update about page

.gitattributes

@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text

pages/About.py

@@ -1,12 +1,10 @@
 import streamlit as st
 import base64
 
-st.sidebar.markdown('''
-# Sections
-- [How it works](#how-it-works)
-- [Schematic overview of ProtHGT](#schematic-overview)
-''', unsafe_allow_html=True)
+st.set_page_config(page_title="About · ProtHGT", page_icon="🧬", layout="wide", initial_sidebar_state="collapsed")
 
+from ui import render_header
+render_header()
 
 st.markdown('''
 ## ProtHGT: Heterogeneous Graph Transformers for Automated Protein Function Prediction Using Knowledge Graphs and Language Models
@@ -17,7 +15,7 @@ st.markdown('''
 st.markdown(
     """
 
-[![github-repository](https://img.shields.io/badge/GitHub-black?logo=github)](https://github.com/HUBioDataLab/ProtHGT)
+[![github-repository](https://img.shields.io/badge/GitHub-black?logo=github)](https://github.com/HUBioDataLab/ProtHGT) [![huggingface-dataset](https://img.shields.io/badge/Hugging%20Face-Dataset-yellow?logo=huggingface)](https://huggingface.co/datasets/HUBioDataLab/ProtHGT)
 
 """)
 
@@ -27,11 +25,11 @@ st.subheader('How it works', anchor='how-it-works')
 
 st.markdown(
     """
-ProtHGT is a **heterogeneous graph transformer-based model** for automated protein function prediction. It integrates diverse biological data—proteins, pathways, domains, GO terms, and more—into **a unified knowledge graph**, enabling accurate and interpretable predictions.
+ProtHGT is a **heterogeneous graph transformer-based model** for automated protein function prediction. It integrates embeddings from **protein language models (PLMs)** with **a large-scale biological knowledge graph** built from 14 data sources, covering 9 biological entity types (proteins, domains, GO terms, pathways, diseases, drugs, and more) and 17 relationship types.
 
-Using transformer-based message passing, ProtHGT models complex biological relationships by propagating information across proteins and their functional associations in a structured graph. The model represents proteins using initial embeddings from **advanced protein language models (e.g., TAPE, ProtT5)** while integrating contextual information from pathways, domains, and molecular interactions. By employing knowledge **graph attention mechanisms**, ProtHGT learns to prioritize the most biologically relevant connections, improving prediction accuracy and interpretability.
+The model jointly learns from sequence-derived PLM embeddings and the multi-relational structure of the graph. Through its **graph attention mechanisms**, ProtHGT learns to prioritize the most influential connections in the knowledge graph, capturing the characteristics of each node and edge type and their specific contributions to protein function. ProtHGT is developed primarily for **Gene Ontology (GO) term prediction**, with dedicated models for all three sub-ontologies. To test whether the same architecture extends to a functional vocabulary with a different structure, it was additionally applied to **EC number prediction**, using a task-specific graph configuration derived from the same knowledge graph. This web application currently provides GO term predictions.
 
-ProtHGT outperforms existing sequence- and graph-based methods, as demonstrated in evaluations on DeepHGAT and PROBE datasets. By incorporating a broader biological context through its knowledge graph, **the model improves function prediction across all Gene Ontology (GO) sub-ontologies**. Additionally, its attention-based framework allows researchers to trace predictions back to key contributing relationships in the graph, making it possible to explore new functional links, validate known annotations, and generate testable biological hypotheses.
+ProtHGT outperforms state-of-the-art methods on the [DeepHGAT](https://ieeexplore.ieee.org/document/10423114/) and [PROBE](https://huggingface.co/spaces/HUBioDataLab/PROBE) benchmarks for GO term prediction, and ablation analyses highlight the **critical role of heterogeneous data integration** in its accuracy. Compared with the original PLM features, ProtHGT-refined embeddings show improved biological coherence. The learned representations and the architecture also generalize beyond GO term prediction: to semantic similarity inference and drug target protein family classification on PROBE, and to EC number prediction, where ProtHGT is competitive with dedicated enzyme function prediction methods on the [CLEAN](https://doi.org/10.1126/science.adf2465) benchmark. Finally, a use case study supports the biological validity of its predictions and shows how researchers can **trace predictions back to their underlying connections** in the graph, making it possible to explore new functional links, validate known annotations, and generate testable biological hypotheses.
 
 Overall workflow of ProtHGT is shown below.
 """)
@@ -40,8 +38,33 @@ st.subheader('Schematic overview of ProtHGT', anchor='schematic-overview')
 st.image('figures/ProtHGT_workflow.png')
 
 st.markdown(
-'<p style="text-align:center"><em><strong>Schematic representation of the ProtHGT framework. a)</strong> Diverse biological datasets, including proteins, pathways, domains, and GO terms, are integrated into a unified knowledge graph; <strong>b)</strong> the heterogeneous graph is constructed, capturing multi-relational biological associations; <strong>c)</strong> feature vectors for each node type are generated using state-of-the-art embedding methods; <strong>d)</strong> protein function prediction models are trained separately for molecular function, biological process, and cellular component sub-ontologies; <strong>e)</strong> heterogeneous graph transformer (HGT) layers process and refine node representations through multi-relational message passing. Final protein function predictions are obtained by linking proteins to GO terms based on learned embeddings and attention-weighted relationships.</em></p>', unsafe_allow_html=True)
+'<p style="text-align:center"><em><strong>Schematic representation of the ProtHGT framework. a)</strong> Biological data sources covering proteins and domains, functional terms and EC numbers, pathways, diseases and phenotypes, and drugs and compounds; <strong>b)</strong> the resulting heterogeneous knowledge graph, shown with its edge and node type counts (log scale), relation schema, and the embedding methods used to initialize each node type (vector size); <strong>c)</strong> prediction models. For GO term prediction, separate models are trained for molecular function (MF), biological process (BP), and cellular component (CC); final protein and GO term representations from the HGT encoder are concatenated and passed to a neural network that outputs the prediction probability. For EC number prediction, the graph is extended with the EC hierarchy and sequence similarity edges, and level-specific heads are trained jointly across EC levels L1–L4; <strong>d)</strong> the HGT encoder. Node features are linearly projected per type, then each HGT layer applies type-aware mutual attention and message passing with edge-type-specific weights across multiple heads, followed by target-specific aggregation, a nonlinear transformation, and a residual connection to produce updated node embeddings.</em></p>', unsafe_allow_html=True)
 
+st.subheader("Upcoming features", anchor="upcoming-features")
 st.markdown("""
+We are actively working on enhancing ProtHGT application with new capabilities:
 
-""")
+- **Real-time data retrieval for new proteins.** Currently, ProtHGT can only generate predictions for proteins that already exist in our knowledge graph. We are developing a new feature that will allow users to **predict functions for entirely new proteins starting from their sequences**. This will work by **retrieving relevant relationship data in real time from external source databases** (e.g., UniProt, STRING, and other biological repositories). The system will dynamically construct a knowledge graph for the query protein, incorporating its interactions, domains, pathways, and other biological associations before running function prediction. This approach will enable ProtHGT to analyze newly discovered or less-studied proteins even if they are not pre-annotated in our dataset.
+- **Expanded embedding options.** Currently, this application represents proteins using **TAPE embeddings**, which serve as the initial numerical representations of protein sequences before being processed in the heterogeneous graph model. We are working on integrating **ProtT5** and **ESM-2** as alternative initial embeddings, allowing users to choose different sequence representations that may enhance performance for specific tasks. A detailed comparison of how these embeddings influence function prediction accuracy will be included in our upcoming publication.
+
+**Recently added**
+
+- **Knowledge graph visualization for interpretability.** Explore the biological relationships behind ProtHGT's predictions interactively. For each query protein, you can inspect protein interactions, GO annotations, domains, pathways, and other key connections in a structured graphical format. Navigate to the **Knowledge graph** tab after generating predictions to try it out.
+
+Stay tuned for updates and future publications!
+""")
+
+st.subheader("Citation", anchor="citation")
+st.markdown("""
+If you use the ProtHGT method or the datasets provided in this repository, please cite:
+
+Ulusoy, E., & Dogan, T. (2025). ProtHGT: Heterogeneous Graph Transformers for Automated Protein Function Prediction Using Biological Knowledge Graphs and Language Models (p. 2025.04.19.649272). bioRxiv. [https://doi.org/10.1101/2025.04.19.649272](https://doi.org/10.1101/2025.04.19.649272)
+""")
+st.code("""@article{ulusoy2025prothgt,
+  title   = {ProtHGT: Heterogeneous Graph Transformers for Automated Protein Function Prediction Using Biological Knowledge Graphs and Language Models},
+  author  = {Ulusoy, Erva and Dogan, Tunca},
+  journal = {bioRxiv},
+  pages   = {2025.04.19.649272},
+  year    = {2025},
+  doi     = {10.1101/2025.04.19.649272}
+}""", language="bibtex")