Protein Sequence and Function Relationship: Three-Layer Mapping, Three Major Dilemmas, and New Solutions in the AI Era

From AlphaFold cracking the challenge of structure prediction, to AI-designed protein drugs entering clinical trials, and to AI-driven enzyme engineering in synthetic biology constantly breaking performance records—protein science is undergoing an unprecedented paradigm shift. At the core of all these technological breakthroughs lies the same fundamental scientific question: the relationship between protein sequences and their functions.

How important is this question? It’s the 'first principle' of protein engineering and the central proposition of bioinformatics—whether it’s boosting the activity of a single enzyme or designing a completely new antibody drug from scratch, essentially, it’s all trying to answer the same thing: why does a specific amino acid sequence correspond to a particular function? And how can it be changed to get the desired function?

But if you’ve actually done protein research, you know this question is far more complicated than it seems. Why is it that even though structure prediction has reached near-atomic precision, we still can’t directly read function from sequence? Why is it so hard to achieve a massive performance jump even after screening tens of thousands to hundreds of thousands of variants in directed evolution? Why can a seemingly conservative point mutation lead to a completely unexpected functional phenotype?

In today’s article, we’ll start from the most fundamental question—'the relationship between protein sequences and functions'—and break down its essence, the challenges, and the paradigm shift happening in the era of large AI models.

1. What is the 'relationship between protein sequences and functions'? It’s not just as simple as the central dogma.

Schematic diagram of the central dogma of molecular biology

The central dogma tells us the flow of information "DNA → RNA → protein," but it only tells us where proteins come from. The question of the "relationship between protein sequence and function" goes deeper: how does a linear chain made up of 20 amino acids ultimately determine the biological function it performs?

From the perspective of information flow, the sequence-function relationship involves at least three layers of mapping:

First layer: Sequence → Structure. This is the most widely recognized layer. Twenty amino acids arranged into a polypeptide chain fold into a 3D structure, which gives it biological function. Revolutionary breakthroughs in tools like AlphaFold and ESMFold have moved our understanding of this layer from "very hard to predict" to "mostly predictable." Structure is the most important bridge between sequence and function, but it’s not everything.

Second layer: Structure → Function. This is more subtle and also more difficult. Two proteins with very similar structures may have vastly different functions; conversely, two proteins with low sequence homology may catalyze the same reaction. Function is not just "can it bind the substrate," but also how strongly it binds, how fast it catalyzes, optimal pH, optimal temperature, stability, solubility… you can’t get all that from a static structure alone.

Third layer: Sequence → Function (end-to-end). This is the ultimate "holy grail": given an amino acid sequence, directly predict its core functional phenotypes—such as enzyme activity, affinity, stability, solubility, and other physicochemical properties governed by the sequence itself. No intermediate structure as a bridge—directly map the 1D sequence to key functional dimensions. This is also the ultimate goal in understanding the protein sequence-function relationship.

So when we talk about "studying the relationship between protein sequence and function," we’re not just looking at a single problem, but at an entire complex network mapping from sequence to the functional space layer by layer. And almost every node in that network is a black box.

2. Why is the relationship between protein sequences and functions so hard to understand? Three core dilemmas

If it were just a linear relationship of "sequence → structure → function," the problem would have been solved long ago. The relationship between real-world protein sequences and functions lies in three fundamental dilemmas.

Dilemma 1: The sequence space is too large, and the functional space is more dimensional

A protein of 300 lengths theoretically has 20^300 possible sequences—a number far exceeding the total number of atoms in the observable universe (about 10^80). Of course, the biological space for sequences that can be "folded into stable structures" is much smaller than this number, and even so, the scope that current experimental methods can explore is still a tiny fraction of this vast space. And what about the functional dimension? Catalytic activity, substrate specificity, optimal pH, optimal temperature, stability, solubility, immunogenicity...... Just counting casually is a dozen or so dimensions.

How low is the probability of finding a small handful of sequences that simultaneously satisfy multiple functional constraints in an almost infinite sequence space? The reason for creating databases and screening for directional evolution is essentially because we can't "calculate" it, only "collide" it.

Dilemma 2: Not one-to-one correspondence, but "many-to-one" and "one-to-many"

The relationship between sequences and functions is not simply a one-to-one mapping:

• Many-to-one: Many different sequences can be folded into similar structures and perform similar functions. Nature has long proven this—homologous proteins from different species may have significant sequence differences but highly conserved functions.

• One-to-many: The same sequence may function completely differently under different conditions, cellular environments, and post-translational modifications. A protein may be both an enzyme and a structural protein, possibly both a transcription factor and a pro-apoptotic factor.

This bidirectional non-one-to-one mapping makes the "sequence read function" extremely difficult.

Dilemma Three: Top-level Effect and Contextual Dependence

Schematic illustration of epistasis

The effects of single-point mutations are often not simply additive. A mutation at site A might double the activity, and a mutation at site B might also double it, but putting both A and B together may not give a 4-fold effect—it could be more, less, or even the opposite. This is called epistasis, which means that the functional effect of a mutation at one site is influenced by the sequence context of one or more other sites, and the combined effect deviates from a simple sum of individual effects. Epistasis can show up as positive synergy (1+1>2), negative antagonism (1+1<2), or even a reversal in sign (beneficial individually but harmful together).

What’s even trickier is that epistasis is highly context-dependent: the same pair of mutations may be beneficial in one protein scaffold but harmful in another. This explains why a rational design that works on protein A might not work on protein B—the mapping from sequence to function isn’t linear, it’s the result of cooperative effects across the entire sequence.

Because of epistasis, studying the relationship between protein sequence and function can’t be simplified to 'find a few key sites and tweak them'; you have to deal with the nonlinear mapping of the whole sequence space. This is also a key reason why protein engineering can’t yet be fully 'calculated' or predicted.

All three of these challenges together explain why, even after years of research, the relationship between protein sequence and function remains one of the most central open questions in all of life sciences.

3. Traditional Methods for Studying the Relationship Between Protein Sequence and Function: Where Are Their Limits?

Before the AI revolution, there were mainly three paths to study the relationship between protein sequences and their functions, each with its own limitations.

Path 1: Structural Biology — 'See what it looks like, then you know what it can do'

From X-ray crystallography to cryo-EM, structural biologists have been constantly pushing the limits of "seeing clearly." Knowing the structure does provide a lot of clues about function—like the location of active sites, the shape of substrate-binding pockets, and the distribution of potential functional residues.

But the limitations of structural biology are obvious: the throughput of experimentally solving structures is far lower than the speed at which sequences are generated; structures are static, while functions are dynamic; structural similarity does not equal functional similarity. Structure is a necessary condition for function, but not a sufficient one.

Path 2: Directed Evolution — 'Let nature do the screening for us'

Since we can’t design it ourselves, we let mutations and screening handle it. The idea of directed evolution is simple, crude, and effective: build a library of mutants, apply functional selection pressure, and after a few rounds, you can get mutants with improved performance.

There are countless success stories with directed evolution—from maturing antibody affinity to industrial enzyme engineering. But its ceiling is also clear: the size of the library you can screen is limited (usually 10^6 to 10^10), which is just a tiny fraction of the sequence space; the types of functions you can optimize are strictly limited by whether you can set up a high-throughput screening system; and even when you get mutants with improved function, you often have no idea why they improved—'you know it works, but not why.'

Directed Evolution Cycle

Directed Evolution Cycle

Pathway 3: Bioinformatics — 'Finding Patterns in Massive Data'

By comparing homologous sequences, analyzing co-evolution, and predicting conserved sites, bioinformaticians uncover links between sequences and functions from vast amounts of natural sequences. Co-evolution analysis is especially powerful — if two sites always mutate together during evolution, it suggests they might be functionally related.

But bioinformatics has its issues: it's good at finding 'correlations' but not 'causation'; it relies on existing data and struggles with new proteins that differ significantly from known families; it can tell you which sites are important but has a hard time telling you 'what changes would be better.'

Each of these three pathways has its own strengths, but none has fundamentally cracked the code linking protein sequences to functions. And the advent of AI large models is changing the rules of this game.

4. In the Era of AI Large Models, the Study of Protein Sequence-Function Relationships Has Been Completely Rewritten

From the breakthrough of AlphaFold2 in 2020 to now, in just a few years, AI has reshaped protein science, expanding from 'structure prediction' to 'function prediction' and 'sequence design.' For studying the relationship between protein sequences and functions, AI large models have brought three fundamental changes.

Change 1: From Statistical Correlation to High-Confidence Quantitative Prediction

Traditional bioinformatics tells you 'this site might be functionally relevant,' while AI large models, trained on massive datasets, can give quantitative mutation effect predictions for protein families with enough data — for example, 'if this site mutates to X, predicted activity will increase/decrease by Y%.' Accuracy varies depending on the protein family and data availability, but in more and more cases, it already provides reliable guidance for experiments.

In the study of protein sequence-function relationships, AI's current core value lies in 'prediction' rather than complete 'explanation' — it can provide sufficiently accurate functional predictions in many scenarios without relying on a full understanding of the mechanism. It's like how we don't need to fully understand why every weight works as it does in deep learning, yet it doesn't stop us from making precise inferences.

Change 2: From 'Single-Point Mutations' to 'Global Design'

Traditional rational design usually only dares to change a few sites, because changing too many can make things 'go out of control.' AI large models, on the other hand, can consider tens or even hundreds of sites simultaneously for synergistic mutations, finding paths to the desired function through a maze of epistatic effects.

This is crucial for understanding the relationship between protein sequences and functions. Functions are not determined by a single site but result from the cooperation of the entire sequence. Only when we can understand sequence-function mapping on a global level can we truly get to the core of the issue.

Change 3: From 'Experiment-Driven' to a 'Design-Validate-Iterate' Loop

Traditional protein engineering is 'experiment first, understand later'; with AI involvement, it becomes 'predict first, validate next, then iterate.' The role of experiments shifts from 'blind screening' to 'validation and feedback,' boosting efficiency by orders of magnitude.

In this wave, a group of AI agents for protein engineering is emerging. They’re no longer satisfied with just 'giving you a prediction' but aim to cover the entire chain from 'functional goals to optimized sequences.' In terms of the closed-loop capability for function prediction and sequence design, MatwingsVenus™ (Xiaowu™) is a typical representative in this field.

5. MatwingsVenus™ (Xiaowu™) Agent: An AI-Powered Full-Chain Tool for Protein Sequence and Function Research

In the study and application of protein sequence-function relationships, the MatwingsVenus™ (Xiaowu™) agent, with AI-driven protein design technology at its core, provides full-chain intelligent support—from functional site prediction and directed evolution design to global sequence optimization. Leveraging self-developed general protein design large models and a wet-dry experiment closed-loop system, MatwingsVenus™ (Xiaowu™) can efficiently predict key sites affecting protein function, simulate the effects of single-point and multi-point mutations on various functional phenotypes such as enzyme activity, affinity, and stability, and reverse-design optimized sequences according to target functional requirements. This upgrades the traditional 'trial-and-error screening' R&D model to 'precision-directed design,' drastically reducing experiment time and costs.

MatwingsVenus™

The MatwingsVenus™ (Xiaowu™) platform offers a whole bunch of standardized tools for protein sequence and function research. This includes protein sequence functional site prediction tools, modules for analyzing various properties like stability, solubility, and immunogenicity, an intelligent system for directed evolution, a massive database of labeled protein sequences, and over 200 functional components including protein structure prediction and analysis tools. Researchers can mix and match these modules on MatwingsVenus™ to quickly set up their own sequence-to-function analysis workflows.

The platform also provides end-to-end custom services, like directed evolution for specific target proteins (think boosting enzyme activity, improving antibody affinity, or increasing stability), designing new functional proteins from scratch, mining and validating rare functional sites, and full-process outsourcing from AI design to wet-lab experiments. Users just need to specify their functional goals, and the platform will provide experimentally validated optimized sequences—basically 'you tell us what function you need, we give you usable sequences.' This really takes protein sequence-function research from 'hard to grasp' to 'predictable, designable, and verifiable.'

6. How far are we from really "getting" the connection between protein sequences and functions?

Honestly, we still have a long way to go to truly "get" the relationship between protein sequences and what they do. AI models can give pretty accurate results, but a lot of the time, we don’t fully understand why they're accurate—it’s like AlphaFold predicting structures without directly showing the physical rules of protein folding. That said, it’s still a super powerful tool. For basic research, understanding the mechanisms is a long-term goal; but for practical applications, just being able to predict and design reliably is already a game changer.

Before AI, the sequence space felt completely dark, and we could only explore bit by bit with directed evolution as our "flashlight." Now, we have a "functional map"—maybe incomplete, but enough to guide us. And smart tools like MatwingsVenus™ (Xiaowu™) are like GPS devices holding that map, helping you move forward fast.

The relationship between protein sequences and functions—the oldest and most fundamental question in life science—is being pushed forward in a totally new way by AI. Maybe in ten years, we’ll look back and see that what we call "understanding" today was really just opening the first door. But opening that door is what’s turning protein engineering from a "craft" into a "design discipline"—and every one of us researchers is living this revolution firsthand.[eos]