Genie2: Diffusion Model for Complex and Multifunctional Protein Design

Genie2 is an improved diffusion model based on the Genie architecture that outperforms current known models on designability, diversity, and novelty, in single and multi-motif scaffolding.

1️⃣ Training Data Augmentation

Genie2 significantly expands its training dataset by using the AlphaFold database (AFDB), which includes over 200 million predicted protein structures. By employing a subsampled AFDB set, filtered for structural similarity and high-confidence predictions (pLDDT > 80), Genie2 utilizes 588,570 structures for training. This marks a substantial increase from the PDB's dataset and not only enriches the model with a broader variety of protein folds but also enhances its performance.

2️⃣ Unconditional Protein Generation

Genie2's capabilities in protein generation without length constraints outperform its predecessors up to its training limit of 256 aa. Beyond this, up to 500 aa, Genie2 demonstrates robust out-of-distribution generative abilities. It demonstrates commendable performance even beyond its training range. Comparatively, it surpasses Chroma and FrameFlow and is positioned as superior to RFDiffusion.

3️⃣ In-Distribution Performance

Genie2 was assessed alongside Chroma and RFDiffusion by generating five structures of every length ranging from 50 to 256 aa (1,035 structures in total). Relative to RFDiffusion and Chroma, Genie2 achieves comparable designability and much higher diversity and novelty, suggesting its ability to best capture foldable protein structure space.

4️⃣ Length-Based Generative Performance

Genie2 shows comparable designability to RFDiffusion but higher diversity in a subset of sequence lengths ranging from 50-500 aa at 100-generated structures. For short proteins (<200 aa), Genie2 exhibits significantly higher diversity, doubling that of RFDiffusion at 100 aa, which is remarkable as shorter lengths constitute smaller design spaces.

5️⃣ Motif Scaffolding (Single/Multi)

Across 24 single-motif scaffolding tasks, Genie2 solved 23 compared to RFDiffusion’s 22 tasks, with the number of unique successes being 1,445 and 889, respectively. In multi-motif scaffolding, Genie2 solved 4 out of 6 tasks. It also successfully solved the RSV G-protein 2D10 site-scaffolding task (5WN9), where RFDiffusion failed. Out of 24 problems, Genie2 failed on only one, similar to RFDiffusion, when scaffolding the RSV F-protein site-0.

6️⃣ Limitations

Genie2’s sampling time is longer, requiring 10x more denoising iterations (1,000) for sampling compared to FrameFlow (100), 2x more than Chroma (500), and 20x more than RFDiffusion (50) using a single A6000 GPU (48GB). Experimental validation of the designed proteins remains to be explored.

Open Questions and Future Directions:

1. Optimizing Training Data:

• How can the training dataset for Genie2 be further expanded or refined to enhance its predictive capabilities and performance?
• What impact would incorporating additional experimental data, alongside the AlphaFold database, have on Genie2’s accuracy and robustness?

2. Advancing Generative Capabilities:

• Can Genie2 be adapted to generate even longer protein sequences with high fidelity beyond its current out-of-distribution limit of 500 amino acids?
• How might Genie2 be improved to maintain or enhance its generative performance with fewer denoising iterations, reducing sampling time?

3. In-Distribution Performance Enhancements:

• What strategies can be employed to further increase the diversity and novelty of in-distribution protein structures generated by Genie2?
• How can Genie2’s designability metrics be improved to consistently surpass other state-of-the-art models across all sequence lengths?

4. Complex Motif Scaffolding:

• What modifications could be made to Genie2 to improve its success rate in multi-motif scaffolding tasks, particularly for more complex and novel motifs?
• How can Genie2 be adapted to solve challenging scaffolding tasks where current models, including RFDiffusion, fail?

5. Experimental Validation:

• What experimental approaches and collaborations could be pursued to validate the structures and functionalities of proteins designed by Genie2?
• How can feedback from experimental validations be integrated into Genie2’s training process to iteratively enhance its predictive accuracy?

6. Real-World Applications:

• In what ways can Genie2 be applied to practical problems in drug discovery, synthetic biology, and protein engineering?
• How might Genie2’s capabilities be leveraged to design multifunctional proteins with applications in industrial biotechnology and medicine?

7. Technical Improvements:

• How can the computational efficiency of Genie2 be enhanced to reduce the resource requirements for large-scale protein design projects?
• What novel machine learning techniques or algorithms could be incorporated into Genie2 to improve its overall performance and scalability?

By addressing these questions, researchers can further explore and expand the capabilities of Genie2, driving forward innovations in protein design and engineering. Continuous development and interdisciplinary collaboration will be key to realizing the full potential of Genie2 in various scientific and industrial domains.

Resources

Code: Genie 2: Designing and Scaffing Proteins at the Scale of the Structural Universe

Paper: Out of Many, One: Designing and Scaffolding Proteins at the Scale of the Structural Universe with Genie 2