ICACT20250017 Slide.20
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-19.jpg) |
Thank you, we are happy to answer any questions about our PSF web application.
|
|
ICACT20250017 Slide.19
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-18.jpg) |
In conclusion, our PSF web application represents a significant advancement in protein scaffold filling technology with several key achievements:
Successfully integrating cutting-edge AI models into a user-friendly platform
Demonstrating exceptional accuracy in gap filling and sequence generation
Creating a flexible system that accommodates both pre-trained model usage and custom model training
The impact on proteomics includes simplifying scaffold gap filling and providing reliable solutions for advancing research in protein function prediction and drug discovery.
While we acknowledge limitations related to computational demands and dependence on high-quality input sequences, our future directions include expanding AI capabilities for diverse protein families and optimizing scalability through parallel computing.
PSF sets a new standard for protein scaffold filling tools, bridging the gap between advanced AI research and practical needs in computational proteomics.
|
|
ICACT20250017 Slide.18
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-17.jpg) |
Our analysis revealed several key insights:
GPT-2 Model: Achieved highest performance across tasks with 100% accuracy in both gap filling and full sequence generation
Transformer Model: Demonstrated balanced performance with strong generalization capability (96.42% gap filling, 96.72% full sequence accuracy)
CDA Model: Proved effective for gap filling with competitive accuracy (100% gap filling, 95.33% full sequence accuracy)
Based on these findings, we recommend:
For high-accuracy gap filling: GPT-2 or CDA models
For complete sequence generation: GPT-2 or Transformer models
For resource-constrained scenarios: Random Forest or CNN-LSTM approaches
|
|
ICACT20250017 Slide.17
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-16.jpg) |
We evaluated our models using two primary metrics:
Gap Filling Accuracy (GFA):
Measures correctness in filling amino acid gaps
Calculated as: GFA = Correct Predictions / Total Gaps
Full Sequence Accuracy (FSA):
Evaluates the accuracy of entire sequence generation
Calculated as: FSA = Correct Predictions / Total Sequence
The figure shows the Mabcampath scaffold we used for testing, highlighting the gaps in the sequence that need to be filled.
|
|
ICACT20250017 Slide.16
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-15.jpg) |
The custom model training functionality extends PSF's utility beyond pre-existing datasets, enabling researchers to train new models on novel protein scaffolds.
The workflow involves:
Uploading scaffold sequences or generating homologous sequences
System preprocessing of input data
Model training with user-selected AI algorithms
Result delivery with performance metrics
Key features include NCBI Protein BLAST API integration for generating homologous sequences and our task queue system for handling multiple training jobs asynchronously.
|
|
ICACT20250017 Slide.15
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-14.jpg) |
Our pre-trained model inference feature allows researchers to utilize models trained on the Mabcampath scaffold dataset for immediate gap filling tasks.
The user interface provides:
Simple input format with dropdown model selection
Direct display of scaffold-filling results on the interface
In our evaluation, these pre-trained models demonstrated high accuracy in gap filling tasks while significantly reducing computational time for known scaffolds.
|
|
ICACT20250017 Slide.14
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-13.jpg) |
This sequence diagram illustrates the interactions between users, our PSF interface, and various backend components.
User inputs include scaffold sequences in FASTA format and options to use pre-trained models or train new ones.
The system workflow follows these steps:
User uploads scaffold data
The system preprocesses and validates the input
The user selects a pre-trained model or opts to train a custom model
The system fills scaffold gaps using AI models
Results are delivered via the web interface or email notification
|
|
ICACT20250017 Slide.13
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-12.jpg) |
Our core algorithm unifies training and prediction in a single framework:
Training Mode: Trains models using user-provided or pre-existing data
Prediction Mode: Fills gaps in scaffolds based on trained models
The process follows these steps:
Preprocessing the input data
Model initialization or loading
Training or prediction operations
Postprocessing the results
We've also implemented automated features like a task queue for handling multiple training jobs asynchronously and email notifications to send scaffold results directly to users upon job completion.
|
|
ICACT20250017 Slide.12
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-11.jpg) |
Our methodology encompasses sophisticated data preprocessing techniques:
Converting protein sequences into numerical representations suitable for machine learning
Using techniques like label encoding of amino acids, dimensionality reduction with SVD, and row averaging for pattern extraction
We implement various AI models:
Machine Learning Algorithms: KNN, Decision Trees, Random Forest - lightweight options suitable for small-scale datasets
Deep Learning Architectures: CNN-LSTM combines convolutional networks and sequence modeling; Convolutional Denoising Autoencoder fills gaps and corrects errors; Transformer and GPT-2 models capture long-range dependencies for full-sequence predictions
|
|
ICACT20250017 Slide.11
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-10.jpg) |
The architecture also includes:
Data Management: Using JSON-based NoSQL storage for scaffolds and user inputs.
Task Queue System: Managing concurrent model training requests and ensuring system responsiveness for multiple users.
External Integration: Our application integrates with the NCBI Protein BLAST API to generate homologous sequences for training, automating data retrieval and preprocessing.
|
|
ICACT20250017 Slide.10
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-9.jpg) |
Our architecture consists of several components:
Frontend Design: Built using HTML, CSS, and JavaScript for a responsive and intuitive user interface with real-time validation for user inputs and multiple model options.
Backend Framework: Developed with Django, following the MVC architectural pattern, handling user requests and connecting the frontend to our AI models.
Server Infrastructure: Using Nginx as the web server for high-performance load balancing and static content delivery, and Gunicorn as the application server for efficient HTTP request processing.
|
|
ICACT20250017 Slide.09
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-8.jpg) |
Our PSF web application is developed using the Django framework for web accessibility with Python programming. It has two core functionalities:
Utilizing pre-trained models for immediate protein scaffold filling
Training custom models on user-provided data
The application is accessible via a web interface at https://psf.ncat.edu, featuring seamless integration of AI-driven techniques.
The figure illustrates our system architecture, showing the workflow from the web client interface through the server components to the AI models and NCBI Protein Blast API integration.
|
|
ICACT20250017 Slide.08
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-7.jpg) |
Recent research has made significant progress in this area:
In 2022, Sturtz and colleagues achieved high accuracy using CNN and LSTM models on the MabCampath scaffold dataset. Their k-mer sequence-based approach has limitations for large gaps or scaffolds with errors.
In 2023, Sturtz et al. introduced a convolutional denoising autoencoder model achieving 100% gap filling accuracy and 95.32% full sequence prediction without an iterative process.
In 2024, Badal et al. achieved 100% gap filling accuracy using different machine learning algorithms on the Mabcampath scaffold data.
Most recently, Qingge et al. developed generative AI models including transformer and GPT-2 architectures for protein scaffold filling, which outperform previous models with very high accuracy.
|
|
ICACT20250017 Slide.07
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-6.jpg) |
Now, we formally define the protein scaffold filling problem:
Given a target protein sequence T (which is unknown) and a protein scaffold S with m contigs (fragments) labeled S1 through Sm, our objective is to fill the missing amino acids between these contigs to obtain a completed sequence S' that maximizes correctly matched amino acids with the target sequence.
As shown in the diagram, we have multiple contigs (S1, S2, S3, S4) with gaps between them. Our challenge is to predict the sequences in these gaps to complete the scaffold.
|
|
ICACT20250017 Slide.06
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-5.jpg) |
This brings us to our research focus. De novo protein sequencing often results in incomplete sequences, what we call protein scaffolds, with gaps that are particularly difficult to predict.
The significance of this problem spans multiple areas:
Completing Protein Sequences: Filling gaps in protein scaffolds is crucial for determining complete protein sequences, especially in proteomics where de novo sequencing is used.
Enhancing Protein Function Understanding: Accurate scaffold gap filling helps in understanding protein structure and function.
Drug Discovery Applications: Properly filling protein scaffolds can aid in discovering new drugs by providing more accurate models of protein-protein interactions.
Our focus is developing PSF as a comprehensive and user-friendly web-based tool that empowers researchers to leverage artificial intelligence for protein scaffold filling.
|
|
ICACT20250017 Slide.05
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-4.jpg) |
Two principal methods dominate protein sequencing historically:
Edman Degradation, developed in the 1950s, was the pioneering technique for protein sequencing. It works by sequentially removing amino acids one at a time from the peptide chain while leaving the other amino acid bonds intact. Its drawbacks include slow speed and the need for substantial sample quantities.
Mass Spectrometry has gained popularity due to technological advancements, offering higher throughput and sensitivity while requiring less sample material. However, piecing together the complete protein sequence from mass spectra remains a challenging task.
|
|
ICACT20250017 Slide.04
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-3.jpg) |
Protein sequencing helps in determining the precise amino acid code, which is critical for understanding protein structures and functions, with wide-ranging applications across biology, chemistry, and medicine.
There are two main approaches to protein sequencing:
De novo protein sequencing, which involves elucidating the complete amino acid sequence from scratch without prior knowledge
Database searching, which compares experimentally obtained mass spectra to reference libraries to find matches and infer the likely sequence
Both approaches have their advantages and challenges in determining the complete protein sequence.
|
|
ICACT20250017 Slide.03
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-2.jpg) |
We would like to begin with some foundational concepts. Proteins are composed of chains of smaller units called peptides, which themselves consist of linked amino acid building blocks. Nature employs 20 different amino acids with molecular masses ranging from approximately 57 to 237 Daltons to construct all peptides and proteins.
The image on the right shows an example of an amino acid's structural diagram, with the central carbon atom connected to an amino group (N), a carboxyl group (C-O), a hydrogen atom (H), and a variable side chain (R) that determines the specific amino acid type.
|
|
ICACT20250017 Slide.02
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-1.jpg) |
Our presentation today covers seven key areas:
Introduction to the protein scaffold filling challenge
Detailed explanation of the protein scaffold gap filling problem
Overview of our PSF system
Our methodology and algorithms
Results and discussion
Comparative analysis
Conclusions
|
|
ICACT20250017 Slide.01
[Big slide for presentation]
[YouTube] |
Chrome Text-to-Speach Click!! |
-images-0.jpg) |
This is our presentation on "PSF: A Web Application Tool for Protein Scaffold Filling," developed by our team at North Carolina Agricultural and Technical State University. Our research team consists of Kushal Badal, Binhai Zhu, Xiaowen Liu, and Letu Qingge, representing a collaboration across multiple institutions including North Carolina A&T State University, Montana State University, and Tulane University.
|
