O2DENet: Robust Enzyme Kinetics Prediction

• O2DENet is a plug-and-play module for enzyme–substrate interaction predictors that enhances out-of-distribution generalization in kinetic parameter prediction.
• It applies biologically informed perturbations and auxiliary consistency regularization to robustly process enzyme sequences and substrate structures.
• Empirical evaluations across multiple baselines show state-of-the-art gains in R², MAE, and AU-GOOD metrics, confirming its efficacy in enzymatic kinetics modeling.

O$^2$DENet is a plug-and-play module for enzyme–substrate interaction (ESI) predictors, designed to enhance out-of-distribution (OOD) generalization in enzymatic kinetic parameter prediction. It operates by introducing biologically and chemically informed perturbations during training and enforcing invariant representation learning through auxiliary consistency regularization. Notably, O$^2$DENet is architecture-agnostic: it does not alter existing predictor backbones or regression heads, enabling straightforward deployment to diverse enzyme kinetics modeling workflows. Empirical evaluation across four representative baselines demonstrates that O$^2$DENet yields state-of-the-art improvements in accuracy and robustness over stringent sequence-identity-based OOD benchmarks for $k_{cat}$ and $K_m$ prediction (Wu et al., 12 Jan 2026).

1. Architectural Principles

O$^2$DENet functions as a wrapper module for existing ESI predictors. The canonical workflow comprises a parallel dual-branch system:

• Raw branch: Processes unperturbed enzyme sequence $E$ and substrate structure $S$ through baseline encoders (e.g., 1D-CNN, Transformer, protein LLM [PLM], SMILES Transformer, or graph neural network [GNN]), producing enzyme embedding $f_E$ and substrate embedding $f_S$. The concatenated embedding $^2$0 is fed into the predictor’s regression head to predict kinetic parameters ($^2$1 for $^2$2 or $^2$3).
• Augmented branch (pseudo-data): Applies masking and randomization to $^2$4 and $^2$5 to generate perturbed variants $^2$6 and $^2$7 via MaskSeq$^2$8 and either SMILESEnum$^2$9 or MaskGraph$^2$0, respectively. These are encoded identically to yield augmented embeddings $^2$1, $^2$2, and concatenated $^2$3.

During training, predictions are made only from $^2$4, but both $^2$5 and $^2$6 inform the auxiliary consistency loss. At inference, only raw inputs are used; O$^2$7DENet’s augmentation machinery is inactive.

2. Pseudodata-Guided Perturbation Augmentation

O$^2$8DENet introduces domain-informed perturbations to both enzyme and substrate representations to create pseudo-examples:

• Enzyme sequence masking: Random positions (fraction $^2$9) in sequence $k_{cat}$0 are replaced by [MASK] tokens, producing $k_{cat}$1. Empirical ablation indicates an optimal $k_{cat}$2.
• Substrate augmentation:
  • SMILES enumeration: Generates alternative, valid SMILES strings $k_{cat}$3 via random atom traversals, preserving molecular connectivity but varying token order.
  • Graph masking: For graph-based substrates $k_{cat}$4, a fraction $k_{cat}$5 of non-core atoms/bonds are masked by replacing features with a learnable MASK embedding ($k_{cat}$6). Optimal $k_{cat}$7.

Encoders process both $k_{cat}$8 and $k_{cat}$9 identically, governed by parameter set $K_m$0: $K_m$1, $K_m$2.

3. Invariant Representation Learning via Consistency Regularization

To foster robustness against distributional shifts, O$K_m$3DENet employs L$K_m$4 consistency regularization:

• Consistency loss:

$K_m$5

• Base prediction loss: For regression of kinetic parameters, mean squared error is applied:

$K_m$6

• Total training objective:

$K_m$7

Optimal hyperparameter settings are $K_m$8, $K_m$9, as determined by ablation studies.

4. Training Protocols and Implementation

O$^2$0DENet has been benchmarked on CatPred-$^2$1 and CatPred-$^2$2, constructed from BRENDA via CatPred-DB, comprising UniProt-standardized enzyme sequences, corresponding 3D structures, substrate SMILES, and experimental $^2$3/$^2$4. Training and testing splits are explicitly constructed by mmseqs2 clustering at 99%, 80%, 60%, and 40% maximum sequence identity, targeting stringent OOD regimes.

Pseudodata generation maintains a 1:1 ratio with original data, and exact baseline encoder architectures are preserved (DLKCat’s CNN, UniKP’s PLM, CatPred’s SMILES Transformer+GNN, OmniESI’s hybrid). Optimization utilizes Adam (β₁=0.9, β₂=0.999), learning rate 1e-4, batch size 32, and early stopping on validation $^2$5. Deployment uses PyTorch 1.13, Ubuntu 20.04, Intel Xeon CPUs, NVIDIA RTX 4090 GPU, 120 GB RAM.

5. Quantitative Performance and Benchmarks

O$^2$6DENet demonstrates marked improvements in OOD generalization across multiple baselines and stratified sequence-identity splits. Key metrics are $^2$7, mean absolute error (MAE), and AU-GOOD (integration over identity thresholds):

Baseline	Split	$^2$8 Gain	MAE Gain	AU-GOOD Gain
DLKCat	99%	+79.5%	–	+59.4%
UniKP	99–40%	+9.7–20%	–	–
CatPred	99–80%	+0.5–17%	–	–
OmniESI	99–60%	+16.1–34%	−3.8–18%	+6.9%

For $^2$9 prediction, OmniESI achieves $E$0 increase from 0.541→0.576 (+6.5%, 99%), with MAE decreasing from 0.639→0.606 (−5.2%, 99%). Improvements are consistently observed across multiple OOD levels (down to 40%), with ablations establishing maximal benefit at mask ratios $E$1; increases above this threshold degrade biochemical signal and performance. AU-GOOD curves further confirm gains (+1–6% $E$2-based, +0.5–5% MAE-based) (Wu et al., 12 Jan 2026).

6. Integration Strategies and Methodological Considerations

O$E$3DENet integration involves retaining existing encoders and regression heads, augmenting enzyme and substrate inputs on-the-fly via residue masking and SMILES or graph augmentation (10% mask ratio), and computing both the standard regression and auxiliary consistency loss per mini-batch. At deployment, model inference operates unchanged.

Hyperparameters for tuning include $E$4, $E$5 (recommended 5–15%), $E$6 (0.1–1), learning rate, and batch size. The methodology presumes encoder compatibility with masking operations; embedding layers must be adapted if not. Excessive masking (>15%) is contraindicated due to information loss. For substrates with atypical chemistry (e.g., metals, cofactors), SMILES enumeration may be insufficient, necessitating graph-based masking. O$E$7DENet does not account for explicit 3D geometry or pH/temperature conditions and should be supplemented with domain-specific augmentations for such cases. Gains are maximal under stringent OOD test settings; for distributions closely matching training, observable improvement may be modest.

7. Contextual Significance

O$E$8DENet provides a generalizable, lightweight strategy for OOD generalization in enzyme kinetics prediction, emphasizing invariant representation learning and robust augmentation of training data using biochemical priors. Its architecture-agnostic design facilitates adoption without disrupting extant model pipelines, and empirical gains in R$E$9, MAE, and integrated metrics underscore its efficacy for diverse baseline architectures (Wu et al., 12 Jan 2026). A plausible implication is broader utility for other sequence-structure-function modeling tasks in protein engineering and computational enzymology, contingent on encoder compatibility and sufficient perturbation diversity.