Joeri Jongbloets bio photo

Joeri Jongbloets

Almost MSc // Modeller in the lab // Fluent in Java, Python, and R

Email Twitter

Evolution in prolonged turbidostat cultivation

Overview

Combined data of turbidostats 2015-05-28 and 2015-07-20

Overview

Time plot of the OD during the 2015-05-28 experiment

Time plot of the OD during the 2015-07-20 experiment

Clean data using interquartile rule (IQR) and limit time span to clean data regions

The cleaned Optical Density (720nm) measurements for experiment 28-05-2015

The cleaned Optical Density (720nm) measurements for experiment 20-07-2015

Combining the two experiments

The Optical density measurments of the two experiments combined

The number of decisions per 5 hours

Growth rate calculation

To calculate growth rates, different methods are available. I will use some of the available methods as I trust these the most (and I have a reliable pipeline for them).

Reliable methods (used in this document):

  • slopes: fits linear models to each decision cycle.
  • dxdt: calculates change in OD over time

Unreliable methods (not used in this document):

  • dilution: calculates dilution rate from dispense volume and estimated vessel volume.

This method is considered to be unreliable because of the huge variation in estimating the volume vessel which caused the dilution rate to vary a lot.

Calculting growth rate with slopes method

Method works by fitting one linear model each dilution cycle describing the following equation. The estimated coefficient (a), standard error of the estimation, the p-value of the estimate (a != 0) and the R squared of the fit, the number of points included in the decision cycle and the number of points used to fit the model are recorded. Each fit is attributed to the time point of the decision (t(D)) that ends to cycle.

log(OD) = a * t + c

Filter statistics

Plots below show the data and colour based on a selection metric. Goal is to determine selection criteria that reduce the spread of growth rates at each light regime. The median is indicated as the true value. Data is shown per experiment (columns) and channel (rows). Points with r squared below 0.5 are already removed.

Colour by r squared

Coloured by fit_size

Coloured by Slope Std. Error.

Coloured by standard deviation in OD values used to fit the data.

Cleaned data with filters

Remove points with R squared < 0.85 and fit size < 6.

Coloured by r squared (Scale changed!)

Calculting growth rate with dxdt method

The dxdt method calculates the growth rate using the formula given below.

mu = (dx) / (dt) = (log(ODmax) - log(ODmin)) / (tmax - tmin)

At least three growth rates are calculated from the data using combinations of points. The median and standard deviation of the collected slopes is returned. The following combinations are calculated (if enough data is available).

  1. t(min) vs t(max) (fit_size > 3)
  2. t(min+1) vs t(max) (fit_size > 4)
  3. t(min) vs t(max-1) (fit_size > 4)
  4. t(min+1) vs t(max-1) (fit_size > 5)
  5. t(min+2) vs t(max) (fit_size > 5)
  6. t(min) vs t(max-2) (fit_size > 5)
  7. t(min+2) vs t(max-1) (fit_size > 6)
  8. t(min+1) vs t(max-2) (fit_size > 6)
  9. t(min+2) vs t(max-2) (fit_size > 7)

Filter statistics

Plots below show the data and colour based on a selection metric. Goal is to determine selection criteria that reduce the spread of growth rates at each light regime. The median is indicated as the true value. Data is shown per experiment (columns) and channel (rows). Points with standard deviation above 0.75 are already removed.

Colour by std. dev of slopes

Coloured by fit size.

Colour by std. dev of OD

Cleaned data with filters

Remove points with Std. dev. of obtained sloped > 0.75 and a fit size < 6.

Merge methods

Analysis of growth data

Sliding window function

Sliding windows of different sizes (5, 10, 20, 50, 100) over the data to calculate the mean

Compare growth rates at begin to end

The following significance labels are used:

  • ns: P > 0.05
  • *: P <= 0.05
  • **: P <= 0.01
  • ***: P <= 0.001
  • ****: P <= 0.0001

Using t-tests

T-test first 200 against last 200 (i.e 50 - 250 hours vs 650 to 850 hours).

Using unaggregated data.

Using data from all window size per sliding window function (mean and median).

Using data per window size of the mean sliding window function.

Using data per window size of the median sliding window function.

Calculate relative growth rate

Using Linear models

Fitting linear model to all data (not aggregated using sliding window)

Fitting linear models using aggregated data from all sliding window sizes

Fitting linear models to means of sliding window

Fitting linear models to medians of sliding window

Calculate relative growth rate

Thesis pictures