Protein De novo design using MatwingsVenus

If you've also had enough of complex toolchains and high barriers, this article might make you smile knowingly.

Let me start with a true story.

Last month, a friend of mine who works in biopharmaceuticals complained to me. He said that their lab wanted to work on a miniaturized binding protein for a relatively new immune target. The boss's idea was, "Don't you have AI? Quickly design a batch from scratch."

So what happened? They spent almost two months and were still adjusting RFdiffusion parameters.

It's not that they're not trying—PhDs with a structural biology background are certainly not dumb. The problem is: you need to run the entire toolchain of RFdiffusion, ProteinMPNN, AlphaFold3, and along the way, handle various format conversions, environment conflicts, parameter tuning… He said his daily situation now is that he spends more time writing code than thinking about protein backbones.

After hearing this, I felt quite emotional.

The field of de novo protein design has indeed advanced tremendously in terms of technology—the 2024 Nobel Prize in Chemistry was directly awarded to the pioneers of AI protein design and structure prediction, and David Baker's team updates knowledge every month in 《Nature》 and 《Science》. But for most ordinary R&D personnel, the fragmentation of the toolchain and the high technical threshold still remain an impassable wall.

Methods for de novo protein design

A less painful way to put it is:

All the good stuff is there, but using it is really exhausting.

1. Industry Consensus: AI is making "creating something from nothing" a daily occurrence

Of course, the big picture is exciting.

In the past, saying "design from scratch" carried a certain idealistic flair—wanting to create a functional protein from nothing without any template in nature? Sounds like showing off. But now it’s different: from GPCR micro agonists/antagonists to programmable protein cages for drug delivery, these AI-designed proteins can already be validated by cryo-EM and match the design models closely.

According to David Baker’s own judgment: in the next 5-10 years, the potential of protein design in medicine, technology, and sustainable development could far exceed that of natural proteins.

An industry consensus has formed: protein design is moving from "making molecules" to "making machines," and AI is making things that previously required extensive expert experience and luck predictable, iterative, and scalable.

But consensus is consensus; the real feeling for most people is—

2. The Practical Dilemma: There are plenty of tools, but researchers are not AI engineers

protein design

Let's go back to the classic workflow: RFdiffusion (build the backbone) → ProteinMPNN (fill in the sequence) → AlphaFold3 (validate the structure). The functionality is indeed powerful, but can someone tell me, why in 2026 are we still manually modifying input file formats? Still arguing over environment setup? Still unable to run the whole process because of a model version update? Fragmented toolchains, high entry barriers, and lack of validation loops—anyone who has personally done de novo design can probably complain about these three problems all day long. Some industry observers have bluntly pointed out the root of the problem: many so-called "one-click" platforms can actually only handle fixed input formats; slightly more complex backbone cooperative design makes computational costs ridiculously high and makes it difficult to meet users' deeper needs. In short, the technical threshold of AI protein design is becoming the biggest bottleneck for the industry's productivity. But the good news is, someone has finally started seriously addressing this problem.

3. New Solution: Finish Protein Design Just by Chatting with "MatwingsVenus™"

In April 2026, Matwings Technology launched a conversational protein R&D agent—MatwingsVenus™ (XiaoWu™)—at "Big Zero Bay" in Shanghai. What is the biggest change? From "here are a bunch of tools, you put them together yourself" to "you talk, it works." MatwingsVenus™ (XiaoWu™) is a one-stop platform centered on an intelligent agent (Agent), integrating over 200 protein design tools, tens of billions of real labeled protein data, more than 50 platform-certified experts, and over 30 expert-tuned Skills. What do you need to do? Just tell it your goal in natural language. For example, you can directly say: "Help me design a mini-binding protein for target XXX, with nanomolar-level affinity and stability higher than the natural ligand."

MatwingsVenus

Then MatwingsVenus™ (Xiaowu™) automatically breaks down the tasks: which backbone diffusion model to adjust, what sequence design strategy to use, whether molecular dynamics simulations are needed, how to integrate subsequent wet experiments… You don’t need to worry about the bothersome format conversions and environment setup in between.

But the most interesting aspect of MatwingsVenus™ (Xiaowu™) actually isn’t the 'one-stop' feature, but the dry-to-wet closed loop. After design completion, the platform can automatically import the results into the experimental workflow, drive robots to complete sample preparation, protein purification, and functional testing, and then feed the experimental results back to the AI model, forming an automatic 'design → validate → iterate' loop.

Professor Hong Liang, the founder of Matwings Technology and a professor at Shanghai Jiao Tong University, gave an analogy: MatwingsVenus™ (Xiaowu™) turns R&D capabilities that were previously only available to large enterprises and research institutes into infrastructure that individuals can also access.

Sounds a bit exaggerated? Let’s look at a real case.

4. Real Case: Designing an Immune Modulatory Receptor from Scratch

This is an innovative drug project targeting a certain immune modulatory receptor—in simple terms, this target is related to cancer, autoimmune diseases, and inflammation, with high clinical value.

The challenges are also obvious:

- The target is new, with no similar drugs to reference for design;

- The target surface has many polar regions and lacks a traditional 'druggable pocket,' making design difficult;

- The natural ligand itself has nanomolar-level high affinity, making it extremely challenging to create a mini-protein that can effectively block its function.

Under traditional paths, this type of project would likely burn a lot of money and take a lot of time. But on the MatwingsVenus™ (Xiaowu™) platform, the process goes much more smoothly:

The Agent automatically completes backbone screening, interface design, sequence optimization, and druggability prediction based on the target’s structure and functional requirements… After the computing end quickly outputs a batch of high-quality binder sequences, the automated wet lab immediately takes over to perform sample preparation and cell activity validation.

Final result: Several dozen new binder molecules with in vitro cell-blocking activity were obtained in one go, and these molecules not only meet activity standards but also possess potential for function inhibition and high affinity.

The full process of designing from scratch, from computation to wet experiments, was successfully validated.

The takeaway from this is: when an Agent can automatically link backbone design, sequence prediction, and activity evaluation, and compress the design-validation iteration cycle to nearly seamless, the time needed to translate a protein that has never existed in nature from concept to experimental validation will be far less than what the industry traditionally perceives.

5. Making Protein Design More 'Democratic'

On a larger scale, conversational agents like MatwingsVenus™ (XiaoWu™) are essentially promoting the democratization of protein research capabilities. Previously, a de novo design project required collaboration across multiple teams including structural biology, computational chemistry, AI algorithms, protein expression and purification, and functional validation. The organizational cost alone deterred many small teams.

Now, through a conversational interface, individual users or small teams can complete the full loop of 'model prediction → experimental validation → iterative optimization.'

Matwings Technology summarizes this approach as: design is validation, and validation is iteration. Moving from 'big platform-driven' to 'accessible to individuals' could be one of the most anticipated changes in the field of protein design in the coming years.

Final Thoughts

Back to the story of that friend mentioned at the beginning. A few days ago, he came to me again, saying that their lab has recently been trying MatwingsVenus™ (XiaoWu™). They no longer need to struggle with those toolchains themselves and can focus more on the protein design itself.

His exact words: 'It feels like switching from a manual to an automatic transmission. Although it's not perfect yet, at least I don't have to take apart the gearbox while driving.'

I remembered this metaphor for a long time. The value of technology is never just about how many top-tier papers it can produce or how many SOTAs it can achieve; it is about whether it can truly be used, allowing more people to do more creative work.

MatwingsVenus™ (XiaoWu™) is still rapidly iterating. If you are also struggling with toolchains or are interested in AI-driven protein design, feel free to check out the case studies on Matwings Technology's official website, or try the tool yourself.

See you in the comments section, and let's talk about which tools have ever trapped you~