ESMAdam: a plug-and-play all-purpose protein ensemble generator 1. ESMAdam introduces a versatile framework for protein ensemble generation, bridging gaps in high-throughput applications. Leveraging the pretrained ESMFold model and Adam optimization, it excels in efficiency and adaptability compared to traditional methods like molecular dynamics simulations. 2. The method supports a wide range of tasks, including conditional ensemble generation, CG-to-all-atom backmapping, binding mode exploration for protein complexes, and reconstructing 3D structures from cryo-EM images. 3. A key innovation is ESMAdam's reliance on optimizing protein embeddings in a continuous space. Unlike deep generative models, ESMAdam requires no retraining, allowing it to accommodate diverse structural constraints with minimal computational overhead. 4. In protein complex studies, ESMAdam identified metastable binding states previously undetected by molecular dynamics simulations, highlighting its potential for novel discovery in protein-protein interactions. 5. Cryo-EM integration demonstrates ESMAdam's robustness under noisy conditions. It reconstructs conformational ensembles while retaining accuracy in secondary structure prediction, even with significant noise levels. 6. The model achieves Boltzmann-distributed conformations, a feat previously challenging without physics-intensive simulations. It enhances studies on protein dynamics, stability, and interactions. 7. Future directions include reducing computational demands via techniques like LoRA for broader accessibility. ESMAdam positions itself as a critical tool for advancing structural biology and computational protein design. 📜Paper: biorxiv.org/content/10.1101/… #ProteinDesign #AI #StructuralBiology #ComputationalBiology #ESMAdam

ECloudGen: Leveraging Electron Clouds as a Latent Variable to Scale Up Structure-based Molecular Design 1. Introducing ECloudGen, a novel generative model that incorporates quantum physics principles by using electron clouds as latent variables. This approach scales up structure-based molecular design by bridging sparse protein-ligand complex data with ligand-only datasets. 2. Key innovation: ECloudGen's use of electron clouds enhances chemical space exploration and introduces model-level interpretability. This enables superior molecular design, outperforming state-of-the-art methods in binding affinity and drug-like properties. 3. Addressing a critical challenge, the "Paradox of Sparse Chemical Space Generation," ECloudGen effectively combines theoretical insights and modern AI techniques, using latent diffusion and contrastive learning to map structured chemical spaces. 4. Benchmark achievements: ECloudGen demonstrated significant improvements in ligand binding efficiency (LBE), QED, and synthetic accessibility, with results surpassing ten baseline models across multiple datasets. 5. Multi-objective optimization: The model excels in designing molecules that meet complex real-world constraints, exemplified in a BRD4 inhibitor case study, which included key molecular properties and functional group preferences. 6. Enhanced interpretability: By leveraging electron clouds, ECloudGen offers insights into the decision-making process, enabling chemists to trace and validate molecular design pathways effectively. 7. Broad utility: The model supports single-objective and multi-objective molecular optimization, showcasing flexibility and scalability for various drug discovery applications. 8. Real-world impact: In a V2R antagonist design, ECloudGen delivered molecules with micromolar-level binding affinities, proving its practical relevance and potential for accelerating therapeutic development. 9. Future directions: The methodology sets a precedent for integrating physical insights into deep learning frameworks, opening possibilities for applications beyond drug discovery, including material science and synthetic chemistry. 📜Paper: biorxiv.org/content/10.1101/… #DrugDiscovery #QuantumChemistry #AIinBiology #GenerativeModels #ComputationalBiology