Design of Ligand-Binding Proteins with Atomic Flow Matching (2409.12080v1)

Abstract: Designing novel proteins that bind to small molecules is a long-standing challenge in computational biology, with applications in developing catalysts, biosensors, and more. Current computational methods rely on the assumption that the binding pose of the target molecule is known, which is not always feasible, as conformations of novel targets are often unknown and tend to change upon binding. In this work, we formulate proteins and molecules as unified biotokens, and present AtomFlow, a novel deep generative model under the flow-matching framework for the design of ligand-binding proteins from the 2D target molecular graph alone. Operating on representative atoms of biotokens, AtomFlow captures the flexibility of ligands and generates ligand conformations and protein backbone structures iteratively. We consider the multi-scale nature of biotokens and demonstrate that AtomFlow can be effectively trained on a subset of structures from the Protein Data Bank, by matching flow vector field using an SE(3) equivariant structure prediction network. Experimental results show that our method can generate high fidelity ligand-binding proteins and achieve performance comparable to the state-of-the-art model RFDiffusionAA, while not requiring bound ligand structures. As a general framework, AtomFlow holds the potential to be applied to various biomolecule generation tasks in the future.

• The paper introduces AtomFlow, a deep generative model that designs ligand-binding proteins without relying on pre-bound ligand data.
• It employs a flow-matching framework with biotoken representations and SE(3) equivariant networks for iterative refinement of protein and ligand conformations.
• Experimental validation shows AtomFlow achieves competitive fidelity to state-of-the-art methods, expanding the capabilities of computational protein design.

Design of Ligand-Binding Proteins with Atomic Flow Matching

The paper presents a significant advancement in the field of computational biology by addressing the longstanding challenge of designing proteins capable of binding to small molecules. The newly introduced framework, AtomFlow, leverages deep generative modeling to innovate the protein design process, specifically targeting the creation of ligand-binding proteins without necessitating pre-existing data on the ligand's binding pose.

AtomFlow is built upon a flow-matching generative model that employs biotoken representations for both proteins and molecules. This allows for a unified approach to model generation, where proteins and ligands are represented by a set of representative atoms defined as biotokens. The flexibility of ligands, often a challenge due to their conformational changes upon binding, is adeptly captured in this framework. AtomFlow is trained using an SE(3) equivariant structure prediction network, which ensures that the model can effectively generate both ligand conformations and protein backbone structures through iterative refinement based on 2D molecular graphs.

The experimental validation of AtomFlow shows that it can generate ligand-binding proteins with fidelity comparable to state-of-the-art methods such as RFDiffusionAA. However, AtomFlow's distinct advantage lies in its ability to function without the necessity of bound ligand structures, thereby overcoming a critical limitation in existing methodologies that depend on known and static conformations. This capability is particularly relevant in scenarios where the target molecules do not bind to any known natural proteins, or where computational resources are constrained, making it impractical to generate exhaustive conformer databases.

By employing a conditional vector field derived from a rectified flow on the representative atoms, AtomFlow is able to optimally balance the structural features of the protein and ligand, ensuring high binding affinity and novel conformer generation. The implication of these findings extends beyond the immediate challenge of ligand-binding protein design, as the unified biotoken representation and flow-matching framework can be adapted for various biomolecule generation tasks.

The paper emphasizes the potential applicability of AtomFlow in broader biomolecular design tasks, suggesting future research directions that might include its extension to other biological molecules like nucleic acids and complex inorganic compounds. Additionally, the versatility of AtomFlow in addressing ligand flexibility and protein-ligand interactions directly may contribute to advancements in drug discovery, synthetic biology, and the design of novel biosensors and catalysts.

In summary, AtomFlow's introduction of a unified modeling framework for protein-ligand interaction, alongside the demonstration of its capabilities against existing benchmarks, positions it as a noteworthy contribution to computational approaches in molecular biology. The potential to reshape methodologies that tackle the intricacies of protein design is apparent and represents an exciting avenue for further exploration and application.