Towards deep mechanistic neural networks for multi-omits modeling. Ameen Eetemadi Tagkopoulos Lab, UC Davis

Transcription

1 Towards deep mechanistic neural networks for multi-omits modeling. Ameen Eetemadi Tagkopoulos Lab, UC Davis

2 Outline Introduction Motivation Recent work Method (TRNN) Approach Architecture Optimization Results Conclusion

3 Movivation GOAL: Predict cellular state and behavior from environmental settings and genetic background of the cell. Genome-scale model Image credit: Erik Jacobsen, Covert Lab

4 Movivation GOAL: Predict cellular state and behavior from environmental settings and genetic background of the cell. Synthetic biology Genome-scale model Metabolic engineering Test gene circuit in host cell without experiment Image credit: Erik Jacobsen, Covert Lab

5 Movivation GOAL: Predict cellular state and behavior from environmental settings and genetic background of the cell. Synthetic biology Genome-scale model Metabolic engineering Test gene circuit in host cell without experiment Find environment that maximizes engineered products Image credit: Erik Jacobsen, Covert Lab

6 Movivation GOAL: Predict cellular state and behavior from environmental settings and genetic background of the cell. Synthetic biology Genome-scale model Metabolic engineering Test gene circuit in host cell without experiment Find environment that maximizes engineered products Image credit: Erik Jacobsen, Covert Lab Predictive medicine Optimization of antibiotic effect from ~100 existing antibiotics

7 Movivation GOAL: Predict cellular state and behavior from environmental settings and genetic background of the cell. Synthetic biology Genome-scale model Metabolic engineering Test gene circuit in host cell without experiment Find environment that maximizes engineered products Image credit: Erik Jacobsen, Covert Lab Predictive medicine Guided experimentation Optimization of antibiotic effect from ~100 existing antibiotics Find the conditions with highest uncertainty

8 Recent work (ECOMICS) - Data Most comprehensive compendium for an organism M. Kim, N. Rai, V. Zorraquino, and I. Tagkopoulos, Multi-omics integration accurately predicts cellular state in unexplored conditions for Escherichia coli, Nature communications (2016)

9 Recent work (ECOMICS) Predictive Model M. Kim, N. Rai, V. Zorraquino, and I. Tagkopoulos, Multi-omics integration accurately predicts cellular state in unexplored conditions for Escherichia coli, Nature communications (2016)

10 Recent work (ECOMICS) Predictive Model M. Kim, N. Rai, V. Zorraquino, and I. Tagkopoulos, Multi-omics integration accurately predicts cellular state in unexplored conditions for Escherichia coli, Nature communications (2016)

11 Outline Introduction Motivation Recent work Method (TRNN) Approach Architecture Training Results Conclusion

12 Approach Transcription Regulatory Neural Network (TRNN) How to Integrate: 1) Transcription regulatory network 2) Transcription Thermodynamics 3) Transcription profiles

13 Approach Transcription Regulatory Neural Network (TRNN) How to Integrate: 1) Transcription regulatory network 2) Transcription Thermodynamics 3) Transcription profiles

14 Approach Transcription Regulatory Neural Network (TRNN) How to Integrate: 1) Transcription regulatory network 2) Transcription Thermodynamics 3) Transcription profiles Understanding Sentences Words Alphabet

15 Architecture TRNN Activation Function (Alphabet) 1 Concentration of TFs 2 Computational Unit 3 mrna concentration (initial prediction) # $ 1 # ' 2!(# $,, # ' ) ) *(+,,), 4 4 Perturbation 5 Decision Unit 6 mrna concentration (adjusted prediction) Example:! # $, # - # $ + = h $ = h *(+,,) $ + z =, + # $6: : h - = h - (a), {0, 1} )

16 Architecture TRNN Gene Module (Word) Words + + +!(#) B C B $ #!(#) + # $ # -!(# $, # - ) B -

17 Architecture TRNN Gene Circuit (Sentence) Sentences (a) # ) B $ B $ + (b) # B $ B - + ) # $ ) $ # B - (c) B $ ) - (d) B C B $ ) $ ) E B $ ) - # D B $

18 Training Strategy Issue: backpropagation is very sensitive to initial values Strategy: Step1: train individual modules separately through a customized LP based method Step2: train full network together Use full-batch, conjugate gradient

19 Training Individual Modules Algorithm 1: Fit f, given C Input : C = {x 1, x 2, y} Output: = {t 0,t 1,t 2,b 1,b 2,p 1,p 2} 1 initialize p =[p 1,p 2]; 2 while loss (C) is not converged do 3 w ArgMin A.w y 1 1, s.t. b i 0; w i 4 hw p ; 5 for j 1:2do 6 j loss (C) ; 7 p j p j. p j ; 8 end i 9 hw p ; 10 end 11 return mx loss (C) = [f ( x i 1, x i 2 ) i= i i 4 5 x 1, i h 1 = e p 1 h 2 = e p 2 i x i y] apple A = i i 1 h 1 ( h 1 y) h2 ( h 2 y) : entry-wise vector multiplication,

20 Outline Introduction Motivation Recent work Method (TRNN) Approach Architecture Training Results Datasets Comparison Conclusion

21 Dataset 46 different datasets: Multiple candidate networks extracted from DREAM4 challenge Number of genes for each network range from 3 to 9. Synthetic data generated thermodynamic simulation Gene Knockouts Parameter perturbations (e.g. transcription rate, binding affinity of TFs, etc.) Different noise levels Validation Mean Squared Error 5-Fold cross validation Comparison with common ANN architectures

22 Results Overall MSE Comparison TRNN 0.03 MLP MLP MLP Average difference of MSE from best performing method in each dataset

23 Outline Introduction Motivation Recent work Method Approach Architecture Training Results Datasets Comparison Conclusion

24 Conclusion TRNN, is a deep learning based framework for prediction of steady state mrna concentration levels under perturbation. TRNN, requires specialized training due to it s architecture. Evaluation of TRNN on real data is key next step. For scaling up TRNN to genome scale modelling, we will make use of highly parallel computation provided by the cluster.