋅mdm/dt = α − δm · m

At steady-state maintenance

m∗=α/δmm* = α / δm

The enzyme is the opposite. It is built from those messages at rate β. We assume it is stable, so nothing degrades it, and the only loss is growth itself. At each division, the enzyme already made is shared between two daughter cells, so the faster the culture grows, the more it dilutes its own protein (Klumpp et al., 2009). This is where the two models are integrated: dilution takes the culture's current growth rate, µ(t)=µmax⋅(1−X/K)µ(t) = µmax · (1 − X/K), rather than a fixed value. Enzyme therefore stays low while the culture divides fast and accumulates as growth slows (Klumpp et al., 2009).

dE/dt=β⋅m−µ(t)⋅E=β⋅m−µmax⋅(1−X/K)⋅EdE/dt = β · m − µ(t) · E = β · m − µmax · (1 − X/K) · E

Five assumptions:

  1. ci2350’s mRNA turns over like a typical E. coli transcript, since its half-lifeThe time it takes for half of the molecules to break down. has not been measured.
  2. The promoter is fully induced at 0.8 mM IPTG, so transcription runs at a constant rate.
  3. The enzyme is stable, so dilution by growth is its only loss.
  4. Expression does not burden the cell, so it does not slow growth.
  5. mRNA and enzyme are in relative units, since absolute counts would require data we do not have.

Bacterial messages break down within minutes of being made (Bernstein et al., 2002). We used a half-life (3 to 8 mins) to predict the decay constants (δ=ln2/t½)(δ = ln2 / t½) range and used their average of 9.53 h⁻¹. The transcription rate (α) and translation rate (β) are placeholders, as neither has been measured for ci2350 (Table 16).

Table 16. Parameters of the expression model
ParameterValueSource
Transcription rate (α)5 (relative units h⁻¹)Placeholder
mRNA decay constant (δm)9.53 h⁻¹Half-life range, midpoint
Translation rate (β)3 h⁻¹Placeholder

Mathematical simulation and sensitivity of the enzyme level

The integrated three-equation system was solved in R using the deSolve package (Soetaert et al., 2010), with the Runge–KuttaA standard numerical method for solving differential equations. method ode45 over 24 h in 100 time steps, starting from X = 0.05, m = 0 and E = 0. Since several parameters are assumptions or literature ranges rather than measurements, we then asked how much the result depends on them, and which the experiments should pin down first (Saltelli et al., 2008). Each of K, µmax, α and δm was swept by ±20% around its baseline individually, a standard window for local sensitivity, and the whole system was re-run at every step, recording culture density, mRNA level and enzyme level after 24 h as a percentage change from baseline. Because all three are recorded together, a single run shows how much the enzyme level moves and how a growth parameter affects expression, and vice versa. β was not varied separately since the enzyme level depends on α and β only through their product.

Uncertainty analysis of the enzyme level

Sweeps change one parameter at a time, whereas in practice all of them are uncertain together. We therefore ran a Monte Carlo analysisRepeating a calculation many times with randomly drawn inputs to see how uncertain the result is. (Saltelli et al., 2008), drawing 500 parameter sets at random, each value taken from a distribution centered on its baseline with a spread of 10%, and recorded the enzyme level after 24 h for each. The 10% is an assumed level of uncertainty, not a measured one.

The analysis was run three times: with only the growth parameters (K, µmax) varied, with only the expression parameters (α, δm) varied, and with all four varied together. Comparing the three shows which layer contributes most to the uncertainty in the enzyme level and we reported the mean and the 95% interval.

The results of both per-parameter sweep and Monte Carlo analyses of enzyme production are presented on the Results page. The scripts for the integrated model are available on GitLab.

Enzyme activity model

Activity model

Testosterone and progesterone are close in shape and compete for the same active site of ci2350, so each slows down the conversion of the other. This is described by Michaelis–Menten kineticsA model of how fast an enzyme works as the amount of substrate changes. for two competing substrates. Each substrate has its own turnover rate, but the two rates share the same denominator, which makes the substrates compete (Michaelis et al., 1913).

vT=(kcat,T/Km,T)⋅Etot⋅T/(1+T/Km,T+P/Km,P)v_T = (kcat,T / Km,T) · Etot · T / (1 + T/Km,T + P/Km,P)
vP=(kcat,P/Km,P)⋅Etot⋅P/(1+T/Km,T+P/Km,P)v_P = (kcat,P / Km,P) · Etot · P / (1 + T/Km,T + P/Km,P)

T and P are the testosterone and progesterone concentrations (µM), and v_T and v_P are their rates of consumption (µM per hour). E_tot is the enzyme concentration in the culture; Km,T and Km,P are the Michaelis constantsThe substrate concentration at which an enzyme works at half its top speed. Lower means tighter binding. (estimated from docking, see Table 18), and kcat,T and kcat,P are the turnover numbersThe number of substrate molecules one enzyme converts per second when fully busy.. The ratio kcat/Km measures how efficiently each substrate is used, so the preference of a variant is the ratio of kcat,T/Km,T to kcat,P/Km,P.

The enzyme availability comes from the integrated model of enzyme production: enzyme per cell multiplied by culture density, converted to a concentration by a single scaling factor (model assumption). The enzyme in the flask therefore rises as the culture grows, and continues to rise more slowly once growth stops, because the enzyme is no longer diluted.

Model Assumption:

Etot=Escale⋅E⋅XEtot = Escale · E · X

The three layers run together as one system of five equations:

dX/dt=µmax⋅X⋅(1−X/K)dX/dt = µmax · X · (1 − X/K)
dm/dt=α⋅sind(t)−δm⋅mdm/dt = α · sind(t) − δm · m
dE/dt=β⋅m−µmax⋅(1−X/K)⋅EdE/dt = β · m − µmax · (1 − X/K) · E
dT/dt=−vT⋅ssub(t)dT/dt = −v_T · ssub(t)
dP/dt=−vP⋅ssub(t)dP/dt = −v_P · ssub(t)

The first three are the growth and expression layers described in the integrated model of enzyme production; the last two are the substrates being consumed. In the earlier model, the expression started at time zero. Here α is multiplied by s_ind, so that mRNA is made only after IPTG is added, and a second switch, s_sub(t), allows the steroids to be consumed only after they have been added. These switches replicate the bench protocol: s_ind(t) turns transcription on when the culture reaches OD600 0.5, which the growth model identifies at 1.58 h, and s_sub(t) turns substrate consumption on four hours later, when substrate is added and enough enzyme has built up. Both steroids start at 50 µM, and the reaction is followed for eight hours (Table 17).

Seven assumptions underlie these equations which are listed below:

  1. Testosterone and progesterone bind the same site and exclude each other (competitive binding).
  2. NADHA molecule that carries electrons and supplies the enzyme with the hydrogen it adds to the steroid. is present in saturating amounts, so it does not limit the rate.
  3. kcat is the same for both steroids and for all variants.
  4. The amount of enzyme per cell is the same for all ci2350 variants, because they are expressed under the same conditions.
  5. The steroids are available to the enzyme at their added concentrations, without limits from uptake into the cells or from solubility.
  6. Product inhibition is not included.
  7. The ci2350 variants do not change folding or stability.
Table 17. Parameters of activity model
ParameterValueSource
tind – time of inductionOD600 0.5 (1.58 h)Assumed experimental design; time from the growth model
tsub – time of substrate addition4 h after inductionAssumed experimental design
T0, P0 – starting concentration of each steroid50 µMAssumed experimental design
Assay time – length of the simulated reaction8 hAssumed experimental design
kcat – turnover number3600 h⁻¹ (1 s⁻¹)Placeholder, 1 s⁻¹ (low catalysis)
kcat,T / kcat,P – ratio of turnover numbers1Assumption
Escale – enzyme concentration per relative unit per OD0.0004 µMPlaceholder
Km,P (wild type) – Km of the wild type for progesterone0.432 µM(Arp et al., 2025)  (ITC Kd of the wild type)

Linking docking to Km

Figure 9. From docking score to Km. Each docking score (ΔG) is converted to a binding constant, Kd = exp(ΔG/RT), and Km is assumed to follow Kd. The wild-type progesterone score is anchored to its measured Kd (0.432 µM), and each variant's Km shifts by its score difference from wild type. A variant with a lower Km for testosterone than for progesterone prefers testosterone. This link between Km and Kd is an assumption, and docking errors are larger than most differences between variants. Picture made with BioRender

The activity model needs a Michaelis constant (Km) for testosterone and progesterone for every ci2350 variant, but none has been published. The only kinetic information on the enzyme is the wild-type dissociation constant (Kd)A measure of binding strength: a smaller value means tighter binding. for progesterone, 0.432 µM, measured by isothermal titration calorimetryA lab method that measures binding by the heat it gives off. (Figure 9) (Arp et al., 2025).

We therefore propose to use our docking results to estimate the missing values. Docking gives a binding energy ΔG, which relates to Kd=exp(ΔG/RT)Kd = exp(ΔG/RT) (Gilson & Zhou, 2007). Here R is the gas constant and RT = 0.616 kcal/mol at 37 °C, so a more negative ΔG means tighter binding and a smaller Kd. Docking scores are not reliable as absolute energies, since Vina can deviate from experiment by 1 to 2 kcal/mol (Eberhardt et al., 2021; Trott & Olson, 2010), so we used only the differences between scores calculated with the same protocol.

We took the wild type with progesterone as the reference point and gave it the measured value, 0.432 µM, as its Km. This treats Km as proportional to Kd, which is a good approximation when catalysis (kcat) is slower than the release of the steroid from the enzyme (Wilhelmová, 1996). Whether this holds for ci2350 is not known, because no kcat has been measured. From the reference, the Km values of all variants follow in two steps:

  1. Km for progesterone is scaled from the wild-type value by the difference between the progesterone scores of the variant and the wild type (Table 18).
  2. Testosterone has no measured value at all; its Km is calculated from the progesterone Km of the same variant, shifted by the difference between the two steroid scores of that variant (Table 18).
Km,P(variant)=0.432µM⋅exp[(ΔGP,variant−ΔGP,wildtype)/RT]Km,P(variant) = 0.432 µM · exp[(ΔG_P,variant − ΔG_P,wild type) / RT]
Km,T(variant)=Km,P(variant)⋅exp(ΔΔG/RT);ΔΔG=ΔGT−ΔGPKm,T(variant) = Km,P(variant) · exp(ΔΔG / RT); ΔΔG = ΔG_T − ΔG_P

A negative ΔΔG_T–P means that the variant binds testosterone more tightly than progesterone, which gives testosterone the lower Km. We assumed the same kcat for both steroids because docking describes binding and not catalysis. The Km values obtained this way are predictions that guide the experiments, not measured constants. The assumptions behind them (Km proportional to Kd, equal kcat) are tested in the sensitivity analysis and are the first targets of the planned kinetic measurements.

Table 18. Docking scores of the wild type and top 5 variants of ci2350, with the Michaelis constants (Km) calculated from them
VariantΔG progesterone (kcal/mol)ΔG testosterone (kcal/mol)ΔΔG (kcal/mol) [ΔG_T − ΔG_P]Km,P (µM)Km,T (µM)
Wild type−9.1−9.2−0.10.430.37
R215A + V263I−8.5−9.4−0.91.140.27
V263K−6.9−7.4−0.515.36.8
R215A + V263I + Q100T−8.0−8.5−0.52.571.14
R215E + V263I−7.4−7.7−0.36.84.2
V263D−9.1−9.3−0.20.430.31

Sensitivity analysis of activity model

The activity layer adds more uncertain values, such as kcat, E_scale, the wild-type Km,P and the docking scores, on top of the uncertain expression parameters, so two checks were run.

For the variant with the largest predicted preference, R215A + V263I (see Results page), each parameter was varied by ±20% with the others held at baseline, and the progesterone docking score was shifted by ±0.3 kcal/mol to see how a small docking error changes the outcome. The expression settings were swept more widely: α from 0.25 to 4 times baseline, β from 0.4 to 12 times, and the induction point from OD600 0.3 to 1.0.

Uncertainty analysis of activity

Most of the values in the model are guesses, so running it once would only tell us what happens if every guess is right. The Monte Carlo analysis instead ran the model 200 times for each variant, picking a different random value for each uncertain input every time.

Each variant was compared against a wild type given the same random expression and kcat values, so the only difference between the two was the docking scores. Every score was given a random error of about 1 kcal/mol, the typical accuracy of Vina, and the analysis was repeated with 0.5 kcal/mol as a more optimistic case. The other values were drawn over the ranges we consider plausible.

What comes out is a percentage for each variant: how often testosterone is preferred, and how often the variant beats the wild type. The wild type was also compared against a second random copy of itself, which shows how often a difference appears purely by chance when there is no real one.

The results of activity modeling are presented on the Results page. The scripts for the integrated model are available on GitLab.

Limitations

  • Single-run docking: Vina's random seedThe starting number of a random process; a different seed gives slightly different results. gives about ±0.5 kcal/mol of variation, which is at or above most selectivity differences (0.1–0.5 kcal/mol). Averaging at least three runs per ligand–receptor pair would be needed to confirm them.
  • AlphaFold structure: ci2350 has no experimental structure, so subtle side-chain positions in the active site may differ from the real enzyme.
  • Rigid receptor: Vina keeps the protein rigid, so induced fitThe way a protein reshapes slightly when a molecule binds. of flexible residues such as Q100 and R215 is not captured.
  • Fast stability scoring: PyRosetta ΔΔG with local repacking only is an approximation, and a FastRelaxA Rosetta protocol that thoroughly adjusts the whole structure to find a lower-energy arrangement. protocol would be more accurate.
  • Single-pose contacts: The contact analysis used only the top-ranked pose (Mode 1), so the real binding mode may involve different residues.
  • Placeholder parameters: The transcription rate, translation rate, carrying capacity and the enzyme amount per cell are placeholders, and the mRNA half-life is poorly known. The absolute times are therefore not predictions. Only the order of conversion and the trends across the sweeps can be trusted.
  • Km from docking: Km values were derived from docking scores by assuming Km is proportional to Kd. This has not been tested for this enzyme.
  • Unmeasured kcat: The turnover rate for each steroid is unknown and was assumed equal. A testosterone turnover about four times slower than progesterone would remove the advantage of R215A + V263I.
  • Missing cell processes: Steroid uptake into the cell, solubility of the steroids and any toxicity to the cells are not modeled.
  • Incomplete uncertainty analysis: The activity Monte Carlo did not vary the growth rate or carrying capacity, and some draws without a defined half-time were excluded.

References

  1. Alford, R. F., Leaver-Fay, A., Jeliazkov, J. R., O’Meara, M. J., DiMaio, F. P., Park, H., Shapovalov, M. V., Renfrew, P. D., Mulligan, V. K., Kappel, K., Labonte, J. W., Pacella, M. S., Bonneau, R., Bradley, P., Dunbrack, R. L., Das, R., Baker, D., Kuhlman, B., Kortemme, T., & Gray, J. J. (2017). The Rosetta all-atom energy function for macromolecular modeling and design. Journal of Chemical Theory and Computation, 13(6), 3031–3048. https://doi.org/10.1021/acs.jctc.7b00125
  2. Alon, U. (2006). An introduction to systems biology: Design principles of biological circuits. Chapman & Hall/CRC.
  3. Arp, G., Jiang, A. K., Dufault-Thompson, K., Levy, S., Zhong, A., Wassan, J. T., Grant, M. R., Li, Y., Hall, B., & Jiang, X. (2025). Identification of gut bacteria reductases that biotransform steroid hormones. Nature Communications, 16(1). https://doi.org/10.1038/s41467-025-61425-6
  4. Berman, H. M., Westbrook, J., Feng, Z., Gilliland, G., Bhat, T. N., Weissig, H., Shindyalov, I. N., & Bourne, P. E. (2000). The Protein Data Bank. Nucleic Acids Research, 28(1), 235–242. https://doi.org/10.1093/nar/28.1.235
  5. Bernstein, J. A., Khodursky, A. B., Lin, P. H., Lin-Chao, S., & Cohen, S. N. (2002). Global analysis of mRNA decay and abundance in Escherichia coli at single-gene resolution using two-color fluorescent DNA microarrays. Proceedings of the National Academy of Sciences of the United States of America, 99(15), 9697–9702. https://doi.org/10.1073/pnas.112318199
  6. Chaudhury, S., Lyskov, S., & Gray, J. J. (2010). PyRosetta: A script-based interface for implementing molecular modeling algorithms using Rosetta. Bioinformatics, 26(5), 689–691. https://doi.org/10.1093/bioinformatics/btq007
  7. Chen, M., & Penning, T. M. (2014). 5β-Reduced steroids and human Δ4-3-ketosteroid 5β-reductase (AKR1D1). Steroids, 83, 17–26. https://doi.org/10.1016/j.steroids.2014.01.013
  8. Chen, M., Drury, J. E., & Penning, T. M. (2011). Substrate specificity and inhibitor analyses of human steroid 5β-reductase (AKR1D1). Steroids, 76(5). https://doi.org/10.1016/j.steroids.2011.01.003
  9. Di Costanzo, L., Drury, J. E., Penning, T. M., & Christianson, D. W. (2008). Crystal structure of human liver Δ4-3-ketosteroid 5β-reductase (AKR1D1) and implications for substrate binding and catalysis. Journal of Biological Chemistry, 283(24), 16830–16839. https://doi.org/10.1074/jbc.M801778200
  10. Eberhardt, J., Santos-Martins, D., Tillack, A. F., & Forli, S. (2021). AutoDock Vina 1.2.0: New docking methods, expanded force field, and Python bindings. Journal of Chemical Information and Modeling, 61(8), 3891–3898. https://doi.org/10.1021/acs.jcim.1c00203
  11. Faucher, F., Cantin, L., Luu-The, V., Labrie, F., & Breton, R. (2008). The crystal structure of human Δ4-3-ketosteroid 5β-reductase defines the functional role of the residues of the catalytic tetrad in the steroid double bond reduction mechanism. Biochemistry, 47(32), 8261–8270. https://doi.org/10.1021/bi800572s
  12. Gilson, M. K., & Zhou, H.-X. (2007). Calculation of protein-ligand binding affinities. Annual Review of Biophysics and Biomolecular Structure, 36, 21–42. https://doi.org/10.1146/annurev.biophys.36.040306.132550
  13. IUPAC-IUB Joint Commission on Biochemical Nomenclature. (1989). The nomenclature of steroids. European Journal of Biochemistry, 186(3). https://doi.org/10.1111/j.1432-1033.1989.tb15228.x
  14. Jumper, J., Evans, R., Pritzel, A., Green, T., Figurnov, M., Ronneberger, O., Tunyasuvunakool, K., Bates, R., Žídek, A., Potapenko, A., Bridgland, A., Meyer, C., Kohl, S. A. A., Ballard, A. J., Cowie, A., Romera-Paredes, B., Nikolov, S., Jain, R., Adler, J., . . . Hassabis, D. (2021). Highly accurate protein structure prediction with AlphaFold. Nature, 596(7873), 583–589. https://doi.org/10.1038/s41586-021-03819-2
  15. Kellogg, E. H., Leaver-Fay, A., & Baker, D. (2011). Role of conformational sampling in computing mutation-induced changes in protein structure and stability. Proteins: Structure, Function and Bioinformatics, 79(3), 830–838. https://doi.org/10.1002/prot.22921
  16. Kim, S., Chen, J., Cheng, T., Gindulyte, A., He, J., He, S., Li, Q., Shoemaker, B. A., Thiessen, P. A., Yu, B., Zaslavsky, L., Zhang, J., & Bolton, E. E. (2025). PubChem 2025 update. Nucleic Acids Research, 53(D1), D1516–D1525. https://doi.org/10.1093/nar/gkae1059
  17. Klumpp, S., Zhang, Z., & Hwa, T. (2009). Growth rate-dependent global effects on gene expression in bacteria. Cell, 139(7), 1366–1375. https://doi.org/10.1016/j.cell.2009.12.001
  18. Kraikivski, P. (2023). Mathematical modeling in systems biology. Entropy, 25(10), Article 1380. https://doi.org/10.3390/e25101380
  19. Le Guilloux, V., Schmidtke, P., & Tuffery, P. (2009). Fpocket: An open source platform for ligand pocket detection. BMC Bioinformatics, 10, Article 168. https://doi.org/10.1186/1471-2105-10-168
  20. Lutz, S. (2010). Beyond directed evolution: Semi-rational protein engineering and design. Current Opinion in Biotechnology, 21(6), 734–743. https://doi.org/10.1016/j.copbio.2010.08.011
  21. Marchler-Bauer, A., Bo, Y., Han, L., He, J., Lanczycki, C. J., Lu, S., Chitsaz, F., Derbyshire, M. K., Geer, R. C., Gonzales, N. R., Gwadz, M., Hurwitz, D. I., Lu, F., Marchler, G. H., Song, J. S., Thanki, N., Wang, Z., Yamashita, R. A., Zhang, D., . . . Bryant, S. H. (2017). CDD/SPARCLE: Functional classification of proteins via subfamily domain architectures. Nucleic Acids Research, 45(D1), D200–D209. https://doi.org/10.1093/nar/gkw1129
  22. Michaelis, L., Menten, M. L., Goody, R. S., & Johnson, K. A. (1913). Die Kinetik der Invertinwirkung/ The kinetics of invertase action. Biochemistry, 49(2).
  23. Morris, G. M., Ruth, H., Lindstrom, W., Sanner, M. F., Belew, R. K., Goodsell, D. S., & Olson, A. J. (2009). Software news and updates AutoDock4 and AutoDockTools4: Automated docking with selective receptor flexibility. Journal of Computational Chemistry, 30(16), 2785–2791. https://doi.org/10.1002/jcc.21256
  24. Saltelli, A., Ratto, M., Andres, T., Campolongo, F., Cariboni, J., Gatelli, D., Saisana, M., & Tarantola, S. (2008). Global sensitivity analysis: The primer. Wiley. https://doi.org/10.1002/9780470725184
  25. Soetaert, K., Petzoldt, T., & Setzer, R. W. (2010). Solving differential equations in R: Package deSolve. Journal of Statistical Software, 33(9), 1–25. https://doi.org/10.18637/jss.v033.i09
  26. Stokes, N. A., & Hylemon, P. B. (1985). Characterization of Δ4-3-ketosteroid-5β-reductase and 3β-hydroxysteroid dehydrogenase in cell extracts of Clostridium innocuum. Biochimica et Biophysica Acta (BBA)/Lipids and Lipid Metabolism, 836(2). https://doi.org/10.1016/0005-2760(85)90073-6
  27. Trott, O., & Olson, A. J. (2010). AutoDock Vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading. Journal of Computational Chemistry, 31(2), 455–461. https://doi.org/10.1002/jcc.21334
  28. Verhulst, P.-F. (1838). Notice sur la loi que la population suit dans son accroissement. Correspondance Mathématique et Physique Publiée par A. Quetelet, 10.
  29. Wilhelmová, N. (1996). Cornish-Bowden, A.:Fundamentals of Enzyme Kinetics. Biologia Plantarum, 38(3). https://doi.org/10.1007/bf02896674