DNA mutatsiooniliste signatuuride õppimine tehisnärvivõrkude abil
Files
Date
2016
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Kõik pahaloomulised vähkkasvajad on põhjustatud organismi rakkudes\n\rtoimuvate mutatsioonide poolt. On leitud, et need mutatsioonid\n\ron moodustatud spetsiifiliste mustrite ehk signatuuride kombinatsioonist,\n\rmille aluseks olevad protsessid on tihti teada. Seetõttu on nende\n\rsignatuuride õppimine andmetest väga tähtis — see võib anda paremat\n\rinformatsiooni vähkkasvajate mehhanismide kohta ja olla abiks\n\rvähi ennetamisel ja teraapial. Antud töö eesmärk on testida ja võrrelda\n\rerinevaid metoodikaid, et parandada mutatsiooniliste signatuuride\n\rleidmist. Me võrdlesime kolme uut meetodit — tehis-närvivõrgud\n\r(NN), mittenegatiivsed faktorvõrgud (RFN) ja teemade modelleerimine\n\r— praegu kasutatava mittenegatiivse maatriksi faktoriseerimisega\n\r(NMF). Me eksperimenteerisime meetoditega kolmel orgaanilisel ja kolmel\n\rsünteetilisel andmestikul ning mõõtsime rekonstrueerimise viga, tulemuse\n\rhõredust ja arvutusteks kulunud aega. Tulemused näitavad, et\n\rNMF annab väikseima veaga tulemuse kergematel andmestikel, kuid ka\n\rRFN-i tulemus on ligilähedane ning kõikidel teistel andmestikel saavutab\n\rsee parema tulemuse. NN esineb sama hästi kui RFN keerulisematel\n\randmestikel ning lisaks saavutab üleüldiselt kõige hõredamad tulemused.\n\rNMF-i eeliseks on stabiilsuse arvutamise funktsionaalsus, mis väga\n\rtäpselt suudab määrata õige signatuuride arvu. Tulevikus tuleb teha\n\redasist arendustööd, et sarnane võimekus ka RFN ja NN meetoditele\n\rlisada, mille järel oleks võimalik nende praktiline kasutamine mutatsiooniliste\n\rsignatuuride õppimisel.
All cancers are caused by mutations in the cells of an organism. It\n\ris found that these mutations result from the combination of specific\n\rmutational signatures, which often have known underlying processes.\n\rThat is why learning these signatures is important — it can give better\n\rinformation about the mechanisms of cancers and also be helpful for\n\rcancer prevention and therapy. The aim of this thesis is to test and\n\rcompare different methodology to improve the discovery of mutational\n\rsignatures. In particular, we compared three new methods of neural\n\rnetworks (NN), rectified factor networks (RFN) and topic modelling\n\rto the currently used non-negative matrix factorization (NMF). We\n\rexperimented with the methods on three organic and three synthetic\n\rdata sets by measuring reconstruction error, sparsity and time taken\n\rand compared them with NMF. The results show that NMF produces\n\rthe smallest error on easier data sets, but error of RFN is comparably\n\rgood also and on all other data sets produces the best result. NN\n\rperforms equally well with RFN on more difficult data sets and overall\n\rproduces the sparsest results. The advantage of NMF is the stability\n\rfunctionality that determines very well the correct number of signatures.\n\rFuture work will be needed to add this capability to RFN and\n\rNN methods which would enable their practical use for the problem of\n\rfinding mutational signatures.
All cancers are caused by mutations in the cells of an organism. It\n\ris found that these mutations result from the combination of specific\n\rmutational signatures, which often have known underlying processes.\n\rThat is why learning these signatures is important — it can give better\n\rinformation about the mechanisms of cancers and also be helpful for\n\rcancer prevention and therapy. The aim of this thesis is to test and\n\rcompare different methodology to improve the discovery of mutational\n\rsignatures. In particular, we compared three new methods of neural\n\rnetworks (NN), rectified factor networks (RFN) and topic modelling\n\rto the currently used non-negative matrix factorization (NMF). We\n\rexperimented with the methods on three organic and three synthetic\n\rdata sets by measuring reconstruction error, sparsity and time taken\n\rand compared them with NMF. The results show that NMF produces\n\rthe smallest error on easier data sets, but error of RFN is comparably\n\rgood also and on all other data sets produces the best result. NN\n\rperforms equally well with RFN on more difficult data sets and overall\n\rproduces the sparsest results. The advantage of NMF is the stability\n\rfunctionality that determines very well the correct number of signatures.\n\rFuture work will be needed to add this capability to RFN and\n\rNN methods which would enable their practical use for the problem of\n\rfinding mutational signatures.