Few-shot learning empowers efficient protein research and development in the biomedical field

In the process of applying artificial intelligence, insufficient data is a common problem. Traditional AI relies on a massive amount of labeled samples for training, and once the number of samples is scarce or the scenario changes, the model's performance will be greatly reduced. This pain point is particularly acute in the fields of biomedicine and biomanufacturing, especially in protein design and development.

For a protein composed of 361 amino acids, replacing just one amino acid can result in nearly 7,000 possible mutations; replacing two amino acids increases the mutation combinations to more than 23 million; if three amino acids are replaced, the possibilities rise to about 53.3 billion. Each mutation scheme must be validated through actual wet-lab experiments. Large-scale experiments not only require enormous costs but also consume a lot of manpower and time, and the traditional R&D model is no longer suitable for the current pace of innovation.

Massive Proteins

In this context, few-shot learning technology has become the key to breaking through industry dilemmas. The MatwingsVenus™ (Xiaowu™) agent launched by Matwings Technology is a benchmark product that deeply implements few-shot learning in the field of protein design. The following will provide a detailed interpretation of this solution by combining technical principles, algorithm capabilities, and practical cases.

1. Why traditional AI is difficult to adapt to protein R&D scenarios

The core logic of traditional deep learning relies on massive amounts of data to train models. Only with sufficient, diverse datasets can the model gradually summarize feature patterns and develop stable decision-making abilities. This mode has two rigid prerequisites: first, the availability of large-scale, high-quality data; second, the data must cover various actual R&D scenarios. Once the samples are insufficient, or new mutation types appear that are not in the training data, the model is prone to prediction deviations.

The field of protein engineering is a typical 'data desert.' On one hand, wet lab processes related to proteins are complex and costly, and obtaining each set of valid experimental data requires extremely high costs; on the other hand, the number of possible protein mutation combinations is nearly infinite, making it practically impossible to cover all possibilities and accumulate massive mutation data.

Data Desert of Protein

Traditional deep learning, in order to achieve high-accuracy protein function prediction, must rely on tens of thousands of mutation data sets for training. This creates a difficult-to-break vicious cycle: researchers hope to use AI to reduce experimental workload, but the prerequisite for using traditional AI is first to complete a massive number of wet experiments to accumulate data. Due to this constraint, traditional AI has long been unable to achieve large-scale adoption in protein design and biomedical research.

2. Few-shot learning: From 'memorizing features' to 'summarizing underlying patterns'

The technical approach of few-shot learning completely breaks away from the logic of traditional deep learning, which 'relies on massive samples to memorize features.' It no longer makes the model mechanically replicate a small number of samples but focuses on cultivating the model's underlying ability to generalize: leveraging pre-accumulated general knowledge to quickly understand and adapt to entirely new research tasks.

This logic can be understood with a simple example: experienced technicians, who have accumulated knowledge of various materials, structures, and characteristics over many years, can quickly grasp the core features and evaluate advantages and disadvantages of a new research subject by referring to only a few samples. Few-shot learning is about building such a vast 'knowledge experience base' for AI.

The model first completes foundational learning through massive general data, thoroughly understanding the underlying relationships between biological sequences, structures, and functions. When tasked with entirely new protein research, even if only a few experimental samples are provided, the model can utilize its existing knowledge reserves to quickly capture the core features of the new samples and complete task adaptation.

Few-shot learning

In the field of protein design, Professor Hong Liang's team at Shanghai Jiao Tong University proposed a small sample learning method for FSFP specifically for protein function prediction. This method integrates meta-learning, sorting learning, and efficient parameter fine-tuning, enabling optimization of protein pre-training models using just a few dozen wet experimental datasets, greatly improving the accuracy of protein mutation and functional property prediction.

In the ProteinGym test containing 87 high-throughput mutation datasets, even in challenging scenarios where the original pre-trained model's predictive correlation was below 0.1, integrating the FSFP method and training the model with any 20 wet experimental data points could significantly improve the predictive correlation to above 0.5.

At the same time, to study the effectiveness of FSFP. In a specific protein Phi29 modification case, wet experiment validation was conducted. Using only 20 wet experimental data points to train the model, FSFP was able to increase the single-point mutation prediction rate of the original pre-trained protein ESM-1v top-20 by 25%, and nearly 10 brand-new positive single-point mutations were identified.

3. MatwingsVenus Agent: Engineering-based application of small-sample learning in protein R&D

The conversational protein R&D agent MatwingsVenus™ ™, officially released in April 2026, integrates the core ideas of few-shot learning methods, relying on a small amount of wet experimental data to enable rapid fine-tuning and precise iteration of the model.

The complete working logic of the MatwingsVenus™ ™ agent is divided into four stages, achieving a closed loop from algorithm theory to R&D practice in few-sample learning:

First, massive pre-training. Before the product launch, the model had already undergone basic training using nearly 10 billion protein sequence datasets, including special sequences that withstand high temperatures and acids and bases in extreme environments such as volcanoes and deep seas. The massive data enables the model to establish a complete foundational understanding of protein sequences, structures, and biological functions.

Second, small sample adaptation. For specific protein modification and new molecule development tasks, users only need to provide dozens or even fewer wet experimental data sets to quickly fine-tune the model, eliminating the need to train the model from scratch, greatly shortening the preparation cycle.

Third, the wet and dry closed loop. The system connects the entire chain of AI design and automated experiments: after AI outputs candidate protein sequences, the system automatically links to the experimental validation stage, and the real results produced are fed back in real time to the next round of AI design. This "design is verification, validation is iteration" model truly implements few-sample learning as an executable R&D process.

Fourth, continuous iteration. During operation, the system continuously absorbs new experimental data and optimizes model judgment accuracy, becoming increasingly tailored to the needs of specialized R&D scenarios.

This architecture brings a revolutionary change: what used to require thousands of wet experiments to select the optimal protein sequence now requires only a few dozen or fewer experiments to quickly lock on targets, achieving a qualitative leap in R&D efficiency.

MatwingsVenus

4. Case Study: AI-Assisted Alkaline-Resistant Protein Engineering Achieves Industrial Application

The value of technology ultimately needs to be validated through actual projects. In 2024,Matwings Technology partnered with Kintor Pharmaceutical, a leading domestic company in the growth hormone field, to carry out alkaline-resistant modification of a key filler protein used in biopharmaceutical production. During the production process, this protein requires the removal of other organic contaminants on the chromatography medium after purifying growth hormone using affinity chromatography. This helps reduce non-specific adsorption, promote medium regeneration, and achieve sterilization, thereby improving purification efficiency and reusability of the medium. Consequently, strong alkali (0.5M NaOH, pH 13–14) is needed to elute these contaminants, but such strong alkaline conditions can disrupt the spatial structures of most proteins, making the research and development extremely challenging.

Relying on its self-developed protein design large model, Matwings Technology combined a small amount of wet lab experimentation with closed-loop iterations. In less than a year, they increased the alkaline resistance of a previously weakly alkaline-tolerant single-domain antibody by fourfold and simultaneously doubled the lifespan of the chromatography filler.

Currently, this modification achievement has been successfully implemented in 5000-liter scale industrial production, saving the partner company tens of millions of yuan in annual operating costs. It is also the world's first protein product designed by a large model to achieve 5000-liter scale mass production.

To date, the company has successfully delivered over 30 protein design projects, with business covering various biopharmaceutical and biomanufacturing subfields such as innovative drugs, in vitro diagnostics, industrial enzymes, and synthetic biology.

Compared with the traditional R&D model, using this AI solution has shortened the project development cycle from the original 2–5 years to 2–6 months, and the experimental success rate has significantly increased from the traditional 0.1%–1% to over 30%.

5. The Transformations Brought by Few-Shot Learning to the Biopharmaceutical Industry

Looking across the entire biomanufacturing and biopharmaceutical sectors, few-shot learning and intelligent products like MatwingsVenus™ (XiaoWu™) agents have brought two significant, tangible industry transformations.

First, they comprehensively expand the application boundaries of AI. In the past, AI technology could only be applied in mature scenarios with abundant data; now, in areas such as new molecule research, novel protein modification, and cutting-edge biological exploration where data is scarce, AI can also find feasible applications. Many R&D projects that were stalled due to "insufficient samples" now have entirely new solution approaches.

Second, they greatly lower the threshold for intelligent applications. Researchers can independently complete the full process of protein design through natural language dialogue with the platform, no longer heavily relying on specialized algorithm teams. The shortage of talent no longer hinders the industry's intelligent transformation, accelerating the adoption of AI technology in the biopharmaceutical field.

MatwingsVenus™ (XiaoWu™) agents from Matwings Technology are continuously deepening along this direction, helping biopharmaceutical practitioners steadily advance intelligent upgrades amidst the R&D challenges of "limited data, tight timelines, and high technical difficulty." From alkali-resistant filler protein modification, to the discovery of plastic-degrading enzymes, to breakthroughs in the development of various innovative drug target proteins and industrial enzymes, creating greater value with fewer experiments demonstrates the tangible application value of few-shot learning in protein engineering and the biopharmaceutical field.