The Five Generations of Affinity Ligand Evolution: From Natural Extraction to Protein Large-Scale Model Design

In the purification workshop of biopharmaceuticals, a chromatography column might just look like a stainless steel tube filled with tiny beads, but the technical barriers behind it could be beyond what many people imagine. What determines the performance limit of an affinity matrix isn't just the bead substrate (pore size, rigidity, hydrophilicity), but more importantly the affinity ligand coupled on the bead surface—it defines the most essential purification behaviors, like binding specificity, affinity, alkali resistance, and elution conditions.

From naturally extracted Protein A, to genetically engineered recombinant affinity ligands, to directionally optimized engineered ligands, and now to AI-designed affinity ligands, each iteration of the ligand has been pushing the performance boundaries of affinity chromatography outward. The rise of protein large model-based ligand design is transforming the making of this "molecular key" from a long process of trial-and-error screening to a programmable, on-demand design era.

1. Tracing the technology: The evolution path of five generations of affinity ligands

 Comparison Chart Showing the Evolution of Fifth-Generation Affinity Ligand Technology

Affinity ligands are functional molecules that can bind specifically and reversibly to target biomolecules (such as antibodies and functional proteins) and are the core functional units for achieving precise separation in affinity chromatography. From natural discovery to artificial design, the technology has progressed through five generations, with each step enhancing technical barriers and industrial performance.

First Generation: Natural Affinity Ligands — The Foundation of Early Industry

The first-generation affinity ligands were directly extracted from natural microorganisms or plants and animals, with typical examples being natural Protein A, Protein G, and other antibody-binding ligands. Their main advantage is good biocompatibility and basic stability in binding specificity, making them core materials for early bioprocessing.

However, natural ligands have inherent unavoidable drawbacks: unstable structures, poor alkali resistance, and they cannot withstand industrial-scale CIP (Clean-in-Place) processes; large batch-to-batch variations and low purity can lead to ligand leaching and non-specific adsorption; dynamic binding capacities are low, seriously limiting large-scale production efficiency.

Second Generation: Recombinant Affinity Ligands — Standardized Upgrade

To address the issues of batch inconsistency and limited output in natural ligands, recombinant affinity ligands emerged. This technology uses genetic recombination to clone the core functional binding domains of natural ligands and prepares standardized ligand products through heterologous microbial expression, removing redundant segments of natural proteins while retaining the core binding functionality, achieving scalable, standardized production.

The core upgrades are: solving batch stability issues of natural ligands, greatly improving purity, and significantly reducing non-specific adsorption. The second-generation recombinant ligands realized standardized mass production, but modifications mainly focused on removing redundant fragments and did not yet deeply optimize core performance aspects like binding interfaces and stability through systematic protein engineering. Improvements in alkali resistance and dynamic binding capacity were still limited.

Third Generation: Engineered Affinity Ligands — Targeted Optimization

While recombinant affinity ligands allowed standardization, performance improvements were limited. To overcome this bottleneck, engineered affinity ligands became the next step. This technology builds on recombinant ligands and applies precise artificial modifications through protein engineering techniques such as site-directed mutagenesis, structural truncation, modular assembly, disulfide bond engineering, and surface charge engineering.

This approach can specifically optimize the core performance of ligands: key site mutations improve alkali resistance to adapt to 0.5-1.0 M NaOH industrial cleaning systems, significantly extending packing lifespan; optimizing amino acid composition and spatial arrangement at the binding interface enhances target molecule binding efficiency; and reduces ligand leaching and non-specific binding.

However, traditional engineered modifications rely on human experience and extensive wet-lab screening, with high costs and long iteration cycles, and can only achieve local fine-tuning, making it difficult to surpass the inherent performance limits of natural sequences.

Fourth Generation: AI-Designed Affinity Ligands — A Paradigm Shift

The performance ceiling of engineered ligands has pushed the industry to seek more efficient technological paths. AI-designed affinity ligands rely on deep learning algorithms and protein structure prediction models, replacing traditional manual modifications and random screening. Essentially, it’s "AI-powered intelligent optimization"—the input is known ligand sequences, and the output is optimized high-performance variants.

Unlike traditional engineering’s "local tweaks," AI can, through massive data training, accurately parse the coupling between ligand sequences, spatial structures, and functional performance, achieving multi-parameter simultaneous optimization: balancing affinity, stability, alkalinity resistance, expression levels, and other indicators all at once, avoiding the single-focus limitation of manual optimization. In terms of R&D efficiency, AI can significantly shorten the ligand optimization iteration cycle, compressing computation design time from months to weeks.

Fifth Generation: Protein Large Model-Designed Ligands — Creating from Scratch

While AI-assisted optimization greatly improves efficiency, it still relies on natural templates. Designing ligands with protein large models represents the highest current technology in affinity ligands—its essence is "AI-driven de novo creation," completely freeing itself from dependency on natural protein templates. Using self-supervised learning based on billions of protein sequences and structural data, it grasps fundamental rules of protein folding, interface binding, and physicochemical properties to design entirely new ligands from scratch.

This technology breaks the performance ceiling of natural ligands, allowing for custom ligands that are extremely alkali-resistant, ultra-high capacity, super-specific, and low-immunogenic—tailored to industrial needs. It not only optimizes existing antibody purification ligands but also rapidly creates exclusive affinity ligands for new fusion proteins, bispecific antibodies, peptide drugs, and other targets.

From natural extraction to large-model design, the evolution of five generations of ligands follows a clear line: R&D mode shifts from "trial and error" to "precise design," iteration cycles shorten from years to months, performance goes from "just usable" to "customized," and innovation boundaries expand from natural sequence limits to unlimited design space.

2. Platform Enablement: MatwingsVenus™ (Xiaowu™) Builds a Full-Chain Closed Loop for Ligand Design

 End-to-end closed-loop AI ligand design

To address key industry pain points like long ligand R&D cycles, slow performance iteration, lack of originality, and difficult implementation, Shanghai Matwings Technology's MatwingsVenus™ (Xiaowu™) AI protein large model platform has built a full-link affinity ligand R&D system. This system covers everything from intelligent ligand optimization, entirely new ligand design, performance simulation and prediction, to closed-loop validation with dry and wet experiments. It includes four core R&D scenarios: recombinant affinity ligand optimization, engineered affinity ligand iteration, AI-designed affinity ligands, and large-model from-scratch originality.

2.1. Data foundation: A 'library of evolutionary materials' with billions of sequences. At the base of the platform is the VenusPod dataset (developed by Professor Hongliang's team at Shanghai Jiao Tong University, recognized as one of the top ten tech advances at the university’s 5th Science & Technology Innovations, and second prize at the 2025 national finals of the "Data Elements ×" competition). This standardized dataset integrates 15 billion protein sequences globally, with 6.5 billion sequences tagged with extreme environment data like temperature, pH, and osmotic pressure. Exclusive resources include deep-sea MEER Mingshui Plan and microbial sequencing from salt lakes and polar regions, providing massive acid/alkali-resistant protein data to support ligand stability predictions (data source: Matwings Technology website news, May 2026; Shanghai Jiao Tong University official announcement, October 2025).

2.2. Algorithm core: dual-driven from optimization to originality. On the engineering optimization side, the platform can intelligently identify key mutation sites on existing recombinant affinity ligands while optimizing alkali resistance, binding load, and expression levels. This allows rapid iteration to produce high-performance engineered affinity ligands suitable for GMP-scale production, effectively solving traditional issues like ligand detachment, short circulation life, and low purification yield.

For high-level original design, the platform leverages protein large model capabilities without requiring natural templates. Based on user-specified target molecules, purification conditions, and performance requirements, it generates entirely new candidate affinity ligands. Through multi-level intelligent screening, unstable, low-specificity, and low-expression sequences are excluded, leaving only high-potential candidates for wet lab validation. This transforms the traditional "massive screening, low hit rate" approach into a new paradigm of "precise design, high success rate implementation."

2.3. Experimental closed loop: the final mile from computation to production. No matter how good the AI design is, experiments are essential for validation. The platform uses a mature dry and wet closed-loop system to quickly complete the full experimental workflow: ligand molecular cloning, protein expression, stability testing, binding activity verification, and resin coupling tests. Experimental data is fed back in real-time for model iteration and optimization, solving the industry problem of "excellent AI design performance but poor experimental implementation."

A typical case is the alkali-resistance modification of single-domain antibody ligands: to develop alkali-resistant affinity resins targeting non-antibody therapeutic proteins, Matwings Technology used the protein large model to increase the alkali resistance of an ordinary non-alkali single-domain antibody by 4 times in under a year and successfully applied it to a 5,000-liter production line.

3. Three Dimensions of Industrial Value

The value of AI-designed affinity ligands and protein large-model-designed ligands is not just about "doing ligand development faster and better." It is reshaping the fundamental logic of the affinity chromatography industry in three dimensions.

First, moving from "one universal resin rules them all" to an "era of customized ligands." The traditional resin industry model is "we make it, you use it" — a few universal ligands cover most applications. But with the rise of new molecules like bispecific antibodies, ADCs, nanobodies, and gene therapy vectors, universal ligands are increasingly unable to meet personalized purification needs. AI design dramatically reduces the cost and development cycle for customized ligands, and in the future, we might see a situation where "each new molecule has its own dedicated ligand."

Second, moving from "following and imitating" to "independent innovation." For a long time, domestic resin ligand technology mainly followed and imitated imported products. AI-designed ligand technology offers a chance to "overtake on a new track" — no need to chase after foreign technology routes, but instead build independent ligand IP and technology systems directly on the new AI design track.

Third, moving from "selling consumables" to "providing purification solutions." When ligand design becomes efficient and low-cost, the core competitiveness of resin companies shifts from the processing capability of "microsphere manufacturing and ligand coupling" to the design ability of "providing optimal purification solutions for customer molecules." Ligand design capability is becoming the core barrier for resin companies.

4. Conclusion

From the first-generation natural Protein A extracted from Staphylococcus aureus cell walls, to second-generation recombinant affinity ligands, third-generation engineered affinity ligands, fourth-generation AI-designed affinity ligand template optimization, and fifth-generation protein large-model-designed ligands created from scratch, the evolution of five generations of affinity ligands represents a complete upgrade in human ability to modify and create functional proteins.

During the natural and recombinant ligand era, domestic companies could only follow and improve on imported products. The AI protein large-model paradigm brings completely new development opportunities. Domestic AI protein large-model platforms, represented by MatwingsVenus™ (XiaoWu™), will provide core technological support for independent and controllable development in China's biopharmaceutical industry during this transformation.

With the continuous implementation of AI for Science technology, China's affinity chromatography consumable industry will gradually break free from dependence on overseas technologies, relying on independently designed ligand IP to solidify upstream supply chain security in biopharmaceuticals, achieving the leap from industry followers to original leaders.