nagam11 / pp2 Goto Github PK

• presentation is the most important!
• compare databases (does redundancy make a difference)
  • neural networks
  • neural translators
• change FFNN architecture
• UMAP embedding
  • architecture for 2D vectors
• bootstrapping + standard error
• add previous efforts and fails
• refine plotting

As Marla has mentioned, the repo is quite messy and it can become quite confusing. I have created a branch called clean_repo in which I remove all the unnecessary directories and files.

• remove biovec
• merge output and results (do we want to keep the results or are we creating new ones?)
• remove redundant scripts

1. new PPI data set
   • PiSITE (write a pisite2fasta parser)
   • constraints on PPI (info for stratification, only for simple FFNN)
2. embedding/dimensionality reduction
   • UMAP github paper
   • "Learned protein embeddings for machine learning" github paper
   • protein embedding with deep recurrent neural network paper github
   • GloVe
   • TSNE
3. model architecture
   • neural translator (LSTM + attention)
   • ResNet (residual network)
   • LSTM (directly from sequence), GRU (might be tricky)
   • attention learning (remembers more of the input)

Models

• simple FFNN + Word2Vec
• simple FFNN + UMAP
• neural translator + no extra embedding (default)

all models run on small and big datasets

How to visualise the training and predictions. This will be included in the presentation.

Training (Validation set)

• accuracy
• loss

Prediction Performance (Test set)

• accuracy, F1, Q2, ...
• ROC + AUC
• AUPR + AUPR
• Confusion matrix

all as barplots with standard errors

Create a new pipeline using your own architecture for protein prediction. Use word-embedding.

• new architecture

• new ML library

• parameters for word2vec

• different classifier

• new validation

• new plotting

Task

• provide and choose several (N) vector sizes
• extract features via biovec for each of the N vector sizes
• train model with cross-validation
• plot different metrics
• analyse and explain results

Plotting

• define input format
• training vs. validation metrics
• validation metrics across different vector sizes