📄 bioengineering

Programming Biomolecular Interactions with All-Atom Generative Model

The paper introduces AnewOmni, a unified all-atom generative framework trained on over 5 million biomolecular complexes that utilizes programmable graph prompts to successfully design functional small molecules, peptides, and nanobodies across diverse molecular modalities, thereby establishing a foundation for general molecular reasoning in regimes where data and intuition are limited.

Original authors: Kong, X., Chen, J., Zhang, Z., Li, G., Zhu, Q., Wei, L., Li, M., Shi, Y., Dai, W., Zhang, Z., Tan, W., Jiao, R., Wang, X., Zheng, J., Yu, Z., Wu, Q., Guo, Z., Zhang, L., Li, W., Huang, Q., Zhu, T., Wa

Published 2026-03-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kong, X., Chen, J., Zhang, Z., Li, G., Zhu, Q., Wei, L., Li, M., Shi, Y., Dai, W., Zhang, Z., Tan, W., Jiao, R., Wang, X., Zheng, J., Yu, Z., Wu, Q., Guo, Z., Zhang, L., Li, W., Huang, Q., Zhu, T., Wang, X., Huang, W., She, Y., Zhang, J., Liu, Y., Liu, K., Ma, J.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are a master architect trying to build a custom key that fits a very specific, complex lock. In the world of biology, these "locks" are proteins inside our bodies that cause diseases, and the "keys" are molecules (like drugs) that can stop them.

For decades, scientists have built these keys using different toolkits for different sizes:

Small keys (tiny chemical molecules) were built by chemists using one set of rules.
Medium keys (peptides, which are short chains of amino acids) were built by biologists using another.
Huge keys (antibodies or nanobodies) were built by immunologists using a third.

The problem? These toolkits didn't talk to each other. They ignored the fact that, deep down, all keys work on the same physical principles: they need to fit the shape of the lock and stick to it chemically.

Enter AnewOmni, a new "Universal Key Architect" created by researchers at Tsinghua University and ByteDance. Here is how it works, explained simply:

1. The "LEGO" Brain (The Core Idea)

Imagine you have a massive box of LEGO bricks. Some are tiny single studs, some are 2x4 bricks, and some are pre-built wheels or windows.

Old way: You had separate instruction manuals for building cars, houses, and spaceships. You couldn't easily use a car wheel to fix a spaceship.
AnewOmni's way: It looks at everything—tiny chemicals, short protein chains, and huge antibodies—as just different arrangements of the same fundamental "atomic LEGO bricks."

The AI was trained on 5 million different examples of how these bricks fit together in nature. It learned that whether you are building a tiny molecule or a giant antibody, the rules of "stickiness" and "shape" are the same.

2. The "Magic Blueprint" (How it Designs)

Instead of guessing, AnewOmni uses a Generative Model. Think of it like a chef who has tasted millions of dishes and can now invent a new recipe from scratch that perfectly matches a specific flavor profile.

The Input: You tell the AI, "I need a key for this specific lock (a disease-causing protein)."
The Process: The AI doesn't just copy existing keys. It starts with a cloud of "potential atoms" and slowly sculpts them into a perfect shape that fits the lock.
The Secret Sauce: It uses a "Latent Space." Imagine a magical 3D map where similar shapes and chemical properties are grouped together. The AI navigates this map to find the perfect spot to build a new key, even if that key has never existed before.

3. The "Remote Control" (Programmable Prompts)

One of the coolest features is that you can give the AI specific instructions, like a remote control for the design process.

"Make it a loop!" (Cyclization): You can tell it to connect the ends of a peptide chain to make it a ring, making it stronger and more stable.
"Add a special ingredient!" (Non-canonical amino acids): You can ask it to insert a rare, custom chemical building block that nature doesn't usually use, to give the drug special powers.
"Grow from here!" (Scaffold growth): You can give it a small piece of a molecule and say, "Start here and grow the rest of the key to fit deeper into the lock."

4. The Real-World Test: The "Undruggable" Locks

To prove it works, the team tried to design keys for two notoriously difficult targets:

Target A: KRAS G12D (The "Undruggable" Cancer Target)
KRAS is a protein that causes many cancers. For years, scientists said it was "undruggable" because its surface is smooth and slippery, with no deep pockets for drugs to grab onto.

The Result: AnewOmni designed three different types of keys for the exact same spot: a tiny molecule, a medium peptide, and a large nanobody.
The Success: They didn't just copy existing drugs; they invented brand new shapes. In the lab, these new keys actually stuck to the cancer protein and stopped it, with success rates ranging from 23% to 75%. This is huge because usually, you need to test thousands of molecules to find one that works; they found winners with very few tries.

Target B: PCSK9 (The Cholesterol Regulator)
This protein controls cholesterol levels. The team wanted to stop it from doing its job.

The Result: They designed a peptide to block the main door (orthosteric) and a tiny molecule to jam a secret back door (allosteric).
The Success: The tiny molecule they designed fit so perfectly that when scientists took an X-ray picture of it, the real molecule matched the AI's drawing almost exactly (within the width of a single atom!). It successfully lowered cholesterol in cells by stopping the protein from being secreted.

Why This Matters

Before this, if you wanted to design a drug, you had to pick a "modality" (small molecule, peptide, or antibody) and stick to it. If that didn't work, you started over with a different toolkit.

AnewOmni changes the game. It is the first system that can seamlessly switch between tiny molecules and giant antibodies, using the same brain. It proves that nature's rules are universal.

The Bottom Line:
Think of AnewOmni as a universal translator and builder. It speaks the language of atoms fluently, regardless of whether you are building a pebble-sized drug or a boulder-sized antibody. It allows scientists to stop guessing and start programming biology, creating custom solutions for diseases that were previously considered impossible to treat.

1. Problem Statement

Biomolecular interactions are fundamental to cellular life, yet current design strategies for modulating these interactions are fragmented. Historically, small molecules, peptides, proteins (antibodies/nanobodies), and nucleic acids have been designed using distinct representations, rules, and protocols. This separation ignores the shared physicochemical principles of molecular recognition, which ultimately occur at the atomic scale.

Limitations of Current Methods: Existing generative models are typically modality-specific (e.g., separate models for small molecules vs. proteins). Some approaches use coarse-grained tokens (e.g., amino acids) which lose atomic detail, while others use purely atomic graphs but fail to capture higher-order structural priors (like conserved patterns in evolution).
The Gap: There is a lack of a unified, modality-agnostic framework that can transfer interaction principles across different molecular scales (from small chemical entities to large biologics) and support the design of hybrid or unconventional chemistries (e.g., non-canonical amino acids, macrocycles) without relying on high-throughput screening.

2. Methodology: AnewOmni

The authors present AnewOmni, a unified generative foundation model trained on over 5 million biomolecular complexes. The core innovation lies in its architecture and training strategy.

A. Unified Representation: Atom-to-Block Latent Space

Instead of treating molecules as separate modalities, AnewOmni decomposes all molecules into chemically coherent building blocks at atomic resolution:

Canonical Units: Amino acids and nucleobases are treated as single blocks.
Fragmented Units: Small molecules and non-canonical amino acids (ncAAs) are decomposed into frequent molecular fragments mined from databases (e.g., ChEMBL). Rare substructures are broken down to individual atoms.
Latent Encoding: An All-Atom Variational Autoencoder (VAE) maps these building blocks into a low-dimensional, continuous latent space. This preserves both atomic interaction geometries and structural priors.

B. Generative Process: Latent Diffusion

Diffusion Model: A latent diffusion model operates on the compressed point clouds of building blocks. It is conditioned on the target binding site's atomic environment.
Training: The model is trained jointly on a massive dataset comprising protein-small molecule, protein-peptide, and protein-antibody complexes. This joint training encourages the model to learn transferable interaction patterns across modalities.

C. Programmable Graph Prompts

A key feature is the ability to steer generation via programmable graph prompts. Users can inject constraints at the node (chemical identity) or edge (connectivity/geometry) level of the latent graph. This supports:

Topology Control: Enforcing cyclization (head-to-tail or disulfide bonds).
Chemical Composition: Incorporating specific motifs or ncAAs from user-defined libraries.
Geometric Constraints: Guiding anchor residues or covalent bonding to specific target sites.
Scaffold Growing: Extending existing molecules or linking fragments within a binding pocket.

3. Key Contributions

First Unified Foundation Model: AnewOmni is the first framework to successfully perform generative design across all molecular modalities (small molecules, peptides, antibodies/nanobodies) within a single atomic-resolution latent space.
Cross-Modality Transfer: The study demonstrates that joint training enhances performance for each modality compared to single-modality training, proving that shared physical principles of recognition can be learned and transferred.
Programmable Design: The introduction of graph prompts allows for fine-grained control over chemistry, topology, and geometry, enabling the design of complex structures (e.g., cyclic peptides, covalent inhibitors) without retraining.
Experimental Validation: The paper provides extensive experimental validation, moving beyond in silico benchmarks to wet-lab synthesis and testing of novel binders for challenging targets.

4. Key Results

A. In Silico Benchmarks

Performance: AnewOmni outperformed state-of-the-art modality-specific models on public benchmarks for small molecules, peptides, and antibodies.
Generalization: The model showed robust zero-shot generalization to unseen targets, including RNA, DNA, glycans, and phosphorylation sites, despite not being trained on RNA-small molecule complexes.
Physical Validity: Generated molecules maintained high physical validity even at molecular weights significantly larger than typical small molecules, a capability often lacking in other generative models.

B. Experimental Case Study 1: KRAS G12D

KRAS G12D is a notoriously "undruggable" target with a shallow binding pocket.

Small Molecules: Designed de novo inhibitors with novel scaffolds (Tanimoto similarity ~0.12–0.15 to known inhibitor MRTX1133). Two compounds showed moderate affinity ( $IC_{50}$ of 24 µM and 36 µM).
Peptides: Achieved success rates of 23% (linear) and 35% (cyclic) for $IC_{50} < 100$ µM. The best linear peptide had an $IC_{50}$ of 2.37 µM.
Nanobodies: Achieved a 75% success rate (3 out of 4 synthesized) with single-digit micromolar affinity. One nanobody showed a $K_d$ of 587 nM.
Significance: Demonstrated that a single framework can design binders of vastly different sizes for the same pocket.

C. Experimental Case Study 2: PCSK9

Orthosteric Peptides: Designed peptides targeting the protein-protein interaction (PPI) interface between PCSK9 and LDLR. Achieved a 57% hit rate, with the best binder having a $K_d$ of 3.19 µM.
Allosteric Small Molecules: Used AlphaFold3 to identify a cryptic allosteric pocket on the C-terminal domain of PCSK9. Designed small molecules targeting this site.
- Success: 3 out of 10 synthesized compounds bound with $K_d < 10$ µM (best $K_d$ = 2.72 µM).
- Mechanism: Cellular assays confirmed these compounds inhibit PCSK9 secretion, thereby upregulating LDLR.
- Structural Accuracy: A crystal structure of a designed binder (PCSK9-compound-3) confirmed the model's accuracy, with a Root Mean Square Deviation (RMSD) of 0.92 Å between the generated pose and the experimental structure.

5. Significance and Future Outlook

Paradigm Shift: AnewOmni challenges the historical divergence in molecular design, proposing that a unified, atomic-resolution approach is superior for capturing the fundamental physics of biomolecular recognition.
Beyond Data Limitations: The model's ability to generalize to unseen modalities (like RNA) and design unconventional chemistries suggests a path toward "general molecular reasoning engines" that can operate in data-scarce regimes.
Drug Discovery Impact: By bypassing the need for modality-specific high-throughput screening and enabling the design of diverse binder types (from small molecules to biologics) for the same target, AnewOmni offers a powerful tool for tackling "undruggable" targets and discovering novel mechanisms of action.

In conclusion, AnewOmni represents a significant step toward a general-purpose foundation model for biomolecular interactions, successfully unifying the design of small molecules, peptides, and biologics through a shared atomic latent space and programmable constraints.