Just as DeepSeek-V4-Flash's official release shook the global general-purpose large model community, China's life science version of DeepSeek followed closely behind.
Recently, GeneLLM, a life science vertical multi-omics large model independently developed by Kindu Biology (founded by 4 Oxford University returnees), has been successively published in top-tier international academic journals – Nature Communications and Advanced Science.
As the world's first multi-omics large model pre-trained directly on raw omics data, GeneLLM is another heavyweight model following Google's AlphaFold and Stanford's EVO 2, filling the gap in China's life science foundational large models. It can be called China's life science version of DeepSeek, allowing AI to begin understanding the multiple 'languages' and the entire 'system' of life.
Predicting the Next Piece of Life Information Like Predicting the Next Token
Disease recognition is just one application scenario for GeneLLM. What Kindu Biology truly aims to do is build the "Claude Code" for the life science field.
The core of large language models like ChatGPT, Claude, and DeepSeek is next-token prediction.
GeneLLM adopts a similar approach, but what it predicts is not text, but life information.
The four bases in RNA sequences – Adenine (A), Uracil (U), Guanine (G), and Cytosine (C) – become the basic tokens for GeneLLM to understand the language of life.
Traditional bioinformatics analysis typically relies on gene annotation, sequence alignment, and manually defined labels. While accurate, this method prematurely limits the model's cognitive scope and may lose vast amounts of unknown biological signals hidden in raw data.
GeneLLM, however, charts a different path, learning life's patterns directly from unprocessed raw sequencing data.
It uses multi-omics raw data such as RNA-seq, proteomics, and metabolomics as training data, allowing the model to autonomously discover disease-related patterns.
Currently, GeneLLM has completed pre-training of a 1.5 billion parameter model with 3.5 trillion base sequences, and the XLarge version has achieved pre-training of a 30 billion parameter model, continuously expanding the technological barrier.
As the world's first multi-omics large model pre-trained on raw sequencing data, GeneLLM includes two main stages:
(1) Unsupervised Pre-training & Prototype Discovery
(2) Patient-level Disease Fine-tuning (Disease Tuning)
GeneLLM first needs to solve a problem:
How to transform complex life data into a language AI can understand?
In natural language processing, the BPE algorithm splits sentences into Tokens; GeneLLM, however, splits RNA sequencing fragments of about 150bp length into life Tokens using a sliding window of 7 bases (7-mer).
Subsequently, the model utilizes the Transformer architecture to directly predict the next base without gene annotation or human labels.
This means the AI is not first consulting a human-compiled "dictionary," but learning directly from the raw signals of life.
During training, GeneLLM processed approximately tens of trillions of RNA reads, trained on a hundred-card cluster of NVIDIA A100 GPUs.
At this point, GeneLLM's generalization and other capabilities began to "emerge."
As a major innovative breakthrough in the life science field, this domestically produced life science model, GeneLLM, can be applied to various fields including new drug R&D, precision medicine, synthetic biology, environmental monitoring, microbiology & bio-agriculture, and protein & molecule design. It is one of the few large models globally that has achieved real-world scenario deployment.
Only after examining the underlying innovations of GeneLLM in "data, architecture, training," can we truly understand: why it represents China's "Bio Version of DeepSeek."
Silicon Valley stacks tens of thousands of H100s for trillion-parameter models, DeepSeek improves computing efficiency through algorithmic innovation, and GeneLLM similarly moves massive life science data with a deft touch.
If AlphaFold allowed AI to see life's "structure" for the first time, and EVO2 allowed AI to begin understanding life's "code," then GeneLLM attempts to further understand life's "system."
What's more remarkable is efficiency. Traditional methods rely on 6Gb deep sequencing, with high costs making deployment difficult. GeneLLM maintains AUC > 0.8 even at an extremely shallow depth of 1Gb (cost reduced by 83%).
This truly holds the promise of making affordable precision medicine a reality.
A new research paradigm begins to emerge: letting AI learn from life data, bringing life science into a new stage of predictable, computable, and scalable exploration.
Not Just a Model, but the "Last Mile" Intelligent Infrastructure
But Kindu Biology's strategy extends beyond the foundational model.
Using the GeneLLM multi-omics large model as the life cognition base, Kindu Biology further constructs an execution system connecting AI intelligence with the physical world. Through the Harness intelligent experiment execution layer and the DBTL (Design-Build-Test-Learn) data feedback loop, it achieves a complete AI for Science closed loop—from understanding life patterns and generating scientific hypotheses, to automated experimental validation and continuous iterative optimization.
As Liam Fedus, former OpenAI VP and head of post-training, said, current LLMs have exhausted the limited text and code on the internet. The next major advance in scientific discovery must rely on experimental iteration.
In other words, AI cannot discover new knowledge solely by reading what humans have already written; it must conduct experiments itself.
But here's the problem—
Internet services inherently have APIs; network information is inherently digital. But research equipment comes from different vendors with varying protocols, complex and expensive. No one can rebuild everything from scratch in a few months.
Most labs today are designed for humans: instrument panels, pipetting actions, sample states, on-the-spot judgments—the vast majority are not translated into machine-readable signals.
For AI to enter the lab, the current challenge is not just model capability; it requires a new set of infrastructure.
Therefore, for AI to enter the lab, the current issue is not just a model capability problem; it also requires a physical Harness: turning the lab into a system that can be compiled, scheduled, monitored, and traced.
This is the true last mile of AI4S.
BioFord Harness, Making the Lab Start "Running Itself"
For this purpose, Kindu developed a system called BioFord Harness.
This is an infrastructure that connects AI with the physical lab.
They are not making robotic arms imitate human hands, but transforming the lab into a system that can be compiled, scheduled, monitored, and traced.
In this process, the physical Harness must accomplish at least three things:
1. Compile scientific intent or experimental DSL into instructions executable by different devices;
2. Complete scheduling, resource and safety constraint management among multiple devices, and handle exceptions;
3. Let experimental results, device logs, and environmental parameters flow back, becoming inputs for the next round of model and experimental design.
Scientific question → AI understanding → experimental plan → device scheduling → execution → data feedback → model optimization → next round.
A flywheel for scientific experimental data is thus set in motion.
Five Intelligent Agents, Turning Research Workflow into an Assembly Line
The BioFord Agent embodied intelligent research platform connects downward to the physical lab layer and upward to the scientist's cognitive layer, bridging the gap between reasoning and execution with physical AI.
At the cognitive layer, BioFord Agent consists of a collaborative network of five intelligent agents, spanning the entire life science research workflow, potentially multiplying research efficiency.
Literature Retrieval Agent
Experimental Design Agent
Science Agent
Experiment Scheduling Agent
Data Analysis Agent
For example, the Literature Retrieval Agent can quickly help you search and read vast amounts of literature, complete literature reviews, and assist in hypothesis generation.
The Experimental Design Agent enabled the research team to shorten the experimental design cycle from several months to one week.
The Experiment Scheduling Agent launched by Kindu Biology, relying on the Universal Instrument Abstraction Layer, breaks protocol barriers between various heterogeneous devices.
Whether it's PCR machines, microplate readers, flow cytometers, or automated liquid handling workstations, they can all achieve unified management and scheduling. The system has built-in dynamic scheduling algorithms, enabling automatic batch scheduling, real-time conflict avoidance, and full recording of experimental parameters, forming a traceable audit trail.
This means AI is no longer stuck in the "suggestions" stage, but truly enters the lab, operates equipment, executes tasks, and becomes a trustworthy "research assistant."
Failure Data Might Be More Valuable Than Success Data
The deeper value this system addresses is: making every experiment—whether successful or failed—data that the system can digest.
In a traditional lab, a failed record might just be a line saying "results not as expected." The experience resides in people's minds; when they leave, the experience is gone.
But in the BioFord system, every failure is valuable training data—the logic of parameter selection, records of environmental conditions, proofs of erroneous paths... all are deposited, becoming part of the system's "experience."
Next time, AI knows: this path is blocked.
Interestingly, in this field, it's not the one with the strongest computing power that wins, nor the one with the largest model.
Computing power can be bought, but research data cannot.
Jin Yongcheng, Founder & CEO of Kindu Biology, stated: "During the R&D process, we gradually realized that AI for BioScience is not simply about stacking models and data. For those doing experiments, the ultimate problem to solve remains the dilemma of not knowing how to proceed after calculation, and not getting it right when proceeding."
This is the most important, and hardest-to-replicate, moat in the AI4S field.
Four Oxford People, Including a Senior Fellow of Luo Fuli
In 2022, when Jin Yongcheng received his PhD in Bioengineering from the University of Oxford, he faced a choice.
His supervisor was Hagan Bayley, a Fellow of the Royal Society and founder of the UK-listed third-generation sequencing giant Oxford Nanopore. The lab had a strong tradition of "translating research into practice." Staying in the UK was a clear, smooth path.
But he chose another path—packing a "prototype technology" from the lab into his suitcase and bringing it back to China. Accompanying him were three Oxford alumni: Biology PhD Deng Siwei, Computer Science Associate Researcher Sha Lei (PhD in Computer Science from Peking University, senior fellow of Luo Fuli, head of Xiaomi's large model), and Zhou Tianyao, skilled in product commercialization. The four possess complementary backgrounds in bioengineering, AI, computational biology, and business operations, forming a perfect team for highly interdisciplinary research. They had previously collaborated on joint research projects, achieving disease prediction and detection by combining AI and transcriptomics technology.
The company name "Kindu Biology": "Kin" is taken from "Oxford," symbolizing their starting point from top-tier academic institutions like Oxford labs; "du" (渡) signifies ferrying or helping others, representing the destination they believe AI for Science should reach.
Currently, Kindu Biology's BioFord Agent physical AI research platform has been deployed in several prestigious domestic universities with significant results, reducing research cycles from several months to one week.
The Wind Rises: 4 Funding Rounds in 1 Year, A Capital "Enchantment" Scene
However, walking a path "untrodden by predecessors" inevitably comes with solitude.
The early days of entrepreneurship were full of difficulties. At that time, AI startups were booming, but "AI+" attempts in the biological sciences were rare. "Those who understand AI may not understand biological science, and most who understand biological science don't understand AI."
Jin Yongcheng admitted: "Investors once couldn't understand what we were doing."
The team chose the "hardest path": starting from a biological foundational large model, a direction pursued by only a handful of companies globally.
In 2025, a turning point emerged.
That year, the State Council issued the "Opinions on Deepening the Implementation of the 'AI+' Initiative," listing AI for Science among the key areas. The field began to gain momentum.
Kindu Biology achieved the feat of "completing 4 funding rounds in one year." The company's main funding timeline is a benchmark for the industry.
Angel+ Round: Led by Sequoia Capital China Seed Fund;
Pre-A+ Round: Led by Chuangdongfang Investment with tens of millions;
Pre-A+ Round: Received tens of millions in investment from Nanshan Zhanxintou;
Series A Round: Led by Gaotejia Investment with nearly 100 million RMB.
Teng Yuhang, Executive Partner at Gaotejia Investment, stated: "Kindu Biology has transformed life science basic research into a subscribable, scalable 'computing power + experiment' infrastructure."
Jin Yongcheng's goal is even more ambitious: "We are not satisfied with just selling software; we want to build the intelligent operating system for the life science field. Just as Intel defined computing power in the PC era, we hope to define the new R&D paradigm for life science in the AI era."
From Oxford labs to Shenzhen, from being misunderstood to receiving heavy investment from top-tier capital, the story of the four Oxford graduates is not just a startup legend, but a soul-searching inquiry about "what we can do for humanity."
When AI learns to "stay up all night" doing research on its own, scientific discovery may no longer rely on the serendipitous inspiration of geniuses, but become a predictable inevitability. This journey has just begun for them.
Industry Landscape: Kindu Charts a Lightweight Physical AI Path
In the broader AI for BioScience landscape, Kindu is not alone, but its entry point is distinctly different.
The first category is Digital AI Scientists.
Stanford-incubated Biomni (commercialized as Phylo) possesses over 150 specialized tools, capable of automatically performing literature review, hypothesis generation, and bioinformatics analysis.
FutureHouse, backed by Eric Schmidt, is dedicated to building AI scientists that can autonomously generate hypotheses and write papers.
However, while powerful, they remain confined to the digital world.
The second category is the Full-stack Autonomous route.
XtalPi deploys over 300 "AI + robotics" workstations globally.
Lila Sciences, incubated by Flagship Pioneering, with $550 million in funding, attempts to let AI completely take over the design, execution, and redesign of experiments, aiming for "scientific superintelligence."
But these approaches are extremely capital-intensive.
The third category is the End-to-end Pipeline route.
Insilico Medicine pushes AI directly into its own innovative drug pipelines, with its first AI-discovered drug entering Phase III clinical trials, but they are "car builders" rather than "road pavers."
Kindu Biology's choice is: focus on building the physical Harness, creating the "last mile" infrastructure between models and physical experimental systems.
Not engaging in the foundational model race, not pursuing end-to-end pipelines, not creating pure digital AI scientists. Focusing only on that last mile is undoubtedly a more lightweight approach.
Jin Yongcheng articulated this judgment clearly: "The true watershed for AI for Science is not how well the model mimics a scientist's answers, but whether the lab can start functioning like a continuously learning system."
Moreover, in the global AI for Science field, Kindu is not simply "only doing the last mile."
More accurately, it uses the last mile as an entry point to compete for the orchestration rights of the entire scientific workflow.
Compared to pure digital AI scientists, it can interact with the physical world; compared to building heavy-asset science factories from scratch, it has the opportunity to take over clients' existing labs; compared to end-to-end AI drug discovery companies, it doesn't have to bet its fate on a single clinical pipeline.
Its true moat will not be parameter count, but the continuously accumulated experimental trajectories, device interfaces, failure experiences, and cross-laboratory execution network.
As Jin Yongcheng said, the myriad signaling pathways within organisms and the reaction mechanisms full of unknowns are fascinating. "As rich as biology is, so is the prospect of AI for Science."
Exploring the mysteries of life with AI's intelligence—the vast universe for these Oxford prodigies has just begun to unfold.
This article is from the WeChat public account "Xinzhiyuan", author: Xinzhiyuan; Editor: Aeneas KingHZ
