Protein Language Model Function Prediction: How to Choose? How to Implement?

Guide: Protein Language Models (PLMs) are evolving from 'structure prediction tools' into 'function prediction engines.' From gene ontology annotation to mutation effect prediction, from enzyme activity forecasting to aptamer design, PLMs are reshaping the research paradigm of computational biology. This article systematically reviews the technical paths, selection methods, and core challenges of PLM function prediction, and introduces how the Shanghai Tianwu Technology MatwingsVenus™ (Xiaowu™) platform integrates function prediction with wet-lab and in silico feedback loops, providing practical reference methods for AI4S, protein engineering, and computational biology researchers.

1. Three core paths of PLM function prediction

 Three Technical Paths for PLM Function Prediction

The current technical approaches for protein language model function prediction can be summarized into three main categories, each with its own pros and cons depending on the task dimension:

1.1 Zero-Shot Prediction: Likelihood-Based Evolutionary Inference

No additional training data is needed. You directly use a pre-trained model to estimate the effect of a mutation by comparing the log-likelihood ratio of the mutated sequence versus the wild-type sequence. The theoretical assumption is that residues frequently conserved through evolution are more likely to retain the protein's natural function (i.e., under purifying selection), so the likelihood ratio can serve as a proxy for destructive mutations.

Use case: Orphan protein families with little experimental data, quickly starting new projects.

Limitations: Limited accuracy for highly variable regions and higher-order combinatorial mutations.

1.2 Fine-Tuning Prediction: Representation-Supervised Transfer Learning

Use PLMs (protein language models) as feature extractors, feeding the sequence embeddings from pre-training into downstream supervised models, and fine-tune them on task-specific datasets. Typical tasks include GO function annotation, enzyme EC number prediction, protein-protein interaction prediction, ligand-binding site prediction, thermostability/solubility prediction, etc.

For tasks with hundreds or more labeled data points, fine-tuned models usually outperform zero-shot methods. Recent results from the CAFA (Critical Assessment of Function Annotation) challenge show that PLM-based methods have consistently ranked among the top.

Use case: Specific functional tasks with some level of labelled data.

Limitations: Performance is capped by the quality of data; tasks with sparse annotation are limited in effectiveness.

1.3 Fitness Landscape Modeling: From Single to Combinatorial Mutations

The most active direction is complete modeling of a protein’s fitness landscape—not just predicting single-point mutation effects but also predicting the combined effects of multiple mutations, directly serving mutant design in directed evolution.

The core challenge is epistasis: nonlinear interactions between mutations make single-point effects non-linearly extrapolatable. Recently, models trained on deep mutational scanning (DMS) data, combined with evolutionary priors from PLMs, continue to break accuracy records on benchmarks like ProteinGym. Top-tier studies have shown that PLMs can model long-range evolutionary trajectories.

Use case: Directed evolution, multi-objective optimization of industrial enzymes/antibodies.

Limitations: Depends on high-quality DMS data; extrapolating to distant sequences remains challenging.

2. How to Choose: The Four-Dimensional Evaluation Method for Function Prediction Models

When picking a PLM function prediction tool, don’t just rely on benchmark numbers from papers. It’s recommended to evaluate comprehensively from four dimensions:

l Benchmark accuracy. Look at mainstream benchmarks like ProteinGym, CAFA, and VenusMutHub, but be careful to distinguish between zero-shot settings and fine-tuning settings—they shouldn't be compared directly.

l Extrapolation ability. Check how the model performs on low-homology proteins and distant mutants—this reflects real-world value better than interpolation in conserved regions, which is exactly the most common scenario in industrial mining applications.

l Interpretability. Can the model point out key residues responsible for a prediction, and does it align with known functional regions? Black-box predictions have limited usability in research, and interpretability is key for building trust.

l Experimental validation rate. The ultimate gold standard is wet-lab validation. Among the Top N mutants recommended by the model, what proportion actually shows functional gains? This is the most critical metric for industrial deployment.

In short, for choosing a model: basic research should focus on benchmark ranking and interpretability, while industrial projects should prioritize experimental validation rate and closed-loop capability.

3. Why Function Prediction Is Harder than Structure Prediction

The success of structure prediction can be misleading—it might make people think protein function prediction is similarly straightforward. In reality, PLM-based function prediction is still an open problem facing three fundamental challenges:

Functional ambiguity. A protein's 'function' is not a scalar; it’s a multi-dimensional vector: the same enzyme can have catalytic activity, substrate specificity, optimal pH, thermal stability, expression levels, and dozens of other measurable attributes. The self-supervised signal in PLMs comes from sequence co-evolution, mixing constraints from folding, catalysis, binding, expression, etc. It doesn’t distinguish these dimensions, while 'function' is a multi-dimensional, highly context-dependent label set. This makes decoupling specific functional dimensions from the same embedding harder than predicting 3D structure.

Scarcity and bias in annotated data. Unlike the hundreds of thousands of experimental structures in the PDB, functional data is fragmented and unevenly distributed. Popular protein families have lots of DMS data, but many families have sparse or even zero annotations. Models may perform well on common families but drop sharply when generalized to distant proteins.

Extrapolation risk in benchmarks. Benchmarks like ProteinGym have advanced the field but also bring new challenges: can improvements on benchmarks translate into success in real projects? Studies show that some methods that do well on benchmarks see significant performance drops when extrapolated to distant regions of sequence space. For industrial applications, extrapolation ability is often more important than interpolation accuracy.

4. MatwingsVenus™ (Xiaowu™): Function Prediction-Driven Engineering Closed Loop

 MatwingsVenus™（晓鹜™）

Shanghai Matwings Technology's independently developed MatwingsVenus™ (XiaoWu™) protein R&D intelligent agent has taken protein language model function prediction from research benchmarks to industrial application. The following data all come from the platform's publicly published academic results and project disclosures.

4.1 Hierarchical Model System: From General Representation to Vertical Tasks

The platform's function prediction capability is built on a hierarchical architecture: the foundation is a Transformer-based general protein large model, pretrained on ultra-large-scale sequence data; on this basis, vertical models are developed for different functional tasks, covering core scenarios such as mutation effect prediction, stability landscape modeling, and functional enzyme mining.

Mutation Effect Prediction. The platform's mutation effect prediction model integrates protein sequence, 3D structure, and homology evolution information to predict and rank the functional impact of candidate mutations. In April 2025, the model achieved first place on the ProteinGym Substitution benchmark. In a 2026 international academic evaluation, using the ProteinGym v1.3 dataset to conduct a unified comparison of over a hundred models, the model achieved Spearman ρ=0.523 under the ranking of N=50 candidate mutations, ranking first among the professional prediction methods compared. The model has already completed wet lab validation in multiple experimental systems including VHH antibodies and DNA polymerases.

Fine-Grained Functional Understanding. VenusX, accepted by ICLR in 2026, focuses on detailed functional analysis at the substructure level of proteins.

Stability Prediction. The stability landscape learning model (Venus-MAXWELL), published in NeurIPS 2025, addresses the efficiency issues of learning mutation stability landscapes in small-sample scenarios.

4.2 Dry-Wet Closed Loop: From "Accurate Prediction" to "Actionable Mutation"

The ultimate value of function prediction is not in numerical accuracy but in whether it can guide mutations that truly improve performance. What sets MatwingsVenus™ (XiaoWu™) apart is integrating prediction capabilities into a complete dry-wet closed loop:

Researchers propose optimization goals through natural language (e.g., "increase enzyme activity 3-fold while maintaining thermal stability"), and the intelligent agent automatically schedules the function prediction models for mutation hotspot identification, combinatorial effect evaluation, and multi-objective optimization, outputting a concise list of candidate mutations; the design results are automatically imported into the experimental orchestration system, where the automated platform completes expression, purification, and functional characterization; experimental data is fed back to the model to start the next iteration.

This "predict-design-validate-iterate" closed loop turns PLM function prediction from an independent computational task into a core engine driving directed evolution research workflows.

Researchers can submit target sequences on the MatwingsVenus™ (XiaoWu™) platform to get preliminary function predictions and mutation optimization plans.

4.3 Practical Validation: Two Peer-Reviewed Typical Cases

The industrial value of function prediction has been confirmed in multiple projects through peer review:

- Cyclodextrin Glucosyltransferase Multi-Objective Directed Evolution. To tackle the synthetic challenge of EPS-G7, a key raw material for pancreatitis diagnosis, the team used PLM function prediction to perform multi-objective optimization of cyclodextrin glucosyltransferase—simultaneously increasing transglycosylation activity, reducing hydrolysis activity, and enhancing regioselectivity. In just one round of model prediction, they obtained the optimal triple-combination mutant. The ratio of the synthesized product increased from 63% to 98%, the transglycosylation/hydrolysis ratio improved 12-fold, and at a 100 L pilot scale, the yield reached 161 g/L with purity above 99%. The results were published in *Bioresource Technology*, with Tianwu Technology as a co-participant.

- PET Hydrolase Function Mining. Using the function mining model, the team directly screened billions of sequences for PET hydrolase candidates with high activity and stability, bypassing the traditional long path of "screen first then optimize." The research results were published in *Nature Communications*.

5. Practical Guide for Researchers: The Four-Step PLM Function Prediction Method

For teams conducting protein engineering and computational biology research, here are four actionable recommendations:

Step 1: Start from zero-shot. For new projects, use zero-shot PLM predictions for preliminary site ranking—it’s the lowest cost and fastest way to start; as experimental data accumulate, gradually move to fine-tuning models and fitness landscape modeling.

Step 2: Begin with small-scale validation. Any PLM function prediction model should first be validated on 10–20 mutants of the target protein. Once prediction accuracy meets expectations, scale up investment.

Step 3: Use multi-objective closed-loop solutions. Single-point predictions are easy, but balancing multiple objectives is hard. When dealing with activity, stability, expression, and other multiple targets, prioritize platforms that allow joint optimization and provide experimental verification loops.

Step 4: Navigate carefully to avoid wrong answers. The core value of PLM function prediction is narrowing the search space. The most effective approach is “PLM prediction → targeted experimental validation → iterative feedback,” forming a positive cycle.

6. Outlook: From ‘Predicting Function’ to ‘Designing Function’

Protein language model function prediction is experiencing a paradigm shift from “assisting annotation” to “active design.” In the next 2–3 years, advances in structure–function joint modeling, breakthroughs in low- and zero-shot capabilities, and deep integration of experimental feedback loops are expected to reshape the field.

Practical application on the MatwingsVenus™ (Xiaowu™) platform demonstrates that when ultra-large-scale protein data, function prediction models, and automated experiments come together, PLM function prediction is no longer just benchmark numbers in papers—it becomes a core engine that truly accelerates proteins from discovery to practical application.