Please enable JavaScript.

Coggle requires JavaScript to display documents.

Diploma - Coggle Diagram

- - - - Comparison with random pairs
      - Comparison with vienna
      - General: accuracy on val and test
    - - Heatmaps, comparison with mfold
        
        Model learned complementarity
        
        Correct predictions tend to be symmetrical
        
        Model learned that pairs tend to form duplexes
      - DMS-seq
        
        mRNAs were not used for training
        
        mRNAs accessibility were found in experiment
        
        mRNAs is another type of sequences (coding)
      - Features captured by attention
    - - Mention preliminary experiment with artificial sequences
  - - - Provide table of tertiary interactions in tRNA-His
      - Make understandable signs on the certainty plots
    - - Errors may be meaningful and can be generated by some features accounting for tertiary interactions
      - Discuss the tendency of prediction of duplexes again
  - - - Description
      - Results
    - - Description
      - Results
- - - - Catalitic activity
      - Белки типа самые важные, но так было не всегда (гипотеза РНК-мира) и сейчас есть рибозимы
      - Structural RNAs (rRNA, tRNA, tmRNA)
      - Regulatory elements which are mostly parts of mRNA
        
        Transcription
        
        terminators
        
        riboswitches
        
        Translation
        
        riboswithes
        
        both levels: attenuation
        
        secondary structure inhibits translation
        
        Splicing
        
        riboswithes if fungi
        
        pan-handles
        
        stability, editing and other processes
        
        viral issues
        
        imitation of host CAP or tRNA
        
        chromatin: long non-coding RNAs TOPHAT, Xist
        
        через запятую первый список (ядерные, ядрышковые)
    - - Description
        
        3-d
        
        2-d (major part of folding energy)
        
        elements (bulbs, hairpins and so on)
      - Process
        
        suboptimal
        
        kinetic (Mironov)
      - experimental data
        
        ideas of experiments
        
        accessibility
        
        crystallography
        
        data are expensive, there are plenty of genomes, we want to predict from sequence
        
        2D structure is good enough
        
        there is a standard approach based of physics of folding
        
        but they are not ideal and we don't know why
        
        so current methods imply limitations due to high computational cost of calculations
        
        NP-complete problem (with pseudoknots)
        
        DL is a modern approach to deal with such problems. There are much data so we want to try to enhance our understanding with DL approach
  - - - so it is easy to overfit data
      - to overcome overfitting people use regularisations of various types heavily limiting the search space for the parameters
      - learns features themselves and it is of specific interest for us
    - - features learned by NN
      - RNA structures
        
        not ideal
        
        still emerging
- - - - Encoding
      - Sourse
      - Splitting