File size: 1,870 Bytes
7170613
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
---
license: apache-2.0
---
# Struct2Seq-GNN

## Model Description
Struct2Seq-GNN is a lightweight, 6-layer Graph Neural Network designed for inverse protein folding (structure-to-sequence prediction). By mapping the 3D spatial coordinates of protein backbones to their corresponding amino acid sequences, this model serves as a foundational tool for computational protein engineering and structural bioinformatics workflows. 

## Intended Uses & Limitations
* **Primary Use:** Computational protein design, generating plausible sequences for novel or heavily modified protein backbones.
* **Limitations:** This is a lightweight architecture built as an independent research project. While it achieves high native sequence recovery, it is not intended for out-of-the-box production of clinical therapeutics without further validation and optimization. 

## Training Data & Procedure
* **Dataset:** Trained on biological protein assemblies from the PDB, clustered at a 30% sequence identity cutoff to prevent data leakage.
* **Data Augmentation:** During training, 0.1 Å standard deviation Gaussian noise was applied to all input atomic coordinates. This critical augmentation prevents the model from "reading out" the native sequence from over-refined crystal artifacts, forcing it to learn the underlying biophysics of the fold.
* **Hardware:** Trained efficiently over ~65 epochs on a 4-GPU HPC cluster.

## Evaluation Metrics
The model demonstrates strong generalization and robust learning of physical constraints:
* **Global Sequence Recovery:** ~33% validation accuracy across all residues. (Achieving >30% sequence identity strongly suggests the generated sequences will reliably adopt the target 3D fold).
* **Convergence:** Validation loss plateaued smoothly at ~2.236.
* *(Optional: Add your 5.0 Å binding pocket recovery metric here once you calculate it!)*