Why Wet Lab Validation Remains Irreplaceable in the Age of AI Structure Prediction

AlphaFold's Transformative Impact

Enginoma Structure was trained on the PDB: a curated collection of experimentally determined X-ray crystallography, cryo-EM, and NMR structures. Its objective is to predict the physically stable, low-energy conformation that a protein adopts under the conditions of the experiments in PDB—the vast majority of which are near-physiological buffers at moderate temperatures. Enginoma Structure excels at this.

What it does not predict are functional properties: catalytic efficiency (k_cat/K_M), substrate specificity, thermal stability at process temperatures, tolerance for organic solvents, expression yield in a heterologous host, proteolytic susceptibility, or aggregation propensity under process conditions. These are the properties that matter for industrial enzyme engineering, and they depend on factors that go well beyond the static 3D coordinates of a folded structure.

The distinction matters in practical terms every day at the bench. A protein can adopt its predicted folded structure and still fail spectacularly as a biocatalyst—because the active site geometry is sensitive to local dynamics, water molecules, or substrate-induced conformational changes that static structure prediction cannot capture.

Structure ≠ Function: The Prediction Gap

A rigorous validation study examined Enginoma Structure predictions against high-quality experimental structural data across a curated benchmark set. The findings were sobering: approximately 20% of AI-predicted side-chain conformations showed significant deviations from experimental electron density, and roughly 7% of predictions were effectively incompatible with the underlying experimental data. These discrepancies were not random noise—they clustered in active site regions, loop dynamics, and allosteric networks precisely the zones most relevant to enzyme function.

For industrial biocatalysis, these error rates matter enormously. An active-site side chain positioned 2–3 Å off from its experimentally determined location can mean the difference between a catalytically competent geometry and one that barely supports turnover. A flexible loop that structure predictors mark as ordered may, in reality, be disordered under process conditions—causing expression failure or aggregation when the protein is produced at high titre.

Consensus Filtering: Traditional Metrics Fall Short, and So Does Structure Prediction Alone

Research into consensus multi-model scoring extended this critique to functional prediction. A central hypothesis was tested: are conventional metrics like sequence similarity (BLAST e-value) and BLOSUM62 scores the best available predictors of enzyme function? The answer was no—and neither, surprisingly, was structure prediction alone.

When Enginoma Structure structure predictions were benchmarked alongside sequence-based metrics for predicting functional similarity across enzyme families, structure prediction improved functional prediction modestly but was far from definitive. The optimal predictors combined structural information (Enginoma Structure), evolutionary context (Enginoma sequence models embeddings), and direct experimental activity measurements in a unified scoring framework. The experimental data acted as a regularizer, correcting cases where structural or sequence similarity was misleading—paralogs with nearly identical structures but divergent functions, convergent enzymes with different folds that catalyze the same reaction.

Specific Failure Modes of AI-Designed Proteins

At CD Biosynsis, we have characterized hundreds of computationally designed enzyme variants in our wet lab. The most common failure modes for AI-generated candidates are not exotic—they are mundane and instructive. Understanding these patterns is essential for designing effective hybrid AI-wet lab programs.

What Experimental Validation Actually Tests

There is no computational shortcut for actually measuring enzyme kinetics. k_cat, K_M, k_cat/K_M, IC₅₀ values, substrate specificity ratios, pH optima, temperature optima, solvent tolerance curves—these are empirical measurements that require real biochemistry. No current AI model predicts these values with the accuracy needed for regulatory submissions, process design, or intellectual property claims.

For pharmaceutical applications, regulatory agencies (FDA, EMA) require experimentally determined activity data for any enzyme used in a manufacturing process. A structure prediction—even a perfect one—is not sufficient for a Drug Master File. The experimental characterization must be done, and done rigorously, with appropriate controls and statistical rigor. This regulatory reality ensures that wet lab work will remain indispensable for commercial enzyme applications regardless of how powerful computational methods become.

The DBTL Cycle: Design-Build-Test-Learn

The most effective strategy we've found at CD Biosynsis is the Design–Build–Test–Learn (DBTL) cycle, applied with particular intensity at the Test and Learn stages. In practice, this means:

1. Design: AI models generate and prioritize variant candidates. Computational filters remove structurally implausible candidates and select for predicted stability or activity improvements.
2. Build: A curated set of prioritized variants (typically 20–100 per round, depending on throughput) are gene-synthesized, cloned, and expressed in the chosen host organism.
3. Test: Purified variants are characterized through a standardized assay panel: activity screening, kinetic parameter determination, thermal stability profiling, solubility assessment, and substrate specificity screens.
4. Learn: Experimental results are used to update surrogate models, refine activity predictors, and adjust the generative model priors for the next design cycle.

Critically, the experimental data is not merely used to validate or invalidate individual variants—it is used to improve the next round of design. Each cycle produces a better model, which produces better candidates, which produce better experimental results. This is where the compounding advantage of the AI-wet lab integration truly lies: not in any single round, but in the trajectory of successive rounds.

Over 3–4 DBTL cycles, programs at CD Biosynsis routinely achieve activity improvements of 5–50× relative to starting templates, with the rate of improvement accelerating as proprietary experimental data accumulates. This would be impractical with traditional directed evolution alone—the screening burden would be prohibitive. It would be impossible with AI design alone—the accuracy gap is too large without experimental feedback.

Building the Data Flywheel at CD Biosynsis

The convergence of these insights shapes our integrated approach: we use AI to generate and prioritize candidate variants with unprecedented efficiency, and we use rigorous wet lab validation to close the gap between computational prediction and functional reality. Every round of data improves the models; every improvement in the models reduces the experimental burden of the next round.

This data flywheel is our core competitive advantage. As we accumulate proprietary experimental datasets—thousands of kinetic measurements, stability profiles, and expression outcomes across diverse enzyme families—our AI models become increasingly accurate predictors of real-world functional performance. Clients who engage CD Biosynsis early in their enzyme programs benefit most from this compounding effect.