AI Zero-Shot Prediction: How Can Protein Engineering Move Beyond Trial-and-Error Mutations?

Traditional protein engineering for modifying enzymes and functional proteins requires constructing tens of thousands of mutants for wet-lab experimentation, with R&D cycles spanning several years and prohibitive costs. AI zero-shot prediction technology brings a paradigm shift. Leveraging protein language models, inverse folding, and multimodal representation algorithms, it enables prediction of mutational effects without any target protein experimental data, substantially reducing experimental workload and shortening R&D timelines. With this technology, protein engineering cycles have been shortened from 2–5 years to 4–6 months, representing a core upgrade solution for industrial enzymes, gene-editing tools, and drug-binding protein development.

In traditional protein engineering, modifying a target enzyme or functional protein often requires constructing thousands to tens of thousands of mutants for individual activity screening, with the entire process consuming months or even years. This "experiment first, see results later" R&D paradigm relies entirely on large-scale screening and luck—researchers cannot predict which mutations will enhance performance in advance and can only depend on massive experimental trial-and-error.

AI zero-shot prediction has fundamentally changed this landscape: without requiring pre-existing mutational experimental data for the target protein, and relying solely on the AI's learned knowledge from vast quantities of natural protein sequences, it can accurately predict the functional consequences of mutations before any experiments are conducted, propelling protein engineering from "passive trial-and-error" into a new era of "active prediction."

I. What Is AI Zero-Shot Prediction? Predicting Protein Mutational Effects Without Experimental Data

Computational Light Probing Protein Active Site Pocket

Core definition of "zero-shot": When performing mutational prediction, the model has never encountered any functional experimental data for the target protein. It neither relies on target protein-specific mutant libraries nor on homologous protein activity assay results. Instead, it performs inference solely based on the evolutionary patterns of vast natural protein sequences learned during the pre-training phase.

Analogy: Traditional directed evolution is like aimlessly exploring an unfamiliar city on foot, where every route must be physically traversed to know the outcome. Zero-shot prediction is equivalent to having a complete city map in advance—even without having ever walked the target route, one can determine the optimal path based on the map's underlying patterns.

Protein Language Models (PLMs) are the core underlying tools for zero-shot prediction

These models draw from natural language processing techniques, performing self-supervised pre-training on hundreds of millions of natural protein sequences. They do not rely on experimental data to learn catalytic or binding mechanisms; instead, they autonomously discover the intrinsic relationships between amino acid sequences, protein three-dimensional structures, and biological functions.

The representative model ESM-2, after large-scale sequence pre-training, can predict the functional impact of single-point mutations from protein sequence input alone. Related studies have confirmed that the 96 high-quality mutant variants output by ESM-2 zero-shot prediction can serve directly as initial candidates for automated protein evolution platforms, substantially reducing early-stage screening workload.

Currently, zero-shot prediction encompasses three major established technical routes, each suited to different protein engineering scenarios:

Protein Language Model Direct Prediction

This approach employs masked language modeling (MLM) for training, calculating the fitness probability of substituting each sequence position with alternative amino acids. Higher probabilities indicate that the mutation conforms to natural protein evolutionary patterns and is more likely to maintain or even enhance catalytic or binding functions. This is the preferred approach for engineering proteins without known structures.

Inverse Folding Model with Constrained Optimization (AiCE Framework)

Unlike language models that start from sequences, inverse folding algorithms begin with protein three-dimensional structures and work backward to derive amino acid sequences that can stably fold into the target structure. AiCE, built upon general-purpose inverse folding models such as ESM-IF1 and ProteinMPNN, integrates structural and evolutionary constraints to batch-screen highly fit single and combinatorial mutations. It requires no per-protein model training, enabling general-purpose zero-shot evolution, and is well-suited for engineering enzymes and nucleic-acid tools with resolved structures.

Multimodal Deep Representation Learning (ProMEP)

ProMEP simultaneously integrates sequence and structure information from approximately 160 million proteins, capturing both evolutionary signatures and physicochemical constraints. It performs mutational prediction without requiring multiple sequence alignment (MSA-free), achieving superior predictive accuracy in gene-editing protein engineering scenarios (Cell Research 2024).

II. Three Core Advantages of AI Zero-Shot Prediction: Directly Addressing Traditional Protein Engineering Pain Points

Data Grid to Protein Crystal Iteration Pipeline

Traditional protein engineering has long suffered from three industry-wide bottlenecks: limited experimental throughput, the vastness of protein sequence space, and the difficulty of predicting combinatorial mutational effects. Zero-shot prediction effectively addresses these challenges:

1. Computational Pre-Screening of Massive Mutations, Retaining Only a Small Number of High-Potential Variants for Experimentation

Traditional directed evolution requires multiple cycles of library construction, protein expression, and activity screening, consuming enormous resources and time. Zero-shot prediction can evaluate tens of thousands of mutational combinations in a single batch, directly filtering out the best-performing candidates and substantially reducing wet-lab scale.

A 2025 Nature Communications study demonstrated this approach: researchers integrated ESM-2 zero-shot prediction with an automated biofoundry to establish a fully automated protein evolution platform. Using tRNA synthetase as a model enzyme, the AI first employed ESM-2 zero-shot prediction on 96 variants to initiate the evolution cycle. Experimental data were fed back to train a multilayer perceptron (MLP) prediction model. Four evolution rounds were completed within 10 days, achieving a 2.4-fold increase in enzyme activity. Achieving comparable results through traditional directed evolution would typically require months, representing a dramatic improvement in R&D efficiency.

2. Accurate Assessment of Combinatorial Mutations, Avoiding Local Optima Traps

Traditional engineering typically introduces only a small number of single-point mutations, making it highly susceptible to local optima while failing to identify globally optimal mutational combinations. Zero-shot prediction can simultaneously evaluate multi-site synergistic mutational effects.

The MODIFY integrated framework combines protein language models with sequence density algorithms to automatically eliminate deleterious mutations. Coupled with Pareto optimization for mutant library design, it screens for highly fit variants while maintaining mutational diversity—balancing performance enhancement with sequence breadth and overcoming the limitations of single-point engineering (Nature 2024).

3. Sequence-Only Analysis Enables Engineering Without 3D Structures

Traditional rational design relies heavily on high-resolution protein crystal structures, rendering numerous "orphan proteins" without resolved structures difficult to engineer. Protein language model-driven zero-shot prediction, however, requires only the amino acid sequence for analysis, substantially broadening the scope of engineerable proteins.

The ProMEP model has been validated in gene-editing enzyme engineering: for TnpB gene-editing enzyme engineering, a 5-site mutant achieved 74.04% editing efficiency compared to 24.66% for the wild type; for TadA base editor optimization, a 15-site mutant achieved 77.27% A-to-G conversion frequency, outperforming the industry-standard ABE8e tool while substantially reducing off-target and bystander mutation risks (Cell Research 2024).

III. Three Foundational Technological Breakthroughs Underpinning Zero-Shot Prediction for Industrial Translation

The successful industrial translation of the above three technical routes is enabled by the following key breakthroughs:

1. Protein Language Models: Autonomous Learning of Intrinsic Protein Evolution Patterns

Self-supervised pre-training enables models to "read" the language of proteins, autonomously recognizing the relationships between amino acid arrangements and function. These models not only capture local sequence features but also identify synergistic and antagonistic relationships among distant amino acid residues, accurately predicting the epistatic effects arising from multiple mutations—all without requiring experimental data to assess combinatorial mutational outcomes.

2. Inverse Folding + Structural Constraint Integration (AiCE Framework)

Built upon general-purpose inverse folding models such as ESM-IF1 and ProteinMPNN, AiCE learns the sequence–structure correspondence. When evaluating mutations, it simultaneously considers protein folding stability. The framework is applicable to diverse protein classes—including deaminases, nucleases, and reverse transcriptases—spanning sequences from dozens to thousands of amino acids in length. Beneficial mutation hit rates can reach substantial levels, though they vary across different systems (Cell 2025).

Multiple base editors optimized via AiCE have demonstrated comprehensively enhanced performance: enABE8e narrowed the editing window to 5 bp (approximately 50% reduction compared to wild-type ABE8e's 9 bp); enSdd6-CBE achieved 1.3-fold improved fidelity; and the mitochondrial editor enDdd1-DdCBE showed up to 14.3-fold enhanced activity (Cell 2025).

3. Sequence + Structure Multimodal Integration (ProMEP)

Both sequence-only and structure-only information have inherent limitations. ProMEP integrates both dimensions, enabling evolutionary information and spatial physicochemical constraints to complement each other. This synergy simultaneously improves mutational effect prediction accuracy and computational speed, establishing ProMEP as a core tool for high-performance gene-editing tool development.

IV. Real Industrial Value of Zero-Shot Prediction: Validated Through Enterprise Deployment

Trial-and-Error Fragments vs. AI-Guided Protein Crystal

Zero-shot prediction is no longer confined to laboratory theory—it has been fully deployed across industrial enzyme, gene-editing, and innovative therapeutic protein R&D, delivering three tangible industrial benefits:

1. Substantial Compression of R&D Timelines

Traditional complete enzyme and functional protein engineering cycles span 2–5 years. With AI zero-shot prediction performing pre-screening, the overall R&D cycle is shortened to 4–6 months.

Matwings Technology, a domestic protein design company, has developed its proprietary MatwingsVenus™ （晓鹜™） platform, pre-trained on billions of natural protein sequences. The platform incorporates mature zero-shot prediction algorithms and establishes a self-closed-loop AI design system spanning from requirement definition to high-quality sequence output. According to Matwings Technology's official website, the platform leverages the powerful generalization capability of its large model to accurately identify high-performance candidate molecules without experimental data, achieving dramatic R&D cycle compression.

2. Substantial Reduction in Wet-Lab Costs

Traditional R&D requires validation of tens of thousands of mutational samples one by one. With AI zero-shot prediction performing pre-screening, only 100 or fewer high-potential candidates need to be experimentally validated—significantly reducing consumable and labor costs associated with strain construction, protein purification, and activity assays.

3. Expanded Boundaries of Protein Engineering

For orphan proteins lacking resolved structures and homologous sequences—which are difficult to engineer using traditional methods—zero-shot prediction enables mutational prediction based solely on primary sequence, filling a gap left by conventional approaches.

4. Enabling Novel Drug and Biomaterial Development

A June 2026 Nature study demonstrated that researchers employing the NISE iterative strategy achieved de novo design of binding proteins targeting two small-molecule drugs—exatecan and apixaban—with success rates of 100% and 83%, respectively. The best binders achieved affinities at the nanomolar to picomolar level (Nature 2026).

After introducing two amino acid substitutions in the exatecan-binding protein, affinity was enhanced a further 100-fold while simultaneously stabilizing the drug's hydrolytically labile lactone ring, thereby extending the drug's active duration. This finding confirms that zero-shot prediction technology is applicable not only to industrial enzyme activity optimization but also to diverse scenarios including drug delivery, biosensing, and catalytic synthesis.

Enterprise Deployment Case Study: Matwings Technology AI Protein Engineering Services

According to Matwings Technology's official website and public reports, the MatwingsVenus™ （晓鹜™） platform incorporates mature zero-shot prediction algorithms and has delivered over 30 custom protein engineering projects spanning four domains: industrial enzyme preparations, in vitro diagnostics, innovative therapeutics, and gene-editing tools. Nearly 10 products have achieved commercial production, with multiple proprietary enzymes matching or exceeding the performance of comparable products from leading international manufacturers.

Leveraging the pre-screening advantages of zero-shot prediction, the platform enables simultaneous multi-objective co-optimization, balancing catalytic activity, thermostability, substrate specificity, acid–base tolerance, and other industrial requirements—providing a one-stop solution to the traditional challenges of prolonged timelines, low success rates, and high costs in protein engineering.

V. Emerging Industry Trends: Zero-Shot Prediction Reshaping the Foundational Paradigm of Protein Engineering

Looking across the evolution of the protein engineering industry, the technology has completed a critical transition from "natural screening and manual trial-and-error" to "AI zero-shot precision prediction," comprehensively covering the entire workflow of mutational prediction, de novo sequence design, and candidate library screening:

Prediction front: Deep learning models such as IECata and ESM-2 enable zero-shot prediction of core kinetic parameters, including kcat/Km, enabling rapid elimination of ineffective mutations (based on a 2025 study published in PMC).

Design front: Diffusion models such as RFdiffusion generate target backbones, while inverse folding models including ProteinMPNN and ESM-IF design amino acid sequences for those backbones, enabling directed optimization of catalytic sites and drug-binding pockets.

Screening front: Machine learning enables direct prediction of protein activity and substrate specificity from primary sequence, eliminating substantial preliminary experimental work.

AI zero-shot prediction technology deployment has progressed through three developmental phases: Phase I—serving as an assistive tool for structural resolution and mechanistic prediction; Phase II—enabling zero-shot generation of entirely new high-activity protein sequences; and the current phase—simultaneously balancing multiple metrics including activity, stability, and selectivity, providing a one-stop solution to complex industrial application constraints.

The traditional protein engineering logic was "experiment first, validate later"—experimentation was the sole screening mechanism. Zero-shot prediction establishes a fundamentally new logic: "compute first, validate later." AI performs the vast majority of screening work, with wet-lab experimentation reserved solely for performance confirmation of a small number of superior variants. The industry significance of this transformation is comparable to the upgrade from manual drafting to CAD computer modeling—converting massive repetitive experimental labor into AI-driven computational intelligence.

A 2025 Cell study on the AiCE framework explicitly articulated the ideal protein engineering approach: achieving optimal engineering outcomes with minimal experimental effort. Zero-shot prediction represents the optimal implementation pathway for this principle—substituting repetitive wet-lab experiments with knowledge learned from massive pre-training sequence data, and replacing empirical trial-and-error with model-based inference.

Looking ahead, as AI protein models continue to advance, protein catalytic activity, drug-binding affinity, editing efficiency, and other performance metrics will gradually become customizable, programmable, and standardized properties. Whether for industrial fermentation enzymes, clinical gene-editing therapeutics, or small-molecule targeted binding proteins, high-performance protein products can be rapidly customized through zero-shot AI prediction, effectively overcoming the capacity, cost, and performance bottlenecks inherent in traditional approaches and driving the large-scale, high-speed development of the synthetic biology and biomanufacturing industries.