Benchmark Scores Don’t Equal Marketable Drugs! An AI Drug Discovery Company Decides to Build Its Own Pipeline

What Would a Traditional Biology Lab Look Like in the AI Era?

On the screen, researchers input a target, and the AI analyzes protein structures and generates candidate sequences; these sequences then proceed to the experimental phase for expression, purification, and functional validation, with the resulting data fed back into the model to initiate the next round of design.

In the past, progressing from a target to candidate molecules often required iterative cycles of structural analysis, sequence design, experimental screening, and optimization. Today, AI is reshaping certain aspects of this process.

However, “designing” a protein does not equate to developing it into a drug. A protein sequence that performs well in silico may still encounter challenges in the laboratory, including issues with expression, stability, binding affinity, and functionality.

As Dr. Wang Yuguang, founder of Toursynbio, stated in an interview with VCBeat: “The true measure of AI’s integration into the pharmaceutical industry is not based on benchmarks, but on its ability to design molecules that are truly functional, druggable, and capable of advancing to the Preclinical Candidate (PCC) stage or even clinical trials.”

This also means that what AI-driven protein design truly needs to bridge is not just the model itself, but the entire pipeline from computational design to experimental validation and then to data feedback.

Founded in 2023, Toursynbio chose to enter the field through AI-driven protein design, connecting protein models, scientific agents, and automated experimental systems. In this system, AI handles design and optimization, experiments conduct validation, and real-world data feeds back into the model, forming a closed loop of “design–validation–iteration.”

Moving from “design capability” to “validation and iteration capabilities” is precisely the industry gap that Toursynbio aims to bridge.

To understand the choices made by Toursynbio, it is necessary to first revisit the professional background of its founder, Dr. Wang Yuguang.

Dr. Wang Yuguang holds a Ph.D. in Mathematics from the University of New South Wales and currently serves as an Associate Professor at Shanghai Jiao Tong University, with research focused on geometric deep learning and graph neural networks. Around 2020, the emergence of AlphaFold 2 drew his attention to the integration of AI and protein science.

In Dr. Wang Yuguang’s view, the integration of AI and protein design offers clear industrial application potential.

On one hand, proteins themselves possess vast sequence and structural spaces, with traditional R&D relying more on screening, mutagenesis, and experimental iteration; on the other hand, large-molecule drugs such as antibodies, antibody-drug conjugates (ADCs), and peptides have already established mature R&D and commercialization systems, providing clear application scenarios for AI-driven protein design.

“Traditional protein and antibody development has relied heavily on screening and trial-and-error. AI offers the opportunity to shift the paradigm from ‘searching for molecules in nature’ to ‘directly creating novel molecules not found in nature that meet specific functional requirements,’” said Dr. Wang Yuguang, recalling his original motivation for starting the company.

From “Discovery” to “Creation”: The Starting Point for Toursynbio’s Initial Choice of AI Protein Design

In April 2023, Toursynbio officially commenced operations; in June of the same year, the company delivered its first commercial order.

Focusing on the interdisciplinary field of protein design, Toursynbio has gradually assembled a team covering artificial intelligence, computational biology, molecular biology, drug development, and automated experimentation. Among the core members disclosed by the company are a CTO with a Ph.D. in Computer Science from Johns Hopkins University, a technical lead from the Laboratory of Molecular Biology at the University of Cambridge, and pharmaceutical experts with extensive experience in innovative drug research and development.

For an AI-driven drug discovery company, such a team structure is not merely a simple aggregation of specialties. Protein design spans algorithms, structural biology, and experimental sciences. From generating candidate sequences via computational models to experimentally validating their expression, stability, and function, each step requires expertise from diverse fields.

Since then, Toursynbio has continuously advanced product development in the field of protein design, progressing from its early Protein Engine to the ProteinNova protein design AI agent, and further to the ToursynClaw superintelligent scientific platform for protein design.

At this point, the entrepreneurial path of Toursynbio has gradually become clear: Leverage AI and mathematical capabilities to enter the field of protein design, and gradually extend these capabilities to real-world drug discovery and development scenarios.

If large protein models address the question of “how to design proteins,” Toursynbio further attempts to solve how to orchestrate different AI capabilities to accomplish a series of tasks—from analysis and design to screening—against specific R&D objectives. This is also what distinguishes ToursynClaw from single-purpose protein design models.

As designed by the company, ToursynClaw comprises an AI Scientist Agent and more than 20 specialized algorithmic modules. In addressing specific R&D tasks, the system can invoke different tools across stages such as target analysis, epitope selection, protein structure and sequence design, and candidate molecule screening, thereby seamlessly integrating with experimental validation.

ToursynClaw Visual Interface

Within this framework, the ADFLIP inverse folding model stands out as a representative capability among numerous specialized modules. “Inverse folding” can be simply understood as follows: given a desired three-dimensional protein structure, one works backward to identify the amino acid sequence capable of adopting that structure. Unlike traditional protein structure prediction, which infers structure from sequence, inverse folding deduces the sequence from the structure.

Toursynbio’s self-developed ADFLIP employs all-atom modeling, incorporating information on atoms, functional groups, and surrounding environments such as ligands and DNA into the design process, while further accounting for certain downstream functional constraints. Building on this foundation, the company has also developed BackFlip to facilitate protein design in scenarios where complete experimental structural information is unavailable.

Toursynbio's Inverse Folding Module ADFLIP

Notably, ADFLIP and BackFlip are just part of Toursynbio’s more than 20 algorithms and professional modules，The company does not rely on a single model; instead, it develops specialized algorithms for different protein design tasks and orchestrates their integration through intelligent agents.

This platform-based approach has been validated in practical design tasks. In the 2026 International Protein Design Competition (ICLR 2026 Workshop, Adaptyv/ProteinBase RBX1 Binder Design Track), Toursynbio competed as ToursynClaw, securing second place globally and first among Chinese teams in the RBX1 disordered region-binding protein design task.。

According to the competition data disclosed by the company: Among the seven design sets submitted by Toursynbio for testing, two achieved binding results in experiments, yielding a hit rate of 28.6%. In contrast, among approximately 300 entries in the entire competition, the baseline hit rate was around 3% (approximately nine hits). The company’s hit rate for this project was thus about 9.5 times the baseline. In the experimental ranking based on KD (Dissociation Constant, a measure of molecular binding affinity where lower values typically indicate stronger binding), Toursynbio ranked second globally. Notably, one de novo-designed binder targeting the disordered region of RBX1 demonstrated an experimentally measured KD of 85 nM.

Toursynbio's design deep-bear-clay KD 85 nM·Medium Binder (Ranked 2nd Nationwide)

The RBX1 target itself increases the difficulty of design, as it is a target containing intrinsically disordered regions (IDRs). Compared to protein regions with relatively stable three-dimensional structures, IDRs lack fixed conformations, posing greater challenges for design methods that rely on well-defined structures.

The company further validated its similar de novo design capabilities on the GPCR (G protein-coupled receptor) target, the β1-adrenergic receptor. Disclosed case studies showed that 16 out of 25 designed molecules achieved binding, resulting in a binding hit rate of 64%, with the best KD reaching 9 nM.

GPCR Case Study: β1-Adrenergic Receptor

From computational design to integrated experimental validation, these cases provide concrete verification of Toursynbio’s AI protein design capabilities. However, for the company, generating candidate molecules via the model is merely the first step; closing the full R&D loop requires feeding experimental results back into the AI system to inform and refine subsequent design iterations.

This also becomes the "wet-dry experimental closed loop" that Toursynbio aims to integrate in its next step.

From computational design to experimental integration, only the first half of the journey is complete. For AI-driven drug discovery, the ability of models to generate candidate molecules is merely the starting point—only through validation by real-world experiments and feeding the resulting data back into the model can a continuous, iterative R&D loop be established.

This is precisely why Toursynbio is simultaneously deploying an automated wet-lab platform alongside ToursynClaw. If ToursynClaw serves as the “brain” responsible for design and decision-making, then the automated experimental system acts as the “hands and eyes” connecting the digital world with the physical world.

Super Intelligent Agent Based on the Integration of Dry and Wet Lab Experiments

In July 2026, Toursynbio announced a joint initiative with China Biopharmaceutical Group, Asmote, and MegaRobo to co-establish a “Super Factory” dedicated to AI-driven biopharmaceutical development. Under this collaboration, Toursynbio Intelligence provides the ToursynClaw scientific super-agent and protein design capabilities; MegaRobo supplies the automated experimental platform; Asmote is responsible for integrated communication, sensing, intelligent computing, and control capabilities; and China Biopharmaceutical Group contributes real-world drug R&D scenarios and industrialization resources.

Dr. Wang Yuguang believes, “No matter how powerful AI is, without the feedback of effective real-world experimental data, its cognition is nothing but a castle in the air.” Drug development involves complex variables; relying solely on limited historical data to train models makes it difficult to cover the various scenarios encountered in actual R&D. In contrast, automated experiments can continuously generate standardized data, providing new “nourishment” for model iteration.

This closed-loop system must ultimately translate into improved R&D efficiency. By integrating ToursynClaw with automated experimental production lines, the cycle for advancing effective antibodies from design to functional validation can be compressed to approximately 45 days. In terms of pipeline advancement, the first batch of 16 pipelines has entered the Preclinical Candidate Compound (PCC) stage; the company plans to cumulatively complete 30 differentiated innovative molecular pipelines covering modalities such as monoclonal antibodies, bispecific antibodies, ADCs, and cyclic peptides by the end of 2026.

Overview of Toursynbio's Pipeline in Development

Of greater significance is the shift in “personnel efficiency.” Dr. Wang Yuguang introduced that whereas experimental screening previously required substantial manpower, a team of approximately 10 can now simultaneously advance around 10 antibody pipelines to the Preclinical Candidate (PCC) stage. The organizational model of the R&D team is transitioning from “labor-intensive” to “algorithm- and automation-driven.”

To understand Toursynbio’s commitment to the “dry-wet closed loop,” one must first clarify its business model—it is not a company that sells software or AI services, but rather a biopharmaceutical company that leverages AI to drive its own pipeline R&D.

Its business logic is a typical “dual-wheel drive”: The first wheel is AI laboratory services, providing clients with services such as antibody screening, affinity maturation, humanization, and de novo design of Minibinders; having served more than 14 targets to date; The second wheel is the core engine—the pipeline of independently innovated drugs, leveraging its proprietary platform to independently design and advance a pipeline of high-value innovative drugs, with a strategy focused on rapidly establishing advantages through monoclonal antibodies while simultaneously expanding into bispecific antibodies, peptides, and ADCs.

“We aim to rapidly build a substantial drug library, stockpiling high-quality molecules at the preclinical candidate (PCC) stage, and then pursue licensing or co-development opportunities with pharmaceutical companies,” revealed Dr. Wang Yuguang. The company plans to complete the design and experimental validation of 10 bispecific antibody pipelines within 2026. It is precisely because the goal is to develop proprietary pipeline assets that Toursynbio must establish a fully integrated dry-wet lab closed loop, rather than merely delivering a software tool.

The viability of this closed-loop business model has garnered support from the capital market. Toursynbio, established less than two years ago, has completed four rounds of financing, investors include Chengmei Capital, Wuxi Capital Group, and Shanghai Angel Group, with Chengmei Capital participating in three consecutive rounds of follow-on investment.

Returning to the question posed at the beginning of this article: What AI-driven drug discovery truly needs to overcome is not merely model capabilities, but the disconnect between models, real-world experiments, and data feedback.

Toursynbio’s exploration connects scientific super-agents, automated experimentation, and data feedback—enabling AI to participate in R&D decision-making, assigning execution to automated systems, and driving the next round of design with real experimental data.

Its goal is to gradually transform the R&D process, which has historically relied on manual expertise and iterative trial-and-error, into an industrialized R&D system driven collectively by algorithms, automated experimentation, and data.

This pathway remains in its early stages. The overall hit rate of de novo design, the multi-parameter collaborative design of complex molecules such as bispecific and multispecific antibodies, and the further progression from PCC to IND and clinical development all require continued validation.

Yet the direction is clear: AI’s role in pharmaceuticals is evolving from a supportive function offering single-point prediction tools to a core engine that deeply participates in R&D decision-making, drives experimental iteration, and directly generates candidate molecules.

Toursynbio may not be the sole explorer on this path, but it offers a deconstructable practical approach—positioning hyper-intelligent agents as the hub for R&D decision-making, automated production lines as the execution system, and real-world data as the fuel for AI’s continuous evolution.

From “Drug Discovery” to “Drug Creation,” This Experiment in Industrialized New Drug R&D Has Only Just Begun.