3D CN-LSTM for prediction of medical nanoparticle properties
Sustar, Vid (2021)
Sustar, Vid
2021
All rights reserved. This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi-fe2021052831980
https://urn.fi/URN:NBN:fi-fe2021052831980
Tiivistelmä
Cancer is an increasing and already one of the most common causes of deathin developed countries. One way to fight cancer tumours is with targeted anti-tumour drug delivering nanoparticles (NPs). NPs can be composed of goldcore covered with a variety of drug and supporting substances (SS) in varyingratios.
Since chemical synthesis of all potential NPs is costly, to find the mostoptimal drug and drug-SS ratios out of many potential candidates, NPs aresimulated in silico in molecular dynamics (MD) simulations. To further lowerthe costs and expand coverage of potential optimal NP compositions, compu-tationally demanding MD simulations of NPs could in part be replaced withDeep Learning (DL) neural networks. Here the properties of NPs at laterstages of MD simulation would be predicted with DL from NP properties fromstarting stages of MD simulations.
As MD simulations are time series and NPs simulated are 3D objects, onecan join two types of DL: recurrent neural networks (RNN) and convolutionalneural networks (CNN) to create a suitable DL network. The scope of thismaster’s thesis is running MD simulations, finding proper DL architecture forthe model, refining the input and assessing the predictions of the refined model.
The architecture giving the best prediction of NP drug exposure is a com-bination of concatenated 3D CNN for NP structure input and dense layers forother types of input fed into Long Short-Term Memory (LSTM) RNN.
Since chemical synthesis of all potential NPs is costly, to find the mostoptimal drug and drug-SS ratios out of many potential candidates, NPs aresimulated in silico in molecular dynamics (MD) simulations. To further lowerthe costs and expand coverage of potential optimal NP compositions, compu-tationally demanding MD simulations of NPs could in part be replaced withDeep Learning (DL) neural networks. Here the properties of NPs at laterstages of MD simulation would be predicted with DL from NP properties fromstarting stages of MD simulations.
As MD simulations are time series and NPs simulated are 3D objects, onecan join two types of DL: recurrent neural networks (RNN) and convolutionalneural networks (CNN) to create a suitable DL network. The scope of thismaster’s thesis is running MD simulations, finding proper DL architecture forthe model, refining the input and assessing the predictions of the refined model.
The architecture giving the best prediction of NP drug exposure is a com-bination of concatenated 3D CNN for NP structure input and dense layers forother types of input fed into Long Short-Term Memory (LSTM) RNN.