RINAMI: A Deep Learning-Based Model for Protein Stability Prediction

RINAMI is an open-source deep learning model for predicting protein folding free energy change (ΔG). It combines graph neural networks and the ProteinMPNN protein language model to provide computational support for fields such as protein engineering and drug design. Its open-source nature fosters collaboration in the scientific community and is expected to accelerate related research progress.

Protein stability directly affects the operation of organisms. ΔG is the core indicator for measuring stability (a positive value indicates that the mutation makes the protein more unstable, while a negative value means it is more stable). Accurate prediction of ΔG is crucial for understanding disease mechanisms, enzyme design, and therapeutic protein development. However, ΔG prediction faces multiple challenges: complex protein structures and non-linear amino acid interactions; scarce experimental data; large differences in characteristics among different protein families. Traditional physical simulations struggle to balance accuracy and efficiency, while pure statistical methods lack physical interpretability.

RINAMI adopts a hybrid architecture:

1. Graph Neural Network Layer: Treats amino acid residues as nodes and spatial proximity or chemical bonds as edges. It captures long-range residue interactions through multi-layer message passing and adapts to proteins of different sizes.
2. ProteinMPNN Feature Fusion: Integrates node representations and output features from ProteinMPNN, combining structural (GNN) and sequence evolution (ProteinMPNN) information to enhance prediction accuracy.
3. Multi-task Learning Framework: Jointly optimized on benchmark datasets such as Mega-scale, Maxwell, and Garcia to learn general representations and reduce the risk of overfitting.

• Environment Configuration: Based on Python, depends on PyTorch and PyTorch Geometric. NVIDIA GPU acceleration (e.g., RTX3080 + CUDA12.1) is recommended, and dependencies are managed via Conda.
• Data Preparation: Uses ESMFold to predict protein structures and generates ProteinMPNN features via scripts; preprocessed files for large-scale datasets can be downloaded from Zenodo.
• Training and Inference: Training scripts encapsulate and simplify the workflow. Inference supports PDB file input and ΔG heatmap generation, and a Colab notebook is provided for users without a GPU.

1. Directed Evolution Guidance: Pre-screen mutants to reduce experimental workload;
2. Disease Mutation Interpretation: Evaluate the pathogenic potential of mutations and provide clues for clinical diagnosis;
3. Protein Design Optimization: Assess the stability of design schemes and guide sequence optimization.

RINAMI's code (GitHub) and data (Zenodo) are both open-source, lowering the threshold for reproduction. Data hosted on Zenodo ensures traceability, and an interpretability analysis function (exporting ΔG contribution matrix) is provided to help understand the basis of the model's decisions.

Limitations: Insufficient training data coverage (accuracy decreases for protein families not present in the training set); only focuses on single-point mutations, with limited prediction for multi-point mutations and insertions/deletions. Future Directions: Expand training set coverage; introduce physical constraints to enhance interpretability; develop specialized models for membrane proteins, antibodies, etc.; optimize inference speed to support large-scale virtual screening.

RINAMI is a cutting-edge achievement of the integration of computational biology and deep learning, providing an efficient and easy-to-use tool for protein stability prediction. Its open-source nature promotes community collaboration and accelerates protein engineering research. With the progress of deep learning and the accumulation of experimental data, such tools will play a more important role in life sciences, driving comprehensive progress from basic research to application development.