AI-driven protein prediction is ushering in an era of precise drug design

AI-Driven Protein Prediction: Ushering in an Era of Precision Drug Design

Proteins are molecular machines that carry out most of the functions in our cells. Their core feature is that their three-dimensional structure, folded from sequences of amino acids, largely determines their function—what reactions a protein can catalyze, which molecules it can bind to, or what signals it can transmit is mainly dictated by its spatial structure. So if we want to understand disease mechanisms, design drugs that precisely target specific proteins, or engineer enzymes to withstand higher temperatures, we have to tackle a fundamental question: what does this protein's structure look like, and how does it change shape?

However, experimentally figuring out protein structures isn’t easy. Techniques like X-ray crystallography and cryo-EM are precise, but determining a single protein’s structure can take months to years and cost millions. X-ray diffraction requires high-quality crystals, which many proteins (like membrane proteins) struggle to form; cryo-EM doesn’t need crystals, but it requires highly stable samples, and the preparation process can easily denature proteins. Both traditional methods have clear limitations. This problem has puzzled biologists for decades and has remained a tough technical bottleneck. By the end of 2020, databases like UniProt catalogued over 190 million known protein sequences, but fewer than 180,000 protein structures had been determined experimentally, showing a huge gap in numbers. For this reason, scientists have been exploring ways to skip the lengthy experiments and directly predict 3D structures from amino acid sequences—this is one of the core goals of protein prediction.

3D Protein Structure and Bottlenecks of Traditional Structure Determination

How does AI predict proteins?

AI protein prediction models take the target amino acid sequence as the main input. The first step is to build a multiple sequence alignment (MSA) through sequence searching, capturing co-evolutionary constraints between residues from homologous sequences in different species. Simply put, over the long evolutionary processes of species, amino acid residues that are close to each other in 3D space tend to undergo adaptive mutations together, maintaining the overall stability of the protein structure. This co-evolution signal observed in homologous sequences across species can indirectly reflect spatial proximity between residues. On this basis, AlphaFold2 uses the Evoformer deep neural network module, iteratively reasoning through the attention mechanism between the MSA evolutionary information and residue pair spatial representations, gradually predicting distances and relative orientations for all residue pairs in 3D space, and eventually generating highly accurate 3D structural models. In CASP14 (the international protein structure prediction competition), the model achieved a median GDT_TS (Global Distance Test Total Score) of 92.4 out of 100 across all targets, and 87.0 in the most challenging free modeling category. Trained on massive experimental structure data from databases like PDB, the model gradually learned the complex rules for mapping 1D amino acid sequences to 3D atomic coordinates, and the structure prediction for a single protein can now be completed in just a few minutes.

By the end of 2025, the AlphaFold database had included over 241 million predicted structures, covering most known protein sequences (in contrast, the experimentally resolved protein structure database PDB contains only about 247,000 structures). Currently, millions of researchers in over 190 countries can access this data for free. This large-scale open-source movement is regarded as another milestone in structural biology, following the Human Genome Project.

Of course, AlphaFold also has clear limitations. In multi-domain proteins, those self-inhibitory proteins whose functions require relative motion between different domains often have experimental structures that are difficult for AlphaFold to accurately reproduce. Additionally, predictions for intrinsically disordered regions (IDRs) in proteins show significant deviations. Overall, AlphaFold tends to predict static, single conformations, whereas protein function often relies on dynamic conformational changes. Current methods still face challenges in capturing this conformational diversity. Therefore, even though prediction accuracy is nearly at the atomic level, the resulting 3D models still need experimental validation for practical applications. Nevertheless, it undeniably shortens the time from sequence to structure and provides 3D models for many proteins that were previously difficult to resolve experimentally.

From Sequence to Structure

From predicting structures to predicting functions, and even designing proteins

AlphaFold mainly tackles the problem of "sequence to structure," but protein prediction goes far beyond that. Another key area is predicting function from structure or sequence—for example, whether a protein can bind a specific ligand, has catalytic activity, or how its function changes after a mutation.

Protein language models are a prime example. Trained on massive amounts of natural protein sequences, these models learn the "evolutionary rules" of amino acid sequences. For instance, models like ESM-1v can predict the impact of single-point mutations on protein function, helping researchers quickly identify candidate mutations that could improve thermal stability or binding affinity, greatly cutting down on trial-and-error experiments. Similarly, for enzyme design, AI can start from sequences to predict substrate affinity or catalytic efficiency, guiding directed evolution experiments. These "sequence-to-function" prediction methods significantly boost the efficiency of functional analysis and mutation screening.

Going even further, protein prediction can work in reverse: generating sequences from desired functions. Given a target function (like "specifically bind a certain target"), AI can try to generate protein sequences that fulfill that function, opening up new possibilities for "de novo" protein design.

Xiaowu Intelligent Agent: Creating an Integrated R&D Loop of Prediction—Design—Validation

As protein prediction technology has advanced to the stage of being able to generate sequences in reverse, there’s also a new demand in the industry: can scattered technical tools be integrated to connect the full process of prediction, design, and validation, forming an automated R&D loop? Matwings Technology’s MatwingsVenus™ (Xiaowu™) intelligent agent is built exactly along this idea as a conversational protein R&D platform.

It’s not just a single structure prediction tool, but an intelligent R&D assistant that understands natural language instructions. Users only need to describe their task goals in natural language, and MatwingsVenus™ (Xiaowu™) will automatically break down the task, call on over 200 tools for structure prediction, function evaluation, developability analysis, and more, and connect the computation results to an automated wet lab platform, forming an iterative loop of “AI design—experimental validation—data feedback—redesign.”

Unlike general prediction tools, MatwingsVenus™ (Xiaowu™) has focused on real pharmaceutical R&D needs from the start, with specialized training for mainstream drug targets like GPCRs, kinases, and antibodies. In core drug R&D scenarios such as ligand-protein binding and antibody-antigen interactions, its specialized models outperform general baseline models in key metrics like binding affinity prediction and epitope recognition, significantly shortening the early discovery cycle for antibodies. As a domestically developed platform, it also offers enterprise-level services, ensuring the security of R&D data through fine-grained permission management and full-process audit tracking. Currently, MatwingsVenus™ (Xiaowu™) has completed full-cycle projects from de novo protein design to experimental validation in multiple cases, pushing drug R&D from “experimental trial-and-error” to “precision design.”

Ongoing Challenges

Even though protein prediction technology has made several groundbreaking advances, there's still a significant gap before we can fully understand the workings of life molecules. Many core problems still need joint efforts from both academia and industry to solve.

Predicting dynamic conformations is currently one of the most prominent challenges. Most mainstream methods today can only output a single static structure of a protein, but a protein's function often relies on changes in its dynamic conformation. It's like trying to capture a clear video of a running object—much harder than taking a single still photo. Coordinated folding of multi-protein complexes is even more complicated; it’s like predicting the overall tactics of a whole team rather than just analyzing the actions of individual players, and the complexity grows exponentially. According to industry stats, membrane proteins are the target of over 60% of small-molecule drugs, but predicting their structures is tricky because of the complex environment of the phospholipid bilayer in cell membranes, so current prediction accuracy is generally lower than that for soluble proteins.

On top of that, we still haven’t fully mapped structure to function accurately. Many proteins have similar structures but very different functions. At the same time, many high-performing AI models are like "black boxes"—they can give predictions, but the biological principles behind how they make decisions are hard for researchers to interpret. This lack of "explainability" also limits how deeply the technology can be applied in high-safety fields like drug development.

Outlook

The Era of AI-Driven Drug Discovery

From AlphaFold opening the door to structure prediction, to protein language models enabling function prediction, and then to industry platforms like MatwingsVenus™ (XiaoWu™) exploring integrated design-to-validation, protein prediction technology is steadily evolving along the path of 'see it—understand it—make it.' The industry generally believes that this technology could cut early-stage drug development cycles by nearly half, significantly reduce R&D costs, and allow more small and mid-sized research teams to have molecular design capabilities that were once only accessible to big pharmaceutical companies and top institutions.

As AI increasingly deciphers the 'origami code' of life, the goal of designing innovative molecules on demand and precisely intervening in diseases is gradually moving from concept to reality.