Update README.md

README.md

@@ -8,36 +8,68 @@ tags:
 - aphasia
 - dna-regression
 - nvidia-blackwell
+- neuroplasticity
 datasets:
 - ARC-Aphasia-Genomics
 metrics:
 - mse
+- mae
 ---
 
-# Aphasia Recovery Potential Model (Genomic Modality v1.0)
+# 🧬 Genomic-Transformer-Aphasia-Recovery (v1.0)
+### *Predicting Clinical Neuroplasticity via Deep Genomic Regression*
 
 [![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/assix-research/stroke-recovery-analyser)
 
-## 🚀 Live Demo
-**Interactive Analysis:** Test genomic sequences and view recovery forecasts at the [Stroke-Recovery-Analyser](https://huggingface.co/spaces/assix-research/stroke-recovery-analyser).
+---
+
+## 🚀 Project Overview
+This model is a specialized **Sequence-to-Scalar Transformer** designed to predict post-stroke recovery potential. By analyzing 1024-nucleotide windows of genomic data, the model estimates a patient's **Western Aphasia Battery-Aphasia Quotient (WAB-AQ)**, providing a biological "ceiling" for language recovery.
+
+The model serves as the **Biological Modality** in a multi-modal fusion pipeline, intended to be integrated with MRI-based lesion volumetry to create personalized 3D recovery forecasts.
+
+
+
+## 🛠 Technical Specifications
+- **Base Backbone:** `InstaDeepAI/nucleotide-transformer-v2-500m-multi-species`
+- **Architecture Head:** Custom Linear Regression Head for continuous score prediction.
+- **Input Context:** 1024 base pairs (bp) tokenized at the single-nucleotide level.
+- **Precision:** BF16 (BFloat16) Mixed-Precision Training.
+- **Hardware:** Fine-tuned on the **NVIDIA DGX Spark** (Grace Blackwell GB10 Architecture).
+
+### Training Metadata
+| Parameter | Value |
+| :--- | :--- |
+| **GPU Architecture** | NVIDIA Blackwell (GB10) |
+| **Unified Memory** | 128GB LPDDR5x |
+| **Bandwidth** | 273 GB/s |
+| **Batch Size** | 16 (Gradient Accumulation = 4) |
+| **Optimizer** | AdamW (Weight Decay = 0.01) |
+
+## 📊 Dataset: ARC-Aphasia
+The model was fine-tuned on a curated genomic subset of the **Aphasia Recovery Cohort (ARC)**, consisting of 902 subjects. The training focus was localized to genetic regions associated with neural growth factors and neurotransmitter regulation, including:
+- **BDNF** (Brain-Derived Neurotrophic Factor)
+- **COMT** (Catechol-O-methyltransferase)
+- **APOE** (Apolipoprotein E)
+
+## 🖥️ Usage & Inference
+To run this model locally on your DGX or a compatible Blackwell environment:
 
-## Overview
-This model is a fine-tuned version of **InstaDeepAI/nucleotide-transformer-v2-500m-multi-species**, specifically adapted for **clinical regression tasks** in neuro-rehabilitation. It predicts the biological plasticity potential of a patient following a stroke, quantified by the **Western Aphasia Battery-Aphasia Quotient (WAB-AQ)** score.
+```python
+from transformers import AutoTokenizer, AutoModelForSequenceClassification
+import torch
 
-## Technical Architecture
-- **Base Architecture:** Transformer-based Nucleotide Encoder (500M Parameters).
-- **Modification:** Replaced the Language Modeling (LM) head with a specialized **Sequence Classification Head**.
-- **Training Objective:** Mean Squared Error (MSE) loss optimization for predicting WAB-AQ scores (Range: 0-100).
-- **Input:** 1024-nucleotide genomic sequences (focused on neuroplasticity-related markers such as BDNF and COMT).
+model_id = "assix-research/genomic-transformer-aphasia-recovery"
+tokenizer = AutoTokenizer.from_pretrained(model_id)
+model = AutoModelForSequenceClassification.from_pretrained(model_id, trust_remote_code=True)
 
-## Hardware & Performance
-Fine-tuned on the **NVIDIA DGX Spark** using the **Grace Blackwell (GB10)** architecture.
-- **Precision:** BF16 Mixed-Precision.
-- **Memory Bandwidth:** 273 GB/s via LPDDR5x Unified Memory.
-- **Optimization:** Gradient accumulation (steps=4) leveraging the 20-core Arm CPU for high-speed dataloading.
+dna_seq = "ATGC..." # 1024bp sequence
+inputs = tokenizer(dna_seq, return_tensors="pt")
+with torch.no_grad():
+    prediction = model(**inputs).logits.item()
 
-## Intended Use
-This model serves as the **Genomic Modality** in a multi-modal fusion pipeline. It is designed to be used alongside MRI lesion maps and DTI tractography to provide a holistic "Recovery Forecast" for aphasia patients.
+print(f"Predicted WAB-AQ: {prediction:.2f}/100")
+```
+⚠️ Clinical Disclaimer
 
-## ⚠️ Clinical Disclaimer
-This system is a biomedical research prototype. Predictions represent statistical probabilities based on a cohort of 902 stroke cases and are intended to assist clinical decision-making, not replace it.
+This model is a Biomedical Research Tool developed for the study of neuroplasticity. It is not an FDA-cleared diagnostic device. Predicted WAB-AQ scores are statistical estimates based on a specific research cohort and must be interpreted by a neurologist in conjunction with structural neuroimaging (MRI/DTI).