# Artificial Intelligence Meets Marine Ecotoxicology: Applying Deep Learning to Bio-Optical Data from Marine Diatoms Exposed to Legacy and Emerging Contaminants

^{1}

^{2}

^{3}

^{4}

^{5}

^{*}

## Abstract

**:**

## Simple Summary

## Abstract

## 1. Introduction

**Random Forests:**This classifier is an ensemble algorithm that uses the majority vote of a set of decision trees to make predictions;**XGBoost:**This classifier is a state-of-the-art ensemble algorithm that uses a set of decision trees optimized with gradient boosting in order to minimize errors;**ROCKET:**A recent state-of-the-art time series classification algorithm that uses random convolution kernels to increase the dimensionality of a dataset and, together with a simple regression classifier (commonly Ridge);**Neural Networks:**Deep learning models based on the original multi layer perceptron. These networks have a large amount and variety of layers and are trained by means of gradient descent. Although for regular artificial neural networks (ANN) the feature engineering part is done a priori, convolutional neural networks (CNN) perform this step using convolutional and pooling layers. CNNs can be used for multi-dimensional data, so they can process 1D, 2D, or 3D.

## 2. Materials and Methods

#### 2.1. Ecotoxicological Assays

#### 2.2. Methods

#### 2.3. Hardware and Software

#### 2.4. Architecture and Parameters

#### 2.5. Experimental Setup

## 3. Results

## 4. Discussion

## 5. Conclusions and Future Work

## Author Contributions

## Funding

## Institutional Review Board Statement

## Informed Consent Statement

## Data Availability Statement

## Conflicts of Interest

## Appendix A

**Table A1.**Kruskal–Wallis p-values comparing all methods of the EC prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

EC | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $1.23{e}^{-10}$ | $2.58{e}^{-11}$ | $2.80{e}^{-11}$ | $1.89{e}^{-08}$ | $2.63{e}^{-11}$ | $2.68{e}^{-11}$ | Test |

XG | $2.24{e}^{-11}$ | ――― | $2.60{e}^{-11}$ | $2.82{e}^{-11}$ | $1.01{e}^{-07}$ | $2.65{e}^{-11}$ | $2.70{e}^{-11}$ | |

Rocket | $1.88{e}^{-11}$ | $1.08{e}^{-05}$ | ――― | $2.59{e}^{-11}$ | $4.04{e}^{-03}$ | $3.64{e}^{-02}$ | $5.79{e}^{-04}$ | |

ANN | $2.82{e}^{-11}$ | $2.26{e}^{-11}$ | $1.90{e}^{-11}$ | ――― | $7.04{e}^{-04}$ | $2.64{e}^{-11}$ | $2.69{e}^{-11}$ | |

1D CNN | $2.34{e}^{-11}$ | $8.89{e}^{-02}$ | $8.08{e}^{-01}$ | $2.36{e}^{-11}$ | ――― | $3.19{e}^{-04}$ | $2.68{e}^{-06}$ | |

CNN | $2.54{e}^{-11}$ | $2.03{e}^{-11}$ | $1.70{e}^{-11}$ | $2.56{e}^{-11}$ | $3.09{e}^{-08}$ | ――― | $6.93{e}^{-03}$ | |

CNN Log | $1.63{e}^{-09}$ | $2.07{e}^{-11}$ | $1.73{e}^{-11}$ | $2.61{e}^{-11}$ | $6.55{e}^{-09}$ | $2.55{e}^{-02}$ | ――― | |

Training |

**Table A2.**Kruskal–Wallis p-values comparing all methods of the DCF concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

EC | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $1.43{e}^{-01}$ | $1.88{e}^{-11}$ | $6.03{e}^{-11}$ | $1.43{e}^{-04}$ | $5.66{e}^{-07}$ | $5.66{e}^{-07}$ | Test |

XG | $5.32{e}^{-03}$ | ――― | $1.40{e}^{-10}$ | $1.89{e}^{-10}$ | $2.98{e}^{-05}$ | $1.73{e}^{-08}$ | $1.73{e}^{-08}$ | |

Rocket | $5.32{e}^{-03}$ | $1.00{e}^{+00}$ | ――― | $6.34{e}^{-01}$ | $1.91{e}^{-03}$ | $9.42{e}^{-08}$ | $9.42{e}^{-08}$ | |

ANN | $2.98{e}^{-01}$ | $1.36{e}^{-04}$ | $1.36{e}^{-04}$ | ――― | $5.28{e}^{-03}$ | $9.88{e}^{-07}$ | $9.88{e}^{-07}$ | |

1D CNN | $9.28{e}^{-01}$ | $5.34{e}^{-03}$ | $5.34{e}^{-03}$ | $2.08{e}^{-01}$ | ――― | $2.92{e}^{-01}$ | $2.92{e}^{-01}$ | |

CNN | $5.32{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $1.36{e}^{-04}$ | $5.34{e}^{-03}$ | ――― | $1.00{e}^{+00}$ | |

CNN Log | $5.32{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $1.36{e}^{-04}$ | $5.34{e}^{-03}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A3.**Kruskal–Wallis p-values comparing all methods of the FX concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

FX | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $3.69{e}^{-01}$ | $2.64{e}^{-06}$ | $1.65{e}^{-06}$ | $1.15{e}^{-02}$ | $1.94{e}^{-02}$ | $2.44{e}^{-01}$ | Test |

XG | $4.97{e}^{-01}$ | ――― | $1.03{e}^{-08}$ | $2.31{e}^{-09}$ | $1.00{e}^{-04}$ | $7.84{e}^{-04}$ | $7.65{e}^{-01}$ | |

Rocket | $6.16{e}^{-05}$ | $5.92{e}^{-04}$ | ――― | $8.63{e}^{-01}$ | $8.09{e}^{-03}$ | $4.35{e}^{-07}$ | $1.26{e}^{-09}$ | |

ANN | $6.54{e}^{-03}$ | $7.45{e}^{-06}$ | $4.94{e}^{-10}$ | ――― | $7.98{e}^{-03}$ | $2.64{e}^{-07}$ | $1.78{e}^{-10}$ | |

1D CNN | $4.13{e}^{-03}$ | $4.52{e}^{-02}$ | $4.02{e}^{-02}$ | $3.34{e}^{-08}$ | ――― | $2.77{e}^{-02}$ | $2.41{e}^{-05}$ | |

CNN | $6.16{e}^{-05}$ | $5.92{e}^{-04}$ | $1.00{e}^{+00}$ | $4.94{e}^{-10}$ | $4.02{e}^{-02}$ | ――― | $6.19{e}^{-08}$ | |

CNN Log | $6.16{e}^{-05}$ | $5.92{e}^{-04}$ | $1.00{e}^{+00}$ | $4.94{e}^{-10}$ | $4.02{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A4.**Kruskal–Wallis p-values comparing all methods of the GLIPH concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

GLIPH | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $2.84{e}^{-03}$ | $7.00{e}^{-06}$ | $8.98{e}^{-01}$ | $1.35{e}^{-02}$ | $8.08{e}^{-04}$ | $3.23{e}^{-01}$ | Test |

XG | $1.27{e}^{-03}$ | ――― | $2.80{e}^{-09}$ | $7.73{e}^{-03}$ | $3.12{e}^{-06}$ | $8.80{e}^{-01}$ | $6.39{e}^{-05}$ | |

Rocket | $1.27{e}^{-03}$ | $1.00{e}^{+00}$ | ――― | $3.38{e}^{-04}$ | $6.72{e}^{-02}$ | $3.69{e}^{-11}$ | $1.41{e}^{-04}$ | |

ANN | $2.65{e}^{-05}$ | $5.40{e}^{-10}$ | $5.40{e}^{-10}$ | ――― | $5.21{e}^{-02}$ | $1.81{e}^{-04}$ | $2.52{e}^{-01}$ | |

1D CNN | $8.34{e}^{-03}$ | $3.17{e}^{-01}$ | $3.17{e}^{-01}$ | $1.13{e}^{-08}$ | ――― | $5.40{e}^{-09}$ | $2.08{e}^{-01}$ | |

CNN | $1.27{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $5.40{e}^{-10}$ | $3.17{e}^{-01}$ | ――― | $2.49{e}^{-10}$ | |

CNN Log | $1.27{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $5.40{e}^{-10}$ | $3.17{e}^{-01}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A5.**Kruskal–Wallis p-values comparing all methods of the IBU concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

IBU | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $4.60{e}^{-03}$ | $1.93{e}^{-01}$ | $5.17{e}^{-04}$ | $1.87{e}^{-03}$ | $9.16{e}^{-01}$ | $5.92{e}^{-01}$ | Test |

XG | $4.27{e}^{-06}$ | ――― | $2.63{e}^{-05}$ | $1.24{e}^{-08}$ | $5.16{e}^{-08}$ | $4.61{e}^{-05}$ | $3.68{e}^{-05}$ | |

Rocket | $3.07{e}^{-04}$ | $4.38{e}^{-08}$ | ――― | $1.40{e}^{-02}$ | $3.97{e}^{-02}$ | $6.04{e}^{-03}$ | $7.58{e}^{-02}$ | |

ANN | $9.75{e}^{-01}$ | $2.88{e}^{-07}$ | $1.11{e}^{-03}$ | ――― | $3.54{e}^{-01}$ | $1.50{e}^{-08}$ | $5.54{e}^{-06}$ | |

1D CNN | $1.09{e}^{-03}$ | $4.02{e}^{-02}$ | $1.08{e}^{-06}$ | $2.07{e}^{-04}$ | ――― | $1.88{e}^{-07}$ | $7.81{e}^{-05}$ | |

CNN | $4.27{e}^{-06}$ | $1.00{e}^{+00}$ | $4.38{e}^{-08}$ | $2.88{e}^{-07}$ | $4.02{e}^{-02}$ | ――― | $6.10{e}^{-02}$ | |

CNN Log | $4.27{e}^{-06}$ | $1.00{e}^{+00}$ | $4.38{e}^{-08}$ | $2.88{e}^{-07}$ | $4.02{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A6.**Kruskal–Wallis p-values comparing all methods of the PROP concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

PROP | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $3.27{e}^{-01}$ | $4.22{e}^{-05}$ | $4.16{e}^{-02}$ | $2.46{e}^{-03}$ | $6.34{e}^{-01}$ | $1.63{e}^{-06}$ | Test |

XG | $5.32{e}^{-03}$ | ――― | $1.09{e}^{-03}$ | $3.62{e}^{-01}$ | $4.54{e}^{-02}$ | $1.71{e}^{-02}$ | $5.79{e}^{-05}$ | |

Rocket | $5.32{e}^{-03}$ | $1.00{e}^{+00}$ | ――― | $1.83{e}^{-02}$ | $2.16{e}^{-01}$ | $2.18{e}^{-10}$ | $2.71{e}^{-01}$ | |

ANN | $9.69{e}^{-01}$ | $2.65{e}^{-03}$ | $2.65{e}^{-03}$ | ――― | $3.78{e}^{-01}$ | $2.79{e}^{-03}$ | $1.72{e}^{-03}$ | |

1D CNN | $2.83{e}^{-01}$ | $4.02{e}^{-02}$ | $4.02{e}^{-02}$ | $2.19{e}^{-01}$ | ――― | $1.03{e}^{-04}$ | $2.72{e}^{-02}$ | |

CNN | $5.32{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $2.65{e}^{-03}$ | $4.02{e}^{-02}$ | ――― | $1.86{e}^{-11}$ | |

CNN Log | $5.32{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $2.65{e}^{-03}$ | $4.02{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A7.**Kruskal–Wallis p-values comparing all methods of the SDS concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

SDS | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $1.89{e}^{-03}$ | $4.23{e}^{-05}$ | $1.76{e}^{-08}$ | $9.15{e}^{-02}$ | $1.17{e}^{-01}$ | $4.13{e}^{-05}$ | Test |

XG | $6.28{e}^{-04}$ | ――― | $1.46{e}^{-09}$ | $8.01{e}^{-11}$ | $2.18{e}^{-05}$ | $2.25{e}^{-02}$ | $9.10{e}^{-01}$ | |

Rocket | $6.28{e}^{-04}$ | $1.00{e}^{+00}$ | ――― | $3.35{e}^{-03}$ | $1.46{e}^{-02}$ | $4.23{e}^{-10}$ | $1.40{e}^{-11}$ | |

ANN | $2.38{e}^{-04}$ | $5.00{e}^{-09}$ | $5.00{e}^{-09}$ | ――― | $5.19{e}^{-06}$ | $5.93{e}^{-11}$ | $1.40{e}^{-11}$ | |

1D CNN | $1.49{e}^{-02}$ | $1.54{e}^{-01}$ | $1.54{e}^{-01}$ | $2.66{e}^{-07}$ | ――― | $6.22{e}^{-04}$ | $2.38{e}^{-08}$ | |

CNN | $6.28{e}^{-04}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $5.00{e}^{-09}$ | $1.54{e}^{-01}$ | ――― | $2.62{e}^{-05}$ | |

CNN Log | $6.28{e}^{-04}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $5.00{e}^{-09}$ | $1.54{e}^{-01}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A8.**Kruskal–Wallis p-values comparing all methods of the TRI concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

TRI | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $7.03{e}^{-04}$ | $9.62{e}^{-11}$ | $4.89{e}^{-08}$ | $4.99{e}^{-07}$ | $4.22{e}^{-03}$ | $6.36{e}^{-02}$ | Test |

XG | $1.05{e}^{-02}$ | ――― | $1.67{e}^{-11}$ | $1.05{e}^{-10}$ | $1.32{e}^{-09}$ | $3.22{e}^{-08}$ | $7.14{e}^{-07}$ | |

Rocket | $1.05{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | $1.96{e}^{-03}$ | $6.46{e}^{-04}$ | $2.89{e}^{-12}$ | $6.91{e}^{-12}$ | |

ANN | $8.89{e}^{-06}$ | $1.59{e}^{-09}$ | $1.59{e}^{-09}$ | ――― | $5.67{e}^{-01}$ | $5.66{e}^{-08}$ | $6.83{e}^{-07}$ | |

1D CNN | $8.16{e}^{-01}$ | $1.06{e}^{-02}$ | $1.06{e}^{-02}$ | $2.67{e}^{-04}$ | ――― | $1.65{e}^{-09}$ | $3.99{e}^{-08}$ | |

CNN | $1.05{e}^{-02}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $1.59{e}^{-09}$ | $1.06{e}^{-02}$ | ――― | $7.93{e}^{-01}$ | |

CNN Log | $1.05{e}^{-02}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $1.59{e}^{-09}$ | $1.06{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A9.**Kruskal–Wallis p-values comparing all methods of the Cu_d concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

Cu_d | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $5.93{e}^{-01}$ | $6.08{e}^{-10}$ | $1.60{e}^{-03}$ | $4.64{e}^{-02}$ | $1.19{e}^{-01}$ | $2.90{e}^{-01}$ | Test |

XG | $5.97{e}^{-05}$ | ――― | $4.43{e}^{-11}$ | $1.53{e}^{-03}$ | $6.12{e}^{-02}$ | $1.63{e}^{-03}$ | $6.50{e}^{-01}$ | |

Rocket | $5.97{e}^{-05}$ | $1.00{e}^{+00}$ | ――― | $3.66{e}^{-07}$ | $1.26{e}^{-10}$ | $1.54{e}^{-12}$ | $1.01{e}^{-12}$ | |

ANN | $6.94{e}^{-09}$ | $1.22{e}^{-11}$ | $1.22{e}^{-11}$ | ――― | $7.96{e}^{-02}$ | $6.24{e}^{-11}$ | $1.69{e}^{-08}$ | |

1D CNN | $5.17{e}^{-04}$ | $3.17{e}^{-01}$ | $3.17{e}^{-01}$ | $2.85{e}^{-10}$ | ――― | $1.94{e}^{-09}$ | $2.36{e}^{-06}$ | |

CNN | $5.97{e}^{-05}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $1.22{e}^{-11}$ | $3.17{e}^{-01}$ | ――― | $7.24{e}^{-06}$ | |

CNN Log | $5.97{e}^{-05}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $1.22{e}^{-11}$ | $3.17{e}^{-01}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A10.**Kruskal–Wallis p-values comparing all methods of the Cu_np concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

Cu_np | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $2.12{e}^{-02}$ | $3.28{e}^{-06}$ | $3.73{e}^{-05}$ | $3.82{e}^{-04}$ | $6.96{e}^{-01}$ | $3.70{e}^{-02}$ | Test |

XG | $2.06{e}^{-02}$ | ――― | $1.20{e}^{-08}$ | $3.32{e}^{-08}$ | $2.82{e}^{-07}$ | $2.64{e}^{-03}$ | $3.08{e}^{-01}$ | |

Rocket | $2.06{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | $2.68{e}^{-01}$ | $2.80{e}^{-02}$ | $1.15{e}^{-07}$ | $9.62{e}^{-10}$ | |

ANN | $4.23{e}^{-09}$ | $1.40{e}^{-10}$ | $1.40{e}^{-10}$ | ――― | $4.85{e}^{-01}$ | $1.91{e}^{-07}$ | $2.78{e}^{-09}$ | |

1D CNN | $8.49{e}^{-01}$ | $4.02{e}^{-02}$ | $4.02{e}^{-02}$ | $4.21{e}^{-08}$ | ――― | $2.90{e}^{-07}$ | $3.52{e}^{-09}$ | |

CNN | $2.06{e}^{-02}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $1.40{e}^{-10}$ | $4.02{e}^{-02}$ | ――― | $7.65{e}^{-03}$ | |

CNN Log | $2.06{e}^{-02}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $1.40{e}^{-10}$ | $4.02{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A11.**Kruskal–Wallis p-values comparing all methods of the Ti_d concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

Ti_d | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $8.88{e}^{-01}$ | $6.10{e}^{-07}$ | $3.86{e}^{-07}$ | $2.36{e}^{-05}$ | $1.72{e}^{-01}$ | $1.16{e}^{-02}$ | Test |

XG | $1.27{e}^{-03}$ | ――― | $1.66{e}^{-06}$ | $8.50{e}^{-07}$ | $3.70{e}^{-05}$ | $1.73{e}^{-01}$ | $3.76{e}^{-02}$ | |

Rocket | $1.27{e}^{-03}$ | $1.00{e}^{+00}$ | ――― | $6.24{e}^{-01}$ | $6.19{e}^{-01}$ | $3.88{e}^{-08}$ | $5.14{e}^{-09}$ | |

ANN | $7.26{e}^{-04}$ | $7.92{e}^{-07}$ | $7.92{e}^{-07}$ | ――― | $2.06{e}^{-01}$ | $5.14{e}^{-10}$ | $1.10{e}^{-10}$ | |

1D CNN | $4.07{e}^{-01}$ | $2.06{e}^{-02}$ | $2.06{e}^{-02}$ | $4.20{e}^{-04}$ | ――― | $1.23{e}^{-05}$ | $2.54{e}^{-08}$ | |

CNN | $1.27{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $7.92{e}^{-07}$ | $2.06{e}^{-02}$ | ――― | $2.53{e}^{-05}$ | |

CNN Log | $1.27{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $7.92{e}^{-07}$ | $2.06{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A12.**Kruskal–Wallis p-values comparing all methods of the Ti_np concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

Ti_np | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $8.23{e}^{-04}$ | $1.12{e}^{-04}$ | $1.47{e}^{-03}$ | $2.21{e}^{-02}$ | $4.84{e}^{-06}$ | $4.86{e}^{-01}$ | Test |

XG | $1.54{e}^{-01}$ | ――― | $8.15{e}^{-10}$ | $1.51{e}^{-08}$ | $4.89{e}^{-07}$ | $4.45{e}^{-11}$ | $5.04{e}^{-03}$ | |

Rocket | $1.54{e}^{-01}$ | $1.00{e}^{+00}$ | ――― | $3.94{e}^{-01}$ | $7.37{e}^{-02}$ | $1.54{e}^{-01}$ | $7.93{e}^{-06}$ | |

ANN | $1.86{e}^{-06}$ | $3.06{e}^{-07}$ | $3.06{e}^{-07}$ | ――― | $3.61{e}^{-01}$ | $4.02{e}^{-02}$ | $1.09{e}^{-04}$ | |

1D CNN | $5.84{e}^{-01}$ | $3.17{e}^{-01}$ | $3.17{e}^{-01}$ | $1.15{e}^{-06}$ | ――― | $5.25{e}^{-03}$ | $9.09{e}^{-04}$ | |

CNN | $1.54{e}^{-01}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $3.06{e}^{-07}$ | $3.17{e}^{-01}$ | ――― | $2.52{e}^{-07}$ | |

CNN Log | $1.54{e}^{-01}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $3.06{e}^{-07}$ | $3.17{e}^{-01}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A13.**Kruskal–Wallis p-values comparing all methods of the Zn_d concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

Zn_d | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $9.88{e}^{-01}$ | $9.85{e}^{-06}$ | $4.13{e}^{-05}$ | $1.13{e}^{-03}$ | $1.00{e}^{+00}$ | $1.90{e}^{-02}$ | Test |

XG | $5.34{e}^{-03}$ | ――― | $1.35{e}^{-07}$ | $1.07{e}^{-06}$ | $1.37{e}^{-04}$ | $8.21{e}^{-01}$ | $1.93{e}^{-03}$ | |

Rocket | $5.34{e}^{-03}$ | $1.00{e}^{+00}$ | ――― | $1.68{e}^{-01}$ | $3.51{e}^{-02}$ | $7.79{e}^{-10}$ | $2.72{e}^{-08}$ | |

ANN | $9.01{e}^{-06}$ | $4.58{e}^{-09}$ | $4.58{e}^{-09}$ | ――― | $1.66{e}^{-01}$ | $6.71{e}^{-11}$ | $1.29{e}^{-05}$ | |

1D CNN | $8.65{e}^{-02}$ | $1.54{e}^{-01}$ | $1.54{e}^{-01}$ | $1.29{e}^{-07}$ | ――― | $6.29{e}^{-07}$ | $3.27{e}^{-02}$ | |

CNN | $5.34{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $4.58{e}^{-09}$ | $1.54{e}^{-01}$ | ――― | $3.28{e}^{-07}$ | |

CNN Log | $5.34{e}^{-03}$ | $1.00{e}^{+00}$ | $1.00{e}^{+00}$ | $4.58{e}^{-09}$ | $1.54{e}^{-01}$ | $1.00{e}^{+00}$ | ――― | |

Training |

**Table A14.**Kruskal–Wallis p-values comparing all methods of the zn_np concentration prediction task. Above the diagonal, we have the results for the test set, and below for the training set. Significant results ($p<0.01$) are represented in green or red when the method on the left is significantly better or worse, respectively.

Zn_np | RF | XG | Rocket | ANN | 1D CNN | CNN | CNN Log | |
---|---|---|---|---|---|---|---|---|

RF | ――― | $8.92{e}^{-01}$ | $1.75{e}^{-06}$ | $1.07{e}^{-05}$ | $1.83{e}^{-03}$ | $3.11{e}^{-03}$ | $1.25{e}^{-01}$ | Test |

XG | $5.07{e}^{-06}$ | ――― | $7.84{e}^{-08}$ | $1.63{e}^{-07}$ | $6.85{e}^{-05}$ | $2.59{e}^{-05}$ | $4.68{e}^{-02}$ | |

Rocket | $1.94{e}^{-05}$ | $3.17{e}^{-01}$ | ――― | $8.92{e}^{-01}$ | $1.72{e}^{-01}$ | $6.42{e}^{-03}$ | $1.89{e}^{-10}$ | |

ANN | $1.05{e}^{-01}$ | $1.26{e}^{-11}$ | $4.89{e}^{-11}$ | ――― | $1.03{e}^{-01}$ | $6.66{e}^{-04}$ | $1.31{e}^{-10}$ | |

1D CNN | $2.74{e}^{-04}$ | $7.81{e}^{-02}$ | $3.13{e}^{-01}$ | $1.80{e}^{-09}$ | ――― | $1.12{e}^{-01}$ | $1.31{e}^{-06}$ | |

CNN | $5.07{e}^{-06}$ | $1.00{e}^{+00}$ | $3.17{e}^{-01}$ | $1.26{e}^{-11}$ | $7.81{e}^{-02}$ | ――― | $1.91{e}^{-08}$ | |

CNN Log | $5.07{e}^{-06}$ | $1.00{e}^{+00}$ | $3.17{e}^{-01}$ | $1.26{e}^{-11}$ | $7.81{e}^{-02}$ | $1.00{e}^{+00}$ | ――― | |

Training |

## References

- Gavrilescu, M.; Demnerová, K.; Aamand, J.; Agathos, S.; Fava, F. Emerging pollutants in the environment: Present and future challenges in biomonitoring, ecological risks and bioremediation. New Biotechnol.
**2015**, 32, 147–156. [Google Scholar] [CrossRef] [PubMed] - Robin, L.; Georg, H.; Michael, A.; John, B.; Anders, B.; Jordi, D.; Ian, D.; Yuriy, D.; Anja, D.; Barak, H.; et al. Marine Strategy Framework Directive Task Group 8 Report: Contaminants and Pollution Effects; Office for Official Publications of the European Communities: Luxembourg, 2010. [Google Scholar]
- CAS. CAS Registry Keeps Pace with Rapid Growth of Chemical Research, Registers 60 Millionth Substance; Chemistry Abstract Service; CAS Customer Center: Columbus, OH, USA, 2011. [Google Scholar]
- Duarte, B.; Gameiro, C.; Matos, A.R.; Figueiredo, A.; Silva, M.S.; Cordeiro, C.; Caçador, I.; Reis-Santos, P.; Fonseca, V.; Cabrita, M.T. First screening of biocides, persistent organic pollutants, pharmaceutical and personal care products in Antarctic phytoplankton from Deception Island by FT-ICR-MS. Chemosphere
**2021**, 274, 129860. [Google Scholar] [CrossRef] [PubMed] - Reis-Santos, P.; Pais, M.; Duarte, B.; Caçador, I.; Freitas, A.; Pouca, A.S.V.; Barbosa, J.; Leston, S.; Rosa, J.; Ramos, F.; et al. Screening of human and veterinary pharmaceuticals in estuarine waters: A baseline assessment for the Tejo estuary. Mar. Pollut. Bull.
**2018**, 135, 1079–1084. [Google Scholar] [CrossRef] [PubMed] - Carvalho, R.C.D.; Feijão, E.; Matos, A.; Cabrita, M.T.; Novais, S.; Lemos, M.; Caçador, I.; Marques, J.C.; Reis-Santos, P.; Fonseca, V.; et al. Glyphosate-Based Herbicide Toxicophenomics in Marine Diatoms: Impacts on Primary Production and Physiological Fitness. Appl. Sci.
**2020**, 10, 7391. [Google Scholar] [CrossRef] - Feijão, E.; Cruz de Carvalho, R.; Duarte, I.A.; Matos, A.R.; Cabrita, M.T.; Novais, S.C.; Lemos, M.F.L.; Caçador, I.; Marques, J.C.; Reis-Santos, P.; et al. Fluoxetine Arrests Growth of the Model Diatom Phaeodactylum tricornutum by Increasing Oxidative Stress and Altering Energetic and Lipid Metabolism. Front. Microbiol.
**2020**, 11, 1803. [Google Scholar] [CrossRef] - Franzitta, M.; Feijão, E.; Cabrita, M.T.; Gameiro, C.; Matos, A.; Marques, J.C.; Goessling, J.W.; Reis-Santos, P.; Fonseca, V.; Pretti, C.; et al. Toxicity Going Nano: Ionic Versus Engineered Cu Nanoparticles Impacts on the Physiological Fitness of the Model Diatom Phaeodactylum tricornutum. Front. Mar. Sci.
**2020**, 7, 1154. [Google Scholar] [CrossRef] - Duarte, B.; Gameiro, C.; Utkin, A.B.; Matos, A.R.; Caçador, I.; Fonseca, V.; Cabrita, M.T. A multivariate approach to chlorophyll A Fluoresc. Data Trace Elem. Ecotoxicological Trials Using A Model Mar. Diatom. Estuarine Coast. Shelf Sci.
**2021**, 250, 107170. [Google Scholar] [CrossRef] - Duarte, B.; Durante, L.; Marques, J.C.; Reis-Santos, P.; Fonseca, V.F.; Caçador, I. Development of a toxicophenomic index for trace element ecotoxicity tests using the halophyte Juncus acutus: Juncus-TOX. Ecol. Indic.
**2021**, 121, 107097. [Google Scholar] [CrossRef] - Anjum, N.; Duarte, B.; Caçador, I.; Sleimi, N.; Duarte, A.; Pereira, E. Biophysical and Biochemical Markers of Metal/Metalloid-Impacts in Salt Marsh Halophytes and Their Implications. Front. Environ. Sci.
**2016**, 4, 24. [Google Scholar] [CrossRef] [Green Version] - Duarte, B.; Pedro, S.; Marques, J.C.; Adão, H.; Caçador, I. Zostera noltii development probing using chlorophyll a transient analysis (JIP-test) under field conditions: Integrating physiological insights into a photochemical stress index. Ecol. Indic.
**2017**, 76, 219–229. [Google Scholar] [CrossRef] - Santos, D.; Duarte, B.; Caçador, I. Unveiling Zn hyperaccumulation in Juncus acutus: Implications on the electronic energy fluxes and on oxidative stress with emphasis on non-functional Zn-chlorophylls. J. Photochem. Photobiol. B Biol.
**2014**, 140, 228–239. [Google Scholar] [CrossRef] [PubMed] - Carvalho, R.C.D.; Feijão, E.; Kletschkus, E.; Marques, J.; Reis-Santos, P.; Fonseca, V.; Papenbrock, J.; Caçador, I.; Duarte, B. Halophyte bio-optical phenotyping: A multivariate photochemical pressure index (Multi-PPI) to classify salt marsh anthropogenic pressures levels. Ecol. Indic.
**2020**, 119, 106816. [Google Scholar] [CrossRef] - Cabrita, M.T.; Gameiro, C.; Utkin, A.; Duarte, B.; Caçador, I.; Cartaxana, P. Photosynthetic pigment laser-induced fluorescence indicators for the detection of changes associated with trace element stress in the diatom model species Phaeodactylum tricornutum. Environ. Monit. Assess.
**2016**, 188, 285. [Google Scholar] [CrossRef] - Duarte, B.; Feijão, E.; Cruz de Carvalho, R.; Duarte, I.A.; Silva, M.; Matos, A.R.; Cabrita, M.T.; Novais, S.C.; Lemos, M.F.L.; Marques, J.C.; et al. Effects of Propranolol on Growth, Lipids and Energy Metabolism and Oxidative Stress Response of Phaeodactylum tricornutum. Biology
**2020**, 9, 478. [Google Scholar] [CrossRef] - Duarte, B.; Cabrita, M.T.; Vidal, T.; Pereira, J.; Pacheco, M.; Pereira, P.; Canário, J.; Gonçalves, F.; Matos, A.; Rosa, R.; et al. Phytoplankton community-level bio-optical assessment in a naturally mercury contaminated Antarctic ecosystem (Deception Island). Mar. Environ. Res.
**2018**, 140, 412–421. [Google Scholar] [CrossRef] - Rodrigues, N.M.; Batista, J.E.; Trujillo, L.; Duarte, B.; Giacobini, M.; Vanneschi, L.; Silva, S. Plotting time: On the usage of CNNs for time series classification. arXiv
**2021**, arXiv:2102.04179. [Google Scholar] - Koh, J.C.O.; Spangenberg, G.; Kant, S. Automated Machine Learning for High-Throughput Image-Based Plant Phenotyping. bioRxiv
**2020**. [Google Scholar] [CrossRef] - Guillard, R.; Ryther, J. Studies of marine planktonic diatoms. I. Cyclotella nana Hustedt, and Detonula confervacea (cleve) Gran. Can. J. Microbiol.
**1962**, 8, 229–239. [Google Scholar] [CrossRef] [PubMed] - Feijão, E.; Gameiro, C.; Franzitta, M.; Duarte, B.; Caçador, I.; Cabrita, M.T.; Matos, A. Heat wave impacts on the model diatom Phaeodactylum tricornutum: Searching for photochemical and fatty acid biomarkers of thermal stress. Ecol. Indic.
**2018**, 95, 1026–1037. [Google Scholar] [CrossRef] - Cabrita, M.T.; Duarte, B.; Gameiro, C.; Godinho, R.; Caçador, I. Photochemical features and trace element substituted chlorophylls as early detection biomarkers of metal exposure in the model diatom Phaeodactylum tricornutum. Ecol. Indic.
**2018**, 95, 1038–1052. [Google Scholar] [CrossRef] - Duarte, B.; Prata, D.; Matos, A.; Cabrita, M.T.; Caçador, I.; Marques, J.; Cabral, H.; Reis-Santos, P.; Fonseca, V. Ecotoxicity of the lipid-lowering drug bezafibrate on the bioenergetics and lipid metabolism of the diatom Phaeodactylum tricornutum. Sci. Total Environ.
**2019**, 650, 2085–2094. [Google Scholar] [CrossRef] [PubMed] - Silva, M.; Feijão, E.; da Cruz de Carvalho, R.; Duarte, I.A.; Matos, A.; Cabrita, M.T.; Barreiro, A.; Lemos, M.; Novais, S.; Marques, J.H.; et al. Comfortably numb: Ecotoxicity of the non-steroidal anti-inflammatory drug ibuprofen on Phaeodactylum tricornutum. Mar. Environ. Res.
**2020**, 161, 105109. [Google Scholar] [CrossRef] [PubMed] - Duarte, B.; Marques, J.C.; Caçador, I. Ecophysiological response of native and invasive Spartina species to extreme temperature events in Mediterranean marshes. Biol. Invasions
**2015**, 18, 2189–2205. [Google Scholar] [CrossRef] - Breiman, L. Random Forests. Mach. Learn.
**2001**, 45, 5–32. [Google Scholar] [CrossRef] [Green Version] - Chen, T.; Guestrin, C. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference, San Francisco, CA, USA, 13–17 August 2016; pp. 785–794. [Google Scholar] [CrossRef] [Green Version]
- Dempster, A.; Petitjean, F.; Webb, G.I. ROCKET: Exceptionally fast and accurate time series classification using random convolutional kernels. Data Min. Knowl. Discov.
**2020**, 34, 1454–1495. [Google Scholar] [CrossRef] - Goodfellow, I.; Bengio, Y.; Courville, A. Deep Learning; The MIT Press: Cambridge, MA, USA, 2016. [Google Scholar]
- Pedregosa, F.; Varoquaux, G.; Gramfort, A.; Michel, V.; Thirion, B.; Grisel, O.; Blondel, M.; Prettenhofer, P.; Weiss, R.; Dubourg, V.; et al. Scikit-learn: Machine Learning in Python. J. Mach. Learn. Res.
**2011**, 12, 2825–2830. [Google Scholar] - Abadi, M.; Agarwal, A.; Barham, P.; Brevdo, E.; Chen, Z.; Citro, C.; Corrado, G.S.; Davis, A.; Dean, J.; Devin, M.; et al. TensorFlow: Large-Scale Machine Learning on Heterogeneous Systems. 2015. Available online: tensorflow.org (accessed on 18 August 2021).

**Figure 1.**Architectures of the different deep learning models used in this work. The first network corresponds to the ANN used, it is composed by a set of BatchNormalization, Dense (fully connected), and Dropout layers. The second corresponds to the 1D CNN, composed by a set of BatchNormalization, 1D Convolution, and 1D MaxPooling layers, followed by a flatten operation to reduce the dimensionality of the data in order to pass it to the two dense layers that serve as a classifier. The last network corresponds to our 2D CNN, composed by 2D Convolution, 2D MaxPooling, a flatten and both Dense and Dropout layers. Each layer from models has its parameters discretized inside the corresponding box.

**Figure 2.**Heatmap depicting the accuracy per class of each best performing model when predicting the different types of ECs.

**Figure 3.**Boxplot showing the distribution of accuracy obtained by each method on the test set, in 30 independent runs of predicting the EC.

**Figure 4.**Spider plot depicting the accuracy per class of each best performing model when predicting the different types of ECs.

**Figure 5.**Boxplots showing the distribution of accuracy obtained by each method on the test set, in 30 independent runs of predicting the concentration of each given EC.

**Figure 6.**Heatmap depicting the accuracy per class of the models when predicting the different concentrations of the different ECs. This heatmap is a concatenation of 13 smaller ones. The x-axis lists all the ECs at different concentrations (see Table 1).

**Figure 7.**Spider plots depicting the accuracy per class of each model when predicting the different concentrations of an EC. Each vertex within the circle corresponds to the concentration of that EC.

**Table 1.**OPTOX dataset summary, containing the contaminants, the different concentration values and the distribution of samples among them. The first section of the table includes pharma, pesticides, and personal care products, while the second section includes metals and nanoparticles.

EC | Concentrations ($\mathsf{\mu}$g L${}^{-1}$) | #Samples | #Samples per Concentration | |
---|---|---|---|---|

Diclofenac | dcf | $[0,0.8,3,40,100,300]$ | 180 | 30 |

Fluoxetine | fx | $[0,0.3,0.6,20,40,80]$ | 180 | 30 |

Glyphosate | gliph | $[0,10,50,100,250,500]$ | 180 | 30 |

Ibuprofen | ibu | $[0,0.3,3,40,100,300]$ | 180 | 30 |

Propranolol | prop | $[0,0.3,8,80,150,300]$ | 180 | 30 |

Sodium Dodecyl Sulphate | sds | $[0,0.1,1,3,10]$ | 150 | 30 |

Triclosan | tri | $[0,0.1,1,10,50,100]$ | 180 | 30 |

Dissolved Ionic Copper | Cu_{d} | $[0,1,5,10]$ | 120 | 30 |

Copper Engineered Nanoparticles | Cu_{np} | |||

Dissolved Ionic Titanium | Ti_{d} | $[0,10,50,200]$ | 120 | 30 |

Titanium Engineered Nanoparticles | Ti_{np} | |||

Dissolved Ionic Zinc | Zn_{d} | $[0,1,5,10]$ | 120 | 30 |

Zinc Engineered Nanoparticles | Zn_{np} |

**Table 2.**Octanol–water partition coefficients (log KOW; NA—not applicable) and concentrations applied of the tested compounds, and respective growth inhibition (%, negative values indicate growth inhibition) and half-maximal inhibitory concentration (IC${}_{50}$) assessed for each exposure trial set.

Exposure | log K (OW) (OECD 107) | Concentration ($\mathsf{\mu}$g L${}^{-1}$) | Growth Inhibition (%) | IC${}_{50}$ ($\mathsf{\mu}$g L${}^{-1}$) |
---|---|---|---|---|

Diclofenac | 1.9 | 0 | 0 | 318.9 |

0.8 | $-8.5$ | |||

3 | $-14.9$ | |||

40 | $-34.9$ | |||

100 | $-38.9$ | |||

300 | $-42.8$ | |||

Propranolol | 3.12 | 0 | 0 | 194.6 |

0.3 | $-19.37$ | |||

8 | 1.7 | |||

80 | $-31$ | |||

150 | $-68.6$ | |||

300 | $-56.1$ | |||

Fluoxetine | 4.65 | 0 | 0 | 47.3 |

0.3 | $-7.42$ | |||

0.6 | $-10.1$ | |||

20 | $-24.8$ | |||

40 | $-38.8$ | |||

80 | $-82.8$ | |||

Ibuprofen | 2.48 | 0 | 0 | 350.6 |

0.8 | $-17.1$ | |||

3 | $-20.1$ | |||

40 | $-18.2$ | |||

100 | $-37.7$ | |||

300 | $-40$ | |||

Glyphosate | $-1.6$ | 0 | 0 | 225.9 |

10 | 0.92 | |||

50 | $-21$ | |||

100 | $-28$ | |||

250 | $-75.5$ | |||

500 | $-89.2$ | |||

SDS | $-2.03$ | 0 | 0 | 11.4 |

0.1 | $-15.5$ | |||

1 | $-21.1$ | |||

3 | $-24.6$ | |||

10 | $-43.8$ | |||

Triclosan | 4.76 | 0 | 0 | 691.7 |

0.1 | $-0.55$ | |||

1 | −1.41 | |||

10 | −1.6 | |||

100 | −2.39 | |||

Dissolved Ionic Copper | NA | 0 | 0 | 14.6 |

1 | −12.2 | |||

5 | −24.1 | |||

10 | −33.6 | |||

Copper Engineered Nanoparticles | NA | 0 | 0 | 8.31 |

1 | $-21$ | |||

5 | $-33.1$ | |||

10 | $-57.7$ | |||

Dissolved Ionic Titanium | NA | 0 | 0 | 577.5 |

10 | $-5.15$ | |||

50 | $-11.9$ | |||

200 | $-18$ | |||

Titanium Engineered Nanoparticles | NA | 0 | 0 | 634.3 |

10 | 12.3 | |||

50 | 0.47 | |||

200 | $-7.46$ | |||

Dissolved Ionic Zinc | NA | 0 | 0 | 73.9 |

1 | $-7.56$ | |||

5 | $-6.51$ | |||

10 | $-8.37$ | |||

Zinc Engineered Nanoparticles | NA | 0 | 0 | 88.4 |

1 | 3.39 | |||

5 | $-4.27$ | |||

10 | $-7.22$ |

**Table 3.**Median overall accuracy obtained by the different methods when predicting the different types of ECs on the training and test set. The best test result is represented in green.

Train | Test | |
---|---|---|

Random Forests | $0.9625$ | $0.7461$ |

XGBoost | $0.9983$ | $0.8076$ |

Rocket | $0.9991$ | $0.9705$ |

ANN | $0.8493$ | $0.9089$ |

1D CNN | $0.9993$ | $0.9641$ |

CNN | $0.9815$ | $0.9732$ |

CNN Log | $0.9807$ | $0.9765$ |

**Table 4.**Median overall accuracy obtained by the different methods when predicting the different concentrations of the ECs on the training and test set. The best test results for each EC are represented in green. The last row reports the median results for all ECs.

RF | XGB | Rocket | ANN | 1D CNN | CNN | CNN Log | ||||||||
---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|

Train | Test | Train | Test | Train | Test | Train | Test | Train | Test | Train | Test | Train | Test | |

DCF | 1 | $0.9$ | 1 | $0.8833$ | 1 | 1 | 1 | 1 | 1 | 1 | 1 | $0.9666$ | 1 | $0.9666$ |

FX | 1 | $0.8333$ | 1 | $0.8333$ | 1 | $0.9666$ | $0.9833$ | $0.95$ | 1 | $0.9333$ | 1 | $0.9$ | 1 | $0.8333$ |

GLIPH | 1 | $0.9666$ | 1 | $0.9333$ | 1 | 1 | $0.9833$ | $0.9666$ | 1 | 1 | 1 | $0.9333$ | 1 | $0.9666$ |

IBU | $0.9916$ | $0.8666$ | 1 | $0.8333$ | $0.9708$ | $0.9$ | $0.9916$ | $0.95$ | 1 | $0.9333$ | 1 | $0.8666$ | 1 | $0.833$ |

PROP | 1 | $0.9666$ | 1 | $0.9666$ | 1 | 1 | 1 | $0.9833$ | 1 | 1 | 1 | $0.9666$ | 1 | 1 |

SDS | 1 | $0.8333$ | 1 | $0.7916$ | 1 | $0.9166$ | $0.9895$ | $0.9583$ | 1 | $0.875$ | 1 | $0.7916$ | 1 | $0.75$ |

TRI | 1 | $0.9$ | 1 | $0.8333$ | 1 | 1 | $0.9833$ | $0.9666$ | 1 | $0.9666$ | 1 | $0.9333$ | 1 | $0.9333$ |

Cu_d | 1 | $0.8333$ | 1 | $0.8888$ | 1 | 1 | $0.9722$ | $0.9444$ | 1 | $0.9444$ | 1 | $0.8333$ | 1 | $0.8888$ |

Cu_np | 1 | $0.8888$ | 1 | $0.8333$ | 1 | $0.9444$ | $0.9722$ | $0.9444$ | 1 | $0.9444$ | 1 | $0.8888$ | 1 | $0.8333$ |

Ti_d | 1 | $0.8888$ | 1 | $0.8888$ | 1 | 1 | $0.9861$ | 1 | 1 | 1 | 1 | $0.8888$ | 1 | $0.8611$ |

Ti_np | 1 | $0.9444$ | 1 | $0.8888$ | 1 | 1 | $0.9861$ | 1 | 1 | 1 | 1 | 1 | 1 | $0.9444$ |

Zn_d | 1 | $0.8888$ | 1 | $0.8888$ | 1 | $0.9722$ | $0.9861$ | $0.9444$ | 1 | $0.9444$ | 1 | $0.8611$ | 1 | $0.9444$ |

Zn_np | $0.9791$ | $0.7777$ | 1 | $0.7777$ | 1 | $0.8888$ | $0.9722$ | $0.9166$ | 1 | $0.8888$ | 1 | $0.8611$ | 1 | $0.7222$ |

Median | 1 | $0.8888$ | 1 | $0.8833$ | 1 | 1 | $0.9861$ | $0.9583$ | 1 | $0.9444$ | 1 | $0.8888$ | 1 | $0.8888$ |

Publisher’s Note: MDPI stays neutral with regard to jurisdictional claims in published maps and institutional affiliations. |

© 2021 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/).

## Share and Cite

**MDPI and ACS Style**

Rodrigues, N.M.; Batista, J.E.; Mariano, P.; Fonseca, V.; Duarte, B.; Silva, S.
Artificial Intelligence Meets Marine Ecotoxicology: Applying Deep Learning to Bio-Optical Data from Marine Diatoms Exposed to Legacy and Emerging Contaminants. *Biology* **2021**, *10*, 932.
https://doi.org/10.3390/biology10090932

**AMA Style**

Rodrigues NM, Batista JE, Mariano P, Fonseca V, Duarte B, Silva S.
Artificial Intelligence Meets Marine Ecotoxicology: Applying Deep Learning to Bio-Optical Data from Marine Diatoms Exposed to Legacy and Emerging Contaminants. *Biology*. 2021; 10(9):932.
https://doi.org/10.3390/biology10090932

**Chicago/Turabian Style**

Rodrigues, Nuno M., João E. Batista, Pedro Mariano, Vanessa Fonseca, Bernardo Duarte, and Sara Silva.
2021. "Artificial Intelligence Meets Marine Ecotoxicology: Applying Deep Learning to Bio-Optical Data from Marine Diatoms Exposed to Legacy and Emerging Contaminants" *Biology* 10, no. 9: 932.
https://doi.org/10.3390/biology10090932