AI-powered protein design tools for biologists worldwide
Artificial intelligence has already proven it can speed up drug development and deepen our understanding of many diseases. But for this revolution to move from theory to real therapies, the most advanced models need to end up directly in the hands of people at the bench: lab scientists, not just machine learning specialists.
That is exactly where OpenProtein.AI comes in, a company that built a no-code platform for AI-guided protein design. The idea is to let biologists and researchers use powerful foundation models to design proteins, predict structure and function, and even train their own models, all without writing a single line of code or setting up heavy compute infrastructure.
The company was founded by Tristan Bepler, who earned his PhD from MIT in 2020, and Tim Lu, former associate professor at MIT and PhD from the same institution in 2007. Today, OpenProtein.AI’s platform is already used by pharma and biotech teams of all sizes and is also available for free to academic scientists, helping bridge the gap between cutting-edge research and the people actually exploring new therapeutic targets.
According to Bepler, this moment is especially exciting because these models not only make protein engineering more efficient – shortening development cycles for therapies and industrial applications – but also expand the ability to create entirely new proteins with tailor-made properties. The long-term vision goes beyond proteins: the company is already considering how to apply similar approaches to other molecule types, building a sort of formal language to describe complex biological systems.
From MIT to OpenProtein.AI: how AI became a language for proteins
The story of OpenProtein.AI starts in 2014, when Tristan Bepler joined MIT’s PhD program in Computational and Systems Biology, advised by Bonnie Berger, a leading name in math applied to biology. That was when he noticed a serious limitation: we still understand far too little about the molecules that form the basis of life to build good predictive models for them.
In practice, that meant it was still hard to reliably predict how a complete genomic circuit would behave or how a protein interaction network would respond to changes. This gap sparked Bepler’s interest in studying proteins at a much finer level, connecting sequence, structure, and function with the help of AI.
During his PhD, he began using evolutionary data to try to predict amino acid chains that form proteins, even before AlphaFold emerged, the Google DeepMind model that is now a reference for protein structure prediction. That work led to one of the first generative models focused on understanding and designing proteins. The team started calling this type of system a protein language model because it treats amino acid sequences in a similar way to how a language model handles text.
The big challenge behind the project was easy to state but far from trivial to solve: if we know a protein’s function depends on its sequence and structure, would it be possible to use foundation models to skip the explicit structure step and go straight from sequence to function? This line of research paved the way for OpenProtein.AI.
Receive the best innovation content in your email.
All the news, tips, trends, and resources you're looking for, delivered to your inbox.
By subscribing to the newsletter, you agree to receive communications from Método Viral. We are committed to always protecting and respecting your privacy.
After defending his PhD in 2020, Bepler joined Tim Lu’s lab as a postdoc in MIT’s Department of Biological Engineering. At that time, the idea of integrating AI and biology was really starting to gain traction. In Lu’s group, Bepler helped develop better computational models for biologics design, demonstrating in practice a clear mismatch: the most advanced models were available, but many biologists could not use them because they did not know how to code or lacked access to GPU infrastructure.
That is where the core concept of OpenProtein.AI was born: build a platform that brings AI models closer to biologists, instead of forcing biologists to become AI experts.
No-code platform for AI-driven protein engineering
OpenProtein.AI was designed from day one as a machine learning in-the-loop environment for protein engineering. Instead of a bundle of loose scripts and libraries, the company developed a full platform with an intuitive web interface that allows users to:
- Upload data such as sequences, experimental measurements, and variant libraries;
- Train models tailored to the user’s own lab data;
- Generate new protein sequences in silico, at scale;
- Predict structure and function using protein language models and structural models;
- Run comparative analyses of variants to prioritize which ones should go to bench validation.
Users do not have to deal with command lines, clusters, or complex scripts. Instead, they work with guided workflows, clear parameters, and results visualized in a friendly way. For those who prefer code, the company also offers APIs for direct integration with custom pipelines, but that part is optional.
One of the core pieces of the platform is PoET (Protein Evolutionary Transformer), the in-house protein language model developed by OpenProtein.AI. PoET was trained on groups of evolutionarily related proteins, which allows the model to:
- Capture key evolutionary constraints relevant to stability and function;
- Generate sets of related proteins, not just random variants;
- Incorporate new experimental data without requiring full retraining, updating its understanding based on bench results.
In practice, this means researchers can use their own data to refine the model, optimize sequences for a given function, and then leverage the rest of the platform’s tools to deepen their analysis. It is common to see groups generating entire virtual sequence libraries, running everything through structure and function predictors, and only then selecting a lean subset of candidates for experimental testing.
This cycle cuts costs, time, and frustration, because the lab ends up spending resources mainly on the most promising candidates. On top of that, it is possible to start from a protein of interest and ask the model to generate new variants with similar or improved properties, which is helpful in both therapeutic projects and industrial applications.
An open toolbox for different protein types
One point the OpenProtein.AI team is keen to emphasize is that the platform was built as a flexible, open toolkit, not as a rigid pipeline tied to a single protein type or function. The idea is to support everything from industrial enzymes to therapeutic antibodies, including signaling proteins and other classes.
The company’s models are trained to understand proteins in a broad way, meaning they learn about the full space of possible sequences, not just narrow, highly specific families. This enables more exploratory workflows where users can test bolder hypotheses, try unusual domain combinations, or tackle poorly studied targets.
At the same time, this open approach does not rule out very targeted use cases. Labs can configure workflows focused on:
- Boosting thermal stability or stability at extreme pH;
- Reducing immunogenicity in human applications;
- Improving affinity for a specific receptor;
- Creating proteins with multiple functions, such as molecules that change behavior after binding to a target.
This flexibility matters because the frontier of biotechnology is shifting toward more complex therapies, more sophisticated protein logic, and dynamic systems that respond conditionally to the cellular environment.
Pharma partnerships and the evolution of PoET-2
The platform’s impact is not limited to academia. In 2025, major pharma company Boehringer Ingelheim began using OpenProtein.AI’s tools in its internal projects. The partnership evolved to the point where the two companies announced an expanded collaboration, embedding OpenProtein.AI’s platform and models directly into Boehringer’s workflows for protein engineering aimed at diseases such as cancer, autoimmune disorders, and inflammatory conditions.
This move shows how the industry is actually adopting state-of-the-art AI models to redesign its discovery and optimization pipelines for biologics.
On top of that, OpenProtein.AI recently launched a new generation of its model, PoET-2. This version can outperform much larger models using only a fraction of the compute and experimental data needed for training. In other words, PoET-2 was built to be more efficient and accessible, which is crucial in a landscape where the cost of training large AI models is rising fast.
The questions driving PoET-2’s development include some very fundamental ones for anyone working with proteins:
- How can we accurately describe the constraints that define a functional protein?
- What specific language can we use to represent these structural and functional limitations?
- How can we incorporate even more evolutionary information into sequence generation?
- How can we represent enzymatic reactions in a way that lets a model create sequences capable of performing exactly those reactions?
These questions all point to a clear goal: build a formal language for proteins that is rich enough to guide AI models in creating sequences that not only look plausible, but actually carry out useful biochemical functions in the real world.
The next step: dynamic proteins and multiple mechanisms
The current focus of many research groups, including those connected to OpenProtein.AI, is to go beyond simple predictions of binding events or static functions. The future vision is to use these models to design dynamic proteins capable of:
- Interacting with two, three, or more biological mechanisms at the same time;
- Changing function after binding to a specific target;
- Acting as components of molecular logic, responding conditionally to signals in the cellular environment.
Tim Lu, who now serves as an advisor to the company, points out that this shift requires models that consider not just a protein’s static state, but how it behaves over time, across different cellular contexts, and in complex interaction networks. That significantly raises the computational bar, but it also opens the door to smarter, more personalized therapies.
Why open access matters in AI for biology
As AI models get larger, more sophisticated, and more expensive to train, there is an increasing risk that they become highly concentrated resources in the hands of a few players with deep pockets. For the scientific field, that would be a serious problem: a lot of discoveries in biology still come from universities, public labs, and smaller research groups.
That is why OpenProtein.AI stresses the importance of maintaining an open ecosystem at the intersection of AI and biology, where at least a significant fraction of tools, models, and infrastructure remains accessible to the broader scientific community. Offering the platform at no cost to academic researchers fits squarely into this vision.
As projects grow more complex – with protein logic, dynamic therapies, and systems that respond in a coordinated way to multiple signals – traditional experimental setups start to show their limits. Having robust, well-trained, and accessible models helps offset these constraints and enables more groups to contribute data, hypotheses, and validations.
By combining AI, protein engineering, and easy-to-use tools, OpenProtein.AI is a strong example of one possible future for computational biology: instead of splitting those who code from those who pipette, bring both worlds together in a platform that works well for each side, speeding up the arrival of the next generation of therapies and biotech applications.
