The Hangzhou Team Named by Nvidia Closes the "Last Mile" of AI for Science

The Hangzhou Team Named by Nvidia Closes the "Last Mile" of AI for Science

From idea to result, achieved in a single conversation

Jessica, reporting from Aofeisi

QbitAI | WeChat Official Account QbitAI

The competition in AI for Science is extending toward the closed loop of scientific research.

Take protein design as an example. Although protein models are becoming increasingly powerful, to actually design a molecule, researchers still need to search literature, look up structures, download weights, configure environments, and submit computational tasks to servers.

After the model produces results, they still need to switch to another tool to view structures and perform screening to determine which candidates are worth pursuing... Once a computation is completed, where to start the next step often still relies on researchers to connect the various stages.

Now, the work hidden before and after the model is being reexamined by AI companies.

In April this year, OpenAI launched GPT-Rosalind for life sciences; in June, Anthropic released Claude Science; in August, Anthropic published research on Agents participating in protein design and connecting with external experimental validation.

These advances point to the same shift: AI is no longer solely responsible for the computational step of the model itself, but is also beginning to participate in the research work before and after the model—analyzing results, adjusting plans, and continuing to advance tasks.

In China, a team has already reached a similar crossroads.

A Hangzhou-based AI protein design company, LevinsLab, recently released Lévin™ Harness, now open for use by the scientific community.

It integrates data, models, plugin tools, computing power, and workflows into a single Agent workspace. Researchers define the goals, methods, and key decisions, while the Agent advances computations, analyzes results, and adjusts the next round's plan based on feedback.

In Lévin™ Harness, users can select specialized scientific models and analysis tools to complete computational tasks such as protein structure prediction and sequence design. However, Lévin™ Harness itself is not bound to any specific model; the general model layer remains open.

New models and scientific capabilities can continue to be added via plugins. Anyone can create dedicated plugins for their own models or models they want to add to the Agent's capabilities.

Moreover, once a research method is validated, it can be consolidated into a Workflow for repeated runs and further modifications.

The core responsibility of Lévin™ Harness is to organize models, tools, computing power, and the research process onto the same work chain.

This precisely addresses the problem AI for Science currently faces, which is far more complex than single-step inference:

How to make these intelligences truly enter the research setting and continue working.

AI Models Solve Scientific Problems, Agents Begin Organizing the Research Closed Loop

Lévin™ Harness, released by LevinsLab, is an Agent-centric protein design application. It can also be roughly understood as an Agent workbench specifically serving life science research.

Within the same workbench, researchers can set goals, view molecular structures, and organize computations, as well as modify plans after results appear. These actions always revolve around the same project, and information obtained in one step can continue to be used in the next.

The Agent here relies on general large language models (open models such as DeepSeek, Kimi, GLM, Qwen, and Xiaomi MiMo) to understand the researcher's needs and plan steps, then calls the corresponding tools to execute tasks. Users only need to describe their needs in natural language; specially trained scientific models and analysis tools handle computational tasks like protein structure prediction and sequence design.

Harness is responsible for connecting these capabilities with the required data and computing power, and retaining the research process.

Researchers only need to be responsible for scientific goals, methods, and key decisions, and can review results, supplement conditions, or correct directions at any time.

This division of labor is essentially to shorten the path from idea to result as much as possible, allowing the complex design process to be completed within the same conversational work environment. Even researchers without programming knowledge can quickly complete protein design.

So how is this "end-to-end" process specifically achieved? Let's use a protein binder design task as an example:

A binder is a protein molecule that can bind to a target protein. Traditional binder design requires researchers to figure out which target it needs to bind to, at what position, and what purpose they hope to achieve through this binding.

After clarifying these questions, the team can propose constraints to the design model, generate candidates, and then screen out molecules worth investing in experiments.

Every step in this path affects subsequent choices. Therefore, from the very beginning of the task, literature, structures, and design conditions need to be considered together.

In Harness, researchers can propose the goal of designing a binder for a specific target. The Agent participates in target research, organizes relevant information, and assists in analyzing selectable binding positions. The researcher then determines the methods and constraints based on the research purpose, giving subsequent computations a clear direction.

However, a potential issue here is that the Agent needs to "understand" the molecule the researcher is working on.

Proteins are formed by linked amino acids, and the amino acid units in the chain are typically called residues. This chain folds into a three-dimensional structure, and residues that are far apart in the sequence may be close to each other in space.

Whether a binder can form a suitable binding interface with the target is directly related to these spatial relationships. Therefore, information such as which residue the researcher selected and which two conformations are being compared also needs to be incorporated into the Agent's scope of work.

To this end, LevinsLab has built a 3D molecular visualization tool into the Agent workspace. Molecular structure information and related operations on the screen can be directly converted into the Agent's context, directing the conversation toward a specific research object.

When a researcher uses vague terms like "here" in the conversation, the Agent can understand the specific spatial location; whether to adjust a certain residue or further examine local interactions also has a basis for joint review, allowing structural analysis to follow the task throughout the design process.

With a basis for judgment, the next step is to translate the design ideas into computation.

Many specialized tools have emerged in the field of protein design. For example, models like AlphaFold are used for structure prediction, ProteinMPNN can design amino acid sequences for a given structure, Pallatom is for all-atom protein generation, and RFantibody serves antibody design... The Agent will autonomously select the corresponding capabilities based on the task intent.

However, once there are many tools, using them itself generates workload. The inputs, outputs, and runtime environments of different models may not be consistent. How to pass files from one computation to the next tool, and how to convert design conditions into parameters, all require someone to handle.

To solve this problem, Lévin™ Harness connects professional capabilities into the same environment through a customized plugin system.

For plugins already integrated, the system can automatically complete tool download, installation, and invocation. The Agent then autonomously selects appropriate tools, configures parameters, and links the corresponding computations based on the goal.

Moreover, users can continue to add new models and scientific capabilities via plugins, building their own dedicated plugins and preserving room for choice in research methods.

At the same time, computing resources are also connected to this work chain.

Structure generation and batch screening often require continuous GPU jobs. Harness can connect to a local workstation or submit tasks to remote GPU servers, allowing large-scale computations to run in the background. After completion, the Agent analyzes the output and advances subsequent steps.

Returning to the binder task, the Agent can now organize candidate generation and evaluation, and provide screening basis combined with structural analysis.

However, when one round of computation ends, the research is usually not yet complete. If candidates do not meet constraints, or perform differently across various evaluation metrics, researchers still need to determine where the problem lies.

This is where Workflow comes in, organizing these capabilities into executable and modifiable research methods. In Harness, this means the researcher determines scientific methods and key decisions, the Agent assists in connecting, running, and debugging, and then participates in adjustments based on results.

For example, if none of the candidates in a round meet expectations, the researcher and the Agent need to review the binding sites and design constraints, examine the generation methods and screening strategies, and then decide which part to modify.

The Agent can propose suggestions and execute new plans, while the researcher reviews these judgments, allowing the task to continue advancing around the original research goal.

After candidates enter experiments, the organized experimental results can also be returned to the project, becoming the basis for the next round of design. If a molecule does not exhibit the expected function, the team needs to reexamine the evaluation methods.

An intuitive understanding of this process is that experimental operations are still completed by the experimental team, while Harness takes on the collaboration between researchers, computational tools, and feedback.

Through multiple rounds of adjustments, validated tool combinations, parameters, and inspection steps can remain in the Workflow for reuse in similar tasks.

The product philosophy of Lévin™ Harness always revolves around how to continue advancing research tasks, connecting different stages around the research goal to form a scientific research closed loop.

Currently, Lévin™ Harness is open to the scientific community and supports Macs with Apple silicon.

As an Agent developed by LevinsLab, it naturally inherits the team's years of accumulation in protein design. To understand what it looks like today, we must go back to where the company started.

Why Does an AI Protein Design Company Build Its Own Harness?

Hangzhou LevinsLab Biotechnology Co., Ltd. was founded in September 2021. The company's name comes from the classic proposition in protein science—the Levinthal Paradox:

A protein theoretically exists in an enormous number of possible conformations. If one relied entirely on random attempts to find its native structure, the theoretical time could far exceed the age of the universe; however, proteins in nature can complete folding in a very short time.

Thus, how to quickly find an effective path from a huge search space has become one of the long-standing questions in protein science.

The problem LevinsLab initially wanted to solve also revolved around proteins. It is understood that their initial goal was to become a universal protein design company—providing design capabilities for different molecules and applications while developing their own products.

"Universal" means the team needs to face different directions such as binding proteins, antibodies, cyclic peptides, and industrial enzymes. These proteins have different uses and requirements. LevinsLab needed to find design capabilities that could be shared, and then put these capabilities into specific projects for testing.

Choosing such a universal route is related to LevinsLab's founding team.

The founder and CEO, Wang Haobo, holds a Ph.D. from Peking University and completed a postdoctoral fellowship at Harvard University. He studied under protein AI scholar Sergey Ovchinnikov, and his research experience spans organic chemistry, structural biology, bioinformatics, and artificial intelligence.

The LevinsLab team itself also has strong interdisciplinary characteristics, bringing together expertise and real experimental capabilities in computer science, structural biology, and chemistry.

Early on, LevinsLab built the Lésign intelligent protein design platform, and subsequently pushed its technological focus further toward all-atom protein generation.

The most representative work among these is Pallatom, the all-atom protein generation model self-developed by LevinsLab.

Since the backbone forms the main structure of a protein and side chains are attached to the backbone, the shape, position, and chemical properties of the backbone, sequence, and side chains all affect how the protein folds and how it interacts with other molecules.

Many protein generation methods first generate the backbone, then design the matching sequence, and finally optimize the side chains. Researchers need to continuously coordinate these steps.

Pallatom, on the other hand, operates from the all-atom level.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.