Bets on Generative AI to Redefine Drug Discovery——IntelliGenAI and their foundation model approach

For decades, new drug development has been constrained by the notorious “double ten” rule: a 10-year timeline, $1 billion in costs, and barely a 10% success rate for bringing a new therapy to market. A newly founded Generative science startup, IntelliGenAI, believes it can upend that paradigm. By leveraging cutting-edge generative AI in structural biology, IntelliGenAI aims to dramatically accelerate drug discovery and improve success odds. The company recently closed an angel funding round reportedly in the “tens of millions of US dollars” to advance its technology and is gearing up for growth.

Merging Structural Biology with Generative AI

IntelliFold’s（which is the foundation model released by the company ） core innovation is a generative AI model for 3D biomolecular structure prediction. In essence, the startup has built a large-scale “base model” akin to DeepMind’s AlphaFold-3, but with broader capabilities tailored for drug R&D like affinity and allosteric site. The IntelliFold model can predict how different biological molecules – proteins, DNA/RNA, small-molecule drugs, ions, etc. – interacts with each other in three-dimensional space with high precision. According to a early version open technical report from the company, IntelliFold’s performance on key protein-structure benchmarks is on par with Google DeepMind’s latest AlphaFold 3, and the latest Pro version of the model has already been shown to comprehensively surpass AF-3 on public test datasets. This means the model not only computes protein folding, but can also anticipate binding conformations and even estimate binding affinity between a protein and a prospective drug molecule – a crucial metric for virtual screening .

A major feature of IntelliFold’s system is its controllability. By applying lightweight, trainable adapters, the base model can be guided toward specific tasks . For example, it can focus on predicting allosteric conformational changes – the subtle shape shifts a protein undergoes when a molecule binds to a distant site – without losing accuracy on the primary conformation . “Given a specific protein sequence, the IntelliFold model can predict its binding conformation and mode with a small molecule,” explains co-founder Ronald Sun, highlighting a key capability that addresses a clear market need in drug discovery. Beyond structure alone, the model can output an affinity value for the binding, potentially boosting the efficiency and accuracy of drug screening by orders of magnitude. These advances provide pharmaceutical researchers with a powerful tool to design and evaluate new therapeutic molecules far more efficiently than before.

The IntelliFold platform was developed in-house by the startup IntelliGenAI, which was founded in late 2024 amid a surge of interest in generative AI ventures . Ronald Sun, IntelliFold’s President, is a former tech venture investor who spent years backing frontier technology projects to win before deciding to build one himself . The chief scientist, Sun Siqi, is a Fudan University researcher who previously worked at Microsoft’s research labs , specializing in advanced large-language models for years after won "SOTA" in structure prediction on CASP12(Critical Assessment of Structure Prediction, 2016) . The founding team’s uncommon mix of AI expertise and structural biology know-how enabled them to create a sophisticated prediction model from the ground up, rather than just wrapping existing tools. Most team members have dual backgrounds in computational biology and deep learning, which Sun notes was critical for developing their own large model for scientific research. An early version of IntelliFold’s server has already been made available to collaborators and testers, showcasing the technology’s potential in real-world drug discovery projects.

“Generative Science” – A New Research Paradigm

IntelliFold’s approach exemplifies what Ronald Sun calls “generative science” – applying generative AI to scientific discovery in ways that fundamentally differ from the traditional research paradigm. For centuries, science has advanced through the painstaking process of formulating theories, deriving equations, and experimentally verifying each step . In drug development, for instance, researchers normally must identify a biological target, design a molecule, and iteratively test and tweak hypotheses in the lab. Generative AI offers a radically different path: instead of explicitly mapping out every molecular interaction with first-principles physics and chemistry, the AI model is trained on massive datasets of sequences, structures, and experimental results. It can then directly generate plausible solutions or predictions, even without a perfect human understanding of every mechanism .

According to Sun, this data-driven generative method can yield outcomes that are “relatively accurate, but absolutely faster and broader” in scope compared to traditional techniques. In other words, a well-trained model might not explain why a particular protein folds or binds the way it does, but it can predict what will happen much more quickly and across vastly more possibilities than any lab could test manually, and is currently one of the most effective and leading approaches for tackling harder and more complex binding problems, such as so-called undruggable targets. The true dawn of this generative science approach was marked by DeepMind’s AlphaFold2, which in 2020 solved the decades-old problem of predicting protein 3D structures from amino acid sequences . AlphaFold3 (announced in 2023) extended that capability to model interactions between proteins and other molecules like nucleic acids, small compounds, and even antibodies – opening the door for AI to guide drug discovery in a meaningful way.

Now, startups like IntelliFold are pushing this trend further. “We’re seeing a potential shift in the first principles of scientific research,” Sun says of the generative AI wave. “For the first time, it may be possible to expand human knowledge ten times faster and broader, even without fully interpretable models for every step.” Sun expects that harnessing AI in this way could boost research efficiency by at least an order of magnitude and allow scientists to explore options that were previously infeasible. In the pharmaceutical context, he notes, an AI-driven paradigm could drastically shorten the discovery cycle and reduce costs per new drug candidate. Success rates might improve “several-fold,” as advanced models uncover viable drug hits that human experts might overlook. By applying generative models directly to scientific exploration, IntelliFold hopes to turn what was once a slow, linear process into something more akin to rapid prototyping – testing countless virtual compounds and scenarios in silico, with only the most promising ones moving to physical trials.

Chasing SOTA by surfing the scaling law

Alongside symbolic language models and world models, scientific models that capture representations of natural laws and deep underlying regularities constitute the third top-level pillar of Artificial General Intelligence (AGI). Across a wide range of natural science domains—spanning the extremely concrete and the extremely abstract, the macroscopic and the microscopic—there exist objective structures that can be formalized, systematized, and ultimately operationalized as tools.

Historically, mathematical principles and empirical scientific experimentation were the primary means by which humanity explained natural forces and unlocked productivity, guiding sustained and transformative progress. However, since the advent of AlphaFold2 (AF-2), a new race toward state-of-the-art (SOTA) performance has emerged—one in which competition between models themselves has become the principal arena of intellectual rivalry.

The CASP competition, which had seen only incremental advances over several decades, entered a new phase at CASP12 in 2016 with the introduction of convolutional neural networks, and was ultimately brought to a decisive turning point at CASP14, where AlphaFold2 effectively solved the long-standing problem of single-protein structure prediction at near-experimental accuracy.

While years of progress in monomeric structure prediction—culminating in AF-2—have been of profound scientific significance and provided industry with far superior starting points for downstream research, they remain insufficient for one of the most critical challenges in early-stage drug discovery: co-folding. In this domain, the industry has long lacked a solution that simultaneously offers reliable interaction awareness and high-throughput efficiency.

The formal release of AlphaFold3 in 2024, with its generative-AI-based capability for complex and composite structure prediction, marked a major inflection point in the industrial value of biological foundation models. Its breakthroughs in predicting structures of diverse molecular complexes—including antigen–antibody systems and protein–small-molecule interactions—opened a new chapter. Building on this capability, AlphaFold-related platforms secured multiple landmark partnerships with multinational pharmaceutical companies such as Novartis and Eli Lilly, involving upfront payments in the tens of millions of US dollars and total deal values ranging from USD 1–2 billion.

In parallel, replication efforts and exploratory improvements have rapidly followed. Yet due to the exceptionally high barrier to entry, requiring deep expertise in both large-scale generative models and structural biology, progress has been incremental rather than explosive. Recent benchmark studies—such as the newly published FoldBench benchmark—have revealed meaningful overall progress in this direction, while also highlighting substantial headroom for further capability gains relative to current SOTA systems. while also highlighting that current SOTA methods still have considerable room for improvement on certain tasks.

Using biology as a representative example, it has become increasingly clear that how domain-specific scientific data and knowledge are tokenized and mapped into generative AI architectures is now the foundational and defining problem of Generative Science. Once domain science has been successfully tokenized and its feasibility validated, the next imperative is scaling—a process that demands iterative advances in model architecture and infrastructure.

Generative science is not simply about stacking more Transformer blocks. To build more powerful and effective models, researchers need a deep understanding of both domain-specific science and model architectures themselves, in order to find the right scaling directions. Otherwise, merely piling on more compute and data will still fail to achieve good results.

This process—scaling toward higher-capacity, higher-efficiency models, achieving stronger performance, integrating increasingly concrete problem settings and scenario-specific data, and forming a self-reinforcing dynamic flywheel of iteration—is rapidly emerging as the standard methodology of Generative Science.

At the same time, this methodology is expanding quickly beyond biology. Led by organizations such as DeepMind, innovators are actively applying the transformative potential of Transformers and generative-science principles to domains including, but not limited to, climate modeling, materials science, and nuclear fusion control.

IntelligenAI: Timeline and Innovations

Building upon architectural innovations and algorithmic evolution within GenAI frameworks, IntelligenAI has delivered a series of results that are on par with, or in some cases surpass, AlphaFold3 and current industry SOTA: 1. The Pro version outperforms AlphaFold3 across multiple key metrics. 2. By inserting LoRA adapters, the model achieves not only excellent performance in canonical (orthosteric) binding scenarios, but also exceptional capability in directed control tasks, including allosteric site targeting, pocket-guided folding, and epitope-guided folding. 3. One of the world’s first foundation models to achieve GenAI-based affinity prediction, demonstrating significantly superior SOTA performance across multiple datasets. 4. Among the validated and effective improvements currently under development (in internal testing, not yet publicly deployed), more than half stem from fundamental rethinking of model architecture and training paradigms, rather than from data scale alone—an approach expected to yield substantial performance gains in the next major release.