Quanquan Gu's team at Geodesic Intelligence releases NovaAtom-Lite…
Geodesic Intelligence, founded by Quanquan Gu, released the all-atom structure prediction model NovaAtom-Lite on November 10, ranking first on both the antibody-antigen and protein-ligand benchmarks of FoldBench and surpassing IsoDDE from DeepMind's Isomorphic Labs. On the antibody-antigen task, its DockQ success rate is 31 percentage points higher than AlphaFold 3 and 3.5 percentage points higher than IsoDDE; in the 0-20% ligand group with the lowest similarity to the training set, its success rate is 47.6%, compared with 24% for AlphaFold 3. The team says NovaAtom-Lite is only the first member of the NovaAtom family, and larger models are already on the way.
Why read this
AI score 69
The original text provides comparisons on multiple FoldBench benchmarks and grouped tests on unfamiliar molecules, which can be used to assess the new model's relative position in structure prediction.
Topics
Body · AI translation
Recently, Anthropic CEO Dario Amodei said something:
AI will compress 50 to 100 years of human progress in biology into 10 years—cancer, Alzheimer's, heart disease, all expected to be conquered.
That's a bold claim, but many people figured it was just talk.
Until a scientist named Gu Quanquan stepped forward and said: 10 years is too slow—we need to compress the drug development cycle to 4 years!
On November 10, his company Geodesic Intelligence, founded just 5 months ago, released the all-atom structure prediction model NovaAtom-Lite.
On FoldBench, the world's most rigorous biomolecular prediction benchmark, NovaAtom-Lite knocked Google DeepMind's Isomorphic Labs flagship model IsoDDE off its perch.
It topped a series of biopharmaceutical benchmarks, taking multiple No. 1 spots!
A 15-person startup, a few hundred GPUs, 5 months. On the other side: Isomorphic Labs, which has raised $2.7 billion cumulatively and is negotiating a new funding round at a $40 billion to $50 billion valuation.
X exploded instantly.
Some marveled: "Dario Amodei predicts cancer will be conquered within 10 years, while NovaAtom's founder says it will take only 4."
Others did the math: "Google DeepMind spun off Isomorphic, burned billions of dollars, and talked its valuation up past $40 billion. Then a UCLA professor with a 15-person startup surpassed it in 5 months."
And this isn't the first time Gu Quanquan has done something like this. Previously, while leading a team at ByteDance, he built SeedFold—the world's first model to surpass AlphaFold 3 on FoldBench.
It was the emergence of SeedFold that forced Isomorphic Labs to bring out IsoDDE and reclaim the throne of structure prediction. And this time, NovaAtom-Lite has knocked IsoDDE off again, and the throne has changed hands once more.
Even more striking: according to information Gu Quanquan revealed on X, NovaAtom-Lite is just an appetizer—a bigger and stronger NovaAtom-Pro is waiting in the wings.
In other words, even if Isomorphic Labs releases a stronger model soon in an attempt to counterattack, what awaits it may be another breakthrough from NovaAtom-Pro.
In this battle for the throne of structure prediction, Gu Quanquan has already won two rounds, and he still holds the cards for the third.
From now on, humanity's timeline for conquering terminal diseases has been moved up again!
The biotech world has already reached its "HBM breakout moment," one investor remarked.
So what exactly does NovaAtom-Lite do?
Why is it so hard to make drugs?
Because the essence of a drug working is getting one molecule to precisely "bite" another molecule.
Antibodies need to bite antigens; small-molecule drugs need to fit into protein pockets.
This "biting" happens at the atomic scale—the angstrom (Å) level, where one angstrom equals one ten-billionth of a meter.
A slight deviation, and the drug doesn't work.
Traditional methods rely on wet-lab experiments, testing one by one. On average, a new drug takes 13 years from lab to market and burns nearly $2 billion, with a success rate below 5%. The value of AI lies in simulating this "biting" in computers first, greatly shortening the trial-and-error cycle.
That is exactly what NovaAtom-Lite does—
Given the sequences of proteins, DNA, and RNA as well as information on small-molecule ligands, it can predict the three-dimensional spatial coordinates of every heavy atom (every atom other than hydrogen) in the entire complex.
Not a blurry outline, but positions accurate to the atomic level.
Moreover, it is not a "rigid" calculator.
NovaAtom-Lite adds a highly engineering-valuable set of "conditional constraints and prompts" features.
And what it outputs is not just a set of coordinates.
The model simultaneously returns local structure confidence pLDDT, predicted alignment error between residues PAE, and the interface reliability metric ipTM—in simple terms, it not only gives you an answer, but also tells you "how confident I am in this answer."
Two heavy punches: 79.07% and 77.06%
To evaluate how strong a structure prediction model is, the pharmaceutical world has two core touchstones: antibody-antigen binding prediction and protein-ligand binding prediction.
The former is the core of large-molecule drugs (such as anti-cancer antibodies)—you need to know how an antibody precisely "locks onto" its target; the latter is the lifeline of small-molecule targeted drugs—the drug molecule must accurately fit into the protein's binding pocket.
Two directions, each harder than the other.
FoldBench is currently recognized as the most rigorous all-atom biomolecular prediction benchmark, containing 1,522 complex structures, all from new structures after the model training cutoff date, eliminating any possibility of "seeing the answers in advance."
First punch: antibody-antigen. FoldBench includes 172 antibody-antigen binding interfaces.
The evaluation criterion is the DockQ score (which can be understood as how similar the predicted binding pose is to the real structure; above 0.23 counts as a successful prediction, and the higher the better).
A 5-month-old company, 31 percentage points higher than AlphaFold 3 and 3.5 percentage points higher than Isomorphic Labs, which has raised $2.7 billion.
Even more striking: if the standard is raised to DockQ ≥ 0.80 (requiring extremely high-precision predictions), NovaAtom-Lite's success rate is 2.2 times that of AlphaFold 3.
Second punch: protein-ligand. This test covers 558 binding interfaces, with even stricter criteria: it requires not only that the predicted position of the small-molecule drug deviate from the true position by less than 2 angstroms (ligand RMSD 0.8). Both conditions must be met simultaneously to count as a success.
On both core metrics, it ranks TOP 1 across the board.
Does it really "understand" physics, or can it only memorize answers?
Whenever a model tops the charts, the first reaction in the field is always: did you peek at the answers?
This doubt is entirely reasonable. The training data that structure prediction models consume is the library of known protein structures. If a model has merely memorized what it has seen, then no matter how high its benchmark score is, it is useless: when it actually reaches the front lines of drug development, it faces entirely novel molecules it has never seen before.
The industry has a brutal test specifically designed to catch "rote memorization," called Runs N' Poses. It divides test samples into 8 groups by similarity to the training set, from "looks similar" to "completely unfamiliar," peeling back the model's cover layer by layer.
In the lowest-similarity 0-20% group (63 ligands with "no relationship whatsoever" to the training set, equivalent to the completely out-of-syllabus questions on an exam), the models' scores are as follows:
AlphaFold 3 drops by half to 24%, while NovaAtom-Lite still holds steady at 47.6%, nearly double the former.
In the highest-similarity 80-100% group, the gap between the two instead narrows to 91.0% vs 88.2%.
This shows that the more unfamiliar the molecule, the greater NovaAtom-Lite's lead.
It is not reciting from a book; it has truly grasped the underlying physical laws of the microscopic atomic world and possesses a transferable "reasoning engine."
Give it more time to think, and it becomes smarter
Everyone working in AI knows a trend called Test-time Scaling—giving the model more compute during inference yields higher-quality answers.
This has long been validated in the large language model field, and NovaAtom-Lite proves that this rule holds equally true in biomolecular prediction.
On the same batch of 172 antibody-antigen interfaces, the team ran an elegant set of experiments:
Oracle upper bound (selecting the best solution after the fact using the experimental answer): 88.4%
From 74.3% to 83.1%—merely by increasing the number of samples, the success rate rose by nearly 9 percentage points.
This means the system has extremely strong "elasticity."
Startups have limited budgets; using a small amount of compute for rapid trial and error can still yield decent results. For pharmaceutical giants, when a targeted drug concerns a market worth tens of billions and the lives of countless patients, you can throw full compute at it and squeeze out every bit of precision—the model's ceiling is far from reached.
In real drug development scenarios, scientists often already have some clues. For example, they know roughly where the antibody and antigen make contact, or roughly which pocket of the protein the small molecule fits into.
The question is: can the model "understand" these hints and make good use of them?
NovaAtom-Lite's answer: not only does it understand, but the results are explosive.
On the 172 antibody-antigen interfaces in FoldBench, the success rate is 79.07% when no hints are given. Just give the model one contact residue pair (equivalent to telling it that these two residues are very close) and the success rate shoots straight up to 91.28%, a gain of 12 percentage points.
On another, larger-scale antibody-antigen test set, PXM-22to25-AB (383 clusters), the improvement is even more striking: from 66.05% all the way to 88.82%, a rise of nearly 23 percentage points.
Even more interesting is the protein-ligand direction.
On Runs N' Poses, the team tested four settings: no constraints, one contact pair, four pocket residues, and the full pocket, progressing layer by layer.
Hints help the hardest samples the most: in the 0-20% interval with the lowest similarity to the training set, giving the full-pocket hint can pull the success rate from 47.6% to 63.5%, a gain of nearly 16 percentage points; while in the high-similarity interval (above 60%), the gain is only 2 to 5 percentage points.
Simply put, the more "out-of-syllabus" the question, the more the model can "grasp" from the clues you give it. Scientists' experience is not wasted; the more you know, the more accurately it computes.
Two "god-tier" cases: from a molecular lasso to cracking a hidden door
No matter how beautiful the numbers are, what people in pharma ultimately want to see is: can you solve the thorniest structural problems in the real world?
Case one: taking a high-definition freeze-frame photo of a wildly thrashing "molecular lasso"
The Middle East respiratory syndrome (MERS) virus is a vicious opponent. Scientists found a nanobody, VHH-T148, to take it on; the core weapon of this antibody is a long CDR-H3 loop composed of 22 amino acids.
Imagine a "soft whip" thrashing wildly in three-dimensional space. It is extremely flexible, lacks evolutionary covariation signals, and is extremely difficult to predict.
This is a nightmare-level problem universally acknowledged in the structure prediction field.
As a result, NovaAtom-Lite generated 25 candidate solutions, all of which were hits. DockQ reached 0.839, and the prediction error for CDR-H3 was only 0.56 Å, covering every one of the 18 epitope residues.
0.56 Å is roughly one-third of the diameter of a carbon atom—this is a freeze-frame shot from a "molecular-level DSLR."
There is a class of inhibitors that work by burrowing into a "hidden pocket" (switch-II pocket) on KRAS, but the problem is that this pocket is normally locked.
In this case (structure ID 8UN5, KRAS G13D mutant), OpenDDE was completely stumped, with a ligand RMSD as high as 17.74 Å—the drug was flung far, far away.
In contrast, NovaAtom-Lite's ligand RMSD was only 0.98 Å.
Moreover, across all 25 samples, the predictions for the switch-II region differed from the crystal structure by only 0.4 to 0.6 Å, and all 12 native polar contacts were reproduced, with no steric clashes between any ligand atom and the protein.
Just three days after NovaAtom-Lite was released, top Silicon Valley venture capital firm a16z published a blockbuster research report—"Big Pharma Can Rival Big Software."
This report contains several core judgments worth savoring:
First, AI pharma has officially entered its "GPT-3 moment."
a16z's exact words were: we believe the current market maturity of Bio × AI is approaching the GPT-3 level. GPT-3 was the turning point at which general AI went from "usable" to "good to use." a16z believes similar capabilities are now emerging in biological models.
Second, pharma companies will in the future be priced like "tech stocks."
The valuation logic for traditional pharma companies is to do discounted cash flow on existing drug portfolios, and the drug R&D engine behind them is barely recognized by capital. But if AI makes drug discovery "highly repeatable and predictable," investors will be willing to pay for pharma companies' AI infrastructure—just as they pay for SaaS companies' platforms.
This is a fundamental reconstruction of valuation logic.
Third, NovaAtom-Lite, released on November 10, has already been included by a16z in this research report.
A model released only three days earlier was written into an industry report by top-tier capital—the speed of the response itself says a lot. This is real attention.
Even more intriguing is that in the industry-landscape illustration in the a16z report, the only company named is Geodesic Intelligence.
The report divides AI pharma startups into two categories: one is "vertically integrated" players like Isomorphic Labs and Recursion that go into drug development themselves, taking on high risk in pursuit of a trillion-dollar ceiling; the other is the "shovel sellers" that provide models, tools, and platforms: suppliers of lab automation, AI agents, and molecular design engines.
Geodesic is clearly placed in the most cutting-edge position within the latter category. Through the Geodesic API and the NovaDDE platform, NovaAtom-Lite's capabilities are being opened up to every pharma company and research team.
It is not yet known whether Geodesic Intelligence will go into drug development itself, but judging from the competitiveness its models display, it clearly has the opportunity to extend its boundaries from a model and tool supplier further into drug R&D itself.
a16z's judgment is not empty talk. Data from the industry side is already validating this trend.
GenScript's AI pharma business has doubled for three consecutive half-years, and Twist Bioscience's AI drug discovery orders have surged by triple digits.
Big pharma companies are rushing to sign deals. Isomorphic Labs alone has cooperation agreements with Novartis and Lilly totaling nearly $3 billion; Anthropic has opened its own wet lab, and OpenAI has directly acquired a biotech company.
The shortest path
Back to the story of Geodesic Intelligence.
Quanquan Gu is a professor in the Department of Computer Science at UCLA and is renowned in the field of AI for Science. In June 2026, he founded Geodesic Intelligence, with a goal that can be stated in one sentence:
Build AGI and find the shortest path from biology to drugs.
Just five months later, with an elite team of 15 people, he delivered his first answer.
And this is only the starting point. The company has clearly stated that NovaAtom-Lite is the first member of the NovaAtom model family, and larger, stronger models are already on the way.
Geodesic's full layout is a trinity closed loop: NovaAtom handles structure prediction, NovaDDE handles scientific reasoning and the drug discovery workbench, and NovaLab handles wet-lab validation—
from posing biological questions and computationally analyzing molecular structures to experimentally validating candidates, all connected end to end.
This brings to mind the line a16z wrote at the end of its report: developing drugs faster and at lower cost will be AI's most profoundly influential application direction.
The time accelerator for drug R&D has truly arrived.
This article comes from the WeChat public account "Xinzhiyuan"; author: ASI Revelation; editors: Aeneas Solomon; published by 36Kr with authorization.