Improved Lipophilicity and Aqueous Solubility Prediction with Composite Graph Neural Networks

Abstract

The accurate prediction of molecular properties, such as lipophilicity and aqueous solubility, are of great importance and pose challenges in several stages of the drug discovery pipeline. Machine learning methods, such as graph-based neural networks (GNNs), have shown exceptionally good performance in predicting these properties. In this work, we introduce a novel GNN architecture, called directed edge graph isomorphism network (D-GIN). It is composed of two distinct sub-architectures (D-MPNN, GIN) and achieves an improvement in accuracy over its sub-architectures employing various learning, and featurization strategies. We argue that combining models with different key aspects help make graph neural networks deeper and simultaneously increase their predictive power. Furthermore, we address current limitations in assessment of deep-learning models, namely, comparison of single training run performance metrics, and offer a more robust solution.

1. Introduction

Oral bio-availability, drug uptake, and ADME-related properties of small molecules are key properties in pharmacokinetics. For drugs to reach their intended target, they need to pass through several barriers either by passive diffusion or carrier-mediated uptake typically mediated by lipophilicity and aqueous solubility. Compounds with poor solubility are unable to achieve that and, therefore, pose a higher risk in attrition and overall cost during development [1].

Methods based on deep-learning have proven successful in predicting molecular properties [2] and are becoming more and more a routine part of the modern computer-aided drug design toolbox for molecular design and med-chem decision support. Since molecules can be represented as graphs, an obvious approach is to employ a graph-based architecture for deep-learning, which leads to the utilization of graph-based neural networks (GNNs). These kinds of networks are capable of learning representations for a specific task in an automated way and, therefore, can eliminate the complicated feature engineering process where domain specialists have to select the list of descriptors themselves [3]. They became increasingly popular in the last few years [4,5,6] especially due to their success in chemical property prediction [5,7,8,9,10,11].

One of the first GNN models used for physicochemical property prediction was introduced by Micheli [12] in 2009. It predicted the boiling point of alkanes with a recursive architecture for structured data input and achieved an improved state-of-the-art performance. Lusci et al. [13] were the first to apply an undirected cyclic graph recurrent neural network on predicting aqueous solubility successfully. In the following years, several recurrent, spatial, and spectral graph-based neural networks were introduced [3,14,15,16]. One of them was the message passing framework, which was extended to include directed edges [3]. This network, called directed-edge message passing network (D-MPNN), is one of the most successful GNNs to predict chemical properties [1].

Despite the success, one important limitation with message passing networks is the graph isomorphism problem, meaning that they are unaware of the structural role of each node or edge [17]. Most standard GNNs, such as the D-MPNN, are incapable of distinguishing between different types of graph structures to determine whether they are topologically identical [18]. Compounds, such as naphthalene and 1,1-bi(cyclopentane), are perceived as the same structure by these networks. This can be problematic because they have vastly different chemical properties. To address this issue, graph isomophism networks (GIN), another group of GNNs, have recently received attention [18,19]. They are capable of distinguishing between these compounds by reformulating the message passing framework to incorporate the Weisfeiler–Lehman (WL) hierarchy. They try to be at least as expressive as the Weisfeiler–Lehman graph isomorphism test (WL-test) [20] and have shown good results in chemical property prediction [18,19] despite often falling short with respect to speed and accuracy to other frameworks, such as the D-MPNN [21]. Inspired by the key property of the GIN and the success of the D-MPNN framework, we combined the key characteristics of both architectures. By doing so we not only address the isomorphism problem but also incorporate one of the most successful and powerful GNN frameworks to improve lipophilicity and aqueous solubility prediction.

When comparing new machine learning architectures with previously published methods, the standard approach is to compare single performance metrics, such as root mean squared error (RMSE) values on a test set to show model performance [21,22]. This can lead to reproducibility issues as stochastic algorithms like neural networks can vary greatly in their prediction, even without changing their hyperparameters, simply by using different training, validation, test set splits or non-deterministic weight initializations [23,24]. One of the reasons for this is the complex landscape that optimizers have to navigate through in modern machine learning models. In real world applications these landscapes can have multiple local minima and it is especially hard for non-deterministic optimization algorithms like stochastic gradient descent to find the global minimum, therefore often retrieving different results when repeated [25]. This problem can be intensified by using small datasets with different random splits for training and evaluation. Such an approach can lead the optimization algorithm into different local minima and makes it almost impossible for the model to generalize [2]. It is, therefore, difficult to compare different deep-learning model architectures with each other even when using the same data [23]. Another challenge is especially prominent in the GNN domain, where the optimal features for node or edge representation are unknown. Deep-learning benchmark studies often use the same data but different representations for their input data which makes it difficult to make a fair comparison between the models [2,3].

To mitigate these problems, we use the exact same data split to train, evaluate, and test each of the used models with different node and edge features, as well as learning strategies to obtain an average performance independent of the used features and training approaches. Such a procedure is time consuming as multiple models have to be evaluated several times. Nevertheless, obtaining a better overview of the behaviour of GNNs under these different constraints will facilitate the understanding of these architectures and ultimately help advance GNNs beyond the current hype to more explainable and robust models.

Our contribution is a novel graph neural network architecture called directed edge graph isomorphism network (D-GIN). It extends the directed edge message passing (D-MPNN) framework [1] by the graph isomorphism network (GIN) [18]. An overview of the D-GIN model is shown in Figure 1. Our novel architecture shows improved performance compared to its individual, less complex networks, and we demonstrate that combining models with different key aspects help make graph neural networks deeper while simultaneously increasing their predictive power. We evaluated our models by applying different learning and featurization strategies and compared their average performance under different constraints.

2. Materials and Methods

This section gives a detailed overview of the used data, molecular representation, and the different machine learning methods used throughout this work. The most common notations are shown in

Table 1. Common notations used throughout this publication.

Notation	Definition
$\tau$	A non-linear function (e.g., sigmoid or relu)
$cat(,)$	Vector concatenation
t	Iterator of t steps
G	A graph
V	Set of nodes
E	Set of edges
v	Node v ∈ V
$e_{uv}$	Edge $e_{uv}$ ∈ E between node u and v
$N\left(u\right)$	Neighbors of node u
$N\left(u\right)/w$	Neighbors of node u except w
n	The number of nodes
m	The number of edges
d	The dimension of a node feature vector
b	The dimension of a edge feature vector
X∈${}^{n\times d}$	Feature matrix of a graph
x${}_v$${\in }^d$	Feature vector of node v
$x^e{}_{uv}$∈${}^b$	Feature vector of edge $e_{uv}$
h${}_v$${\in }^c$	Hidden feature vector of node v
m${}_v$${\in }^c$	Message feature vector to node v
h${}_G$${\in }^c$	Feature vector of the graph G
h${}_{uv}$${\in }^d$	Hidden feature vector of edge $e_{uv}$
m${}_{uv}$${\in }^d$	Message feature vector to edge $e_{uv}$
W	Weight matrix of a neural network
A$\in {\{1,0\}}^{\left|n\right|x\left|n\right|}$	Adjacency matrix
RMSE	Root mean squared error
GNN	Graph neural network
GIN	Graph isomorphic network as in [18]
$ϵ$	Epsilon as described in [18]
D-MPNN	Directed-edge message passing network as in [1]
D-GIN	Directed-edge graph isomorphic network
CI	95% confidence interval calculated via bootstrapping
$f(·)$	Feed forward neural network

.

2.1. Experimental Data

A total of 10,617 molecules annotated with experimentally derived logD and logP values or logS and logP values were used for model training and predictions. The selected molecules were derived from the Delaney lipophilicity dataset containing experimentally evaluated logD and logP values at pH 7.4 [26] and an aqueous solubility set with logS and logP values [27]. Each dataset was evaluated and molecules were neutralized in both sets. For the aqueous solubility data, salts were stripped off and molecules with logS values lower than −10.0 or higher than 0.0 were removed. The original preprocessed and post-processed data can be found in the GitHub repository [28]. The splitting of each dataset into three subsets for training, evaluation, and testing was completed randomly in a ratio of 81:9:10 for the (training, evaluation, and testing). The data splitting was performed with the same seed for each of the models to be able to compare them using the exact same training, evaluation, and test data. The minimum value of each of the logD, logP, and logS properties was used as an offset to ensure only positive property values. The resulting lipophilicity dataset consisted of 4174 compounds. In total, 3380 were used for training, 376 for evaluating and model selection, and 418 for testing. The post processed solubility dataset contained 6443 molecules. Overall, 5219 compounds were allocated for training, 579 for evaluation, and model selection, and 645 for testing.

2.2. Training Approaches

The training strategies differ in the used dataset and the training target (logD, logP, or logS). Under these constraints, seven different types of strategies were used. The first multi-task learning strategy used a combined approach of logD, logP, and logS values referred to as “logD/P/S”. Three additional multi-task strategies utilized a combination of two physicochemical properties and are notated as either “logD/P”, “logD/S”, or “logS/P”. Three other single task strategies are only learned on a single physicochemical property and are referred to as either “single task logD”, “logP”, or “logS”. When physicochemical properties from different datasets were used, the individual datasets were first split into training, evaluation, and test sets. Afterwards, each physicochemical property was evaluated and tested individually so that the evaluation and test results of the multi-task learning approaches can be compared to those with a single-task learning strategy.

When testing either single-, or multi-task models, the combined root mean squared error (RMSE) for all properties was calculated as the measure for the best model. For logP, we only used the results from either the first multi-task approach (“multi-task logS/D/P”) or the single-task approach with logP values. The reasoning behind this was to use the same test and evaluation data for all models while trying to avoid an unbalanced data bias in favor of logP values. When training with two physicochemical properties where one was logP, we only used the data that had both properties. For example, when training on the lipophilicity dataset which had logP and logD values, we did not include logP compounds from the aqueous solubility dataset and vice versa.

2.3. Molecular Graphs

A graph is defined as $G=(V,E)$, where V is a set of nodes and E denotes a set of edges. Let $v\in V$ be a node with feature vector $x_v$ and $e{}_{uv}\in E$ be an edge pointing from u to v with feature vector $x^e{}_{uv}$. The adjacency matrix A shows the connectivity of the nodes and in our case it was binary as we did not weigh any connections. It is defined as a $n\times n$ matrix with $A{}_{uv}=1$ if $e{}_{uv}\in E$ and $A{}_{uv}=0$ if $e{}_{uv}\notin E$. We use directed, heterogeneous graphs where $e{}_{uv}\ne e{}_{v}u$. Heterogeneous graphs contain different types of nodes and edges with their corresponding featurizations.

2.4. Molecular Featurization

Five different types of edge and vertex featurizations X were being used for the GNNs. The detailed description of x and $x^e$ can be found in

Table A1. Node and edge featurization of type 3. Each node or edge featurization vector consisted of a concatenation of the different one-hot encoded or floating point feature vectors according to their possible states (if present) and corresponding size-36 for nodes and 20 for edges.

Feature	Possible States	Size
chemical element	H, C, N, O, S, F, P, Cl, Br, I	10
calculated formal charge	−2, −1, 0, 1, 2	5
CIP configuration	R, S, None, either	4
hybridization state	sp, sp2, sp3, sp3d, sp3d2, none	6
amide center	yes, no	2
present in aromatic ring	yes, no	2
ring size	1/size	1
nr. of hydrogens	0, 1, 2, 3, 4, 5	6
bond order	1, 2, 3	3
conjugated	yes, no	2
rotate−able	yes, no	2
amide bond	yes, no	2
present in aromatic ring	yes, no	2
present in ring	yes, no	2
ring size	1/size	1
CIP configuration	none, E, Z, trans, cis, either	6

,

Table A2. Node and edge featurization of type 4. Each node or edge featurization vector consisted of a concatenation of the different one-hot encoded or floating point feature vectors according to their possible states (if present) and corresponding size-45 for nodes and 29 for edges.

Feature	Possible States	Size
chemical element	H, C, N, O, S, F, P, Cl, Br, I	10
calculated formal charge	−2, −1, 0, 1, 2	5
CIP configuration	R, S, None, either	4
hybridization state	sp, sp2, sp3, sp3d, sp3d2, none	6
amide center	yes, no	2
present in aromatic ring	yes, no	2
ring size	0, 3, 4, 5, 6, 7, 8, 9, 10, 11	1
nr. of hydrogens	0, 1, 2, 3, 4, 5	6
bond order	1, 2, 3	3
conjugated	yes, no	2
rotate-able	yes, no	2
amide bond	yes, no	2
present in aromatic ring	yes, no	2
present in ring	yes, no	2
ring size	0, 3, 4, 5, 6, 7, 8, 9, 10, 11	1
CIP configuration	none, E, Z, trans, cis, either	6

,

Table A3. Node and edge featurization of type 5. Each node or edge featurization vector consisted of a concatenation of the different one-hot encoded or floating point feature vectors according to their possible states (if present) and corresponding size-15 for nodes and 3 for edges.

Feature	Possible States	Size
chemical element	H, C, N, O, S, F, P, Cl, Br, I	10
calculated formal charge	−2, −1, 0, 1, 2	5
CIP configuration	-	-
hybridization state	-	-
amide center	-	-
present in aromatic ring	-	-
ring size	-	-
nr. of hydrogens	-	-
bond order	1, 2, 3	3
conjugated	-	-
rotate-able	-	-
amide bond	-	-
present in aromatic ring	-	-
present in ring	-	-
ring size	-	-
CIP configuration	-	-

,

Table A4. Node and edge featurization of type 6. Each node or edge featurization vector consisted of a concatenation of the different one-hot encoded or floating point feature vectors according to their possible states (if present) and corresponding size-15 for nodes and 20 for edges.

Feature	Possible States	Size
chemical element	H, C, N, O, S, F, P, Cl, Br, I	10
calculated formal charge	−2, −1, 0, 1, 2	5
CIP configuration	-	-
hybridization state	-	-
amide center	-	-
present in aromatic ring	-	-
present in ring	-	-
nr. of hydrogens	-	-
bond order	1, 2, 3	3
conjugated	yes, no	2
rotate-able	yes, no	2
amide bond	yes, no	2
present in aromatic ring	yes, no	2
present in ring	yes, no	2
ring size	1/size	1
CIP configuration	none, E, Z, trans, cis, either	6

,

Table A5. Node and edge featurization of type 7. Each node or edge featurization vector consisted of a concatenation of the different one-hot encoded or floating point feature vectors according to their possible states (if present) and corresponding size-36 for nodes and 3 for edges.

Feature	Possible States	Size
chemical element	H, C, N, O, S, F, P, Cl, Br, I	10
calculated formal charge	−2, −1, 0, 1, 2	5
CIP configuration	R, S, None, either	4
hybridization state	sp, sp2, sp3, sp3d, sp3d2, none	6
amide center	yes, no	2
present in aromatic ring	yes, no	2
ring size	0, 3, 4, 5, 6, 7, 8, 9, 10, 11	1
nr. of hydrogens	0, 1, 2, 3, 4, 5	6
bond order	1, 2, 3	3
conjugated	-	-
rotate-able	-	-
amide bond	-	-
present in aromatic ring	-	-
present in ring	-	-
ring size	-	-
CIP configuration	-	-

and

Table A6. Node and edge featurization of type 8. Each node or edge featurization vector consisted of a concatenation of the different one-hot encoded or floating point feature vectors according to their possible states (if present) and corresponding size-26 for nodes and 7 for edges.

Feature	Possible States	Size
chemical element	H, C, N, O, S, F, P, Cl, Br, I	10
formal charge	−2, −1, 0, 1, 2	5
CIP priority rule	R, S, None, either	4
hybridization state	sp, sp2, sp3, sp3d, sp3d2, none	6
amide center	-	-
aromaticity	-	-
ring size	float (1/size)	1
nr. of hydrogens	-	-
bond order	1, 2, 3	3
conjugated	-	-
rotate-able	yes, no	2
amide bond	-	-
aromaticity	-	-
present in ring	yes, no	2
ring size	-	-
CIP priority rule	-	-

in the Appendix A. The feature vectors for the non-GNN models consist of 8 different settings-fingerprints (ECFP or MACCSKeys-shown in

Table A7. Non-GNN featurization. The identifier is used as reference.

Identifier	Fingerprint	Radius	nr. Bits	Descriptor
10	ECFP	4	1024	No
11	ECFP	4	1536	No
12	ECFP	4	2048	No
13	MACCSKeys	-	-	No
14	ECFP	4	1024	Yes
15	ECFP	4	1536	Yes
16	ECFP	4	2048	Yes
17	MACCSKeys	-	-	Yes

in the Appendix A) used either in combination with standardized RDKit [29] descriptors or without the descriptors. The descriptors were a combination of all possible and standardized RDKit descriptors, which had a total length of 208. The parametrization of the ECFP was either 1024, 1536, or 2048 bits with a radius of 4. Featurization 3 (Table A1 in the Appendix A) and 4 (Table A2 in the Appendix A) only differ in the way the size of ring systems are being represented. Either as a float value calculated by 1 divided by the size of the ring or as a one-hot encoding with 10 possibilities. The node and edge featurization in 5 (Table A3 in the Appendix A) includes two node features (chemical element and formal charge) and one edge feature (bond order). Featurization 6 (Table A4 in the Appendix A) includes the same node description as 5 and the edge featurization of 3. Featurization 7 (Table A5 in the Appendix A) has the same node featurization as 3 and the same edge featurization as 5. Featurization 8 (Table A6 in the Appendix A) includes a set of optimized node and edge features. This was performed by using a trained D-GIN model and then removing one node or edge feature at a time and observing the RMSE of the prediction. The five node features and the three edge features that had the biggest impact on the RMSE were then taken as the featurization. The graphs and their featurization were implemented using python version 3.7.8 and the toolkit CDPKit [30].

2.5. Directed-Edge GIN (D-GIN) and Reference Models

D-GIN is an extension of the directed-edge message passing neural network of Yang et al. [1] without the additional feature engineering in combination with the graph isomorphic network (GIN) of Xu et al. [18]. Its high level representation can be seen in Figure 1. The principle construction of the network can be seen in the Equations (1)–(8). First, the directed edges were initialized as

$$h_{uw}^0=\tau \left(W_{init}\left(cat(x_u,x^e{}_{uw})\right)\right)$$

(1)

followed by a $t\in 1,\dots ,T$ iteration of

$$m_{uw}^{(t+1)}=\left\{\begin{array}{cc}{\displaystyle \sum _{k\in N\left(u\right)/w}}{h}_{ku}^0,\hfill & \mathrm{if}t==0.\hfill \\ {\displaystyle \sum _{k\in N\left(u\right)/w}}{h}_{ku}^t,\hfill & \mathrm{otherwise}.\hfill \end{array}\right.$$

(2)

$$h_{uw}^{(t+1)}=\tau (h_{uw}^0+W_m{m}_{uw}^{t+1})$$

(3)

after which the messages for each directed-edge was being summed as

$$m_u=\sum _{w\in N\left(u\right)}{h}_{uw}^T$$

(4)

then the message $m_u$ was being concatenated as

$$\begin{array}{c}\hfill h_u=\left\{\begin{array}{cc}cat(m_u,x_u),\hfill & \mathrm{if}\mathrm{D}-\mathrm{GIN}.\hfill \\ (W_{agg}\left(cat(m_u,x_u)\right),\hfill & \mathrm{if}\mathrm{D}-\mathrm{MPNN}.\hfill \end{array}\right.\end{array}$$

(5)

and another message passing over $l\in 1,\dots ,T_2$ was performed by

$$h_u^{\left(l\right)}=\left\{\begin{array}{cc}{\displaystyle \sum _{w\in N\left(u\right)}}{h}_w,\hfill & \mathrm{if}\mathrm{D}-\mathrm{GIN}.\hfill \\ x_u,\hfill & \mathrm{if}\mathrm{GIN}.\hfill \end{array}\right.$$

(6)

$$\begin{array}{cc}\hfill & h_u^{(l+1)}=(W_{agg}(1+ϵ)h_u^0+h_u^{\left(l\right)})\hfill \end{array}$$

(7)

afterwards the updated feature vectors $h_{node}^T$ of each node were aggregated over the whole molecule as

$$\begin{array}{cc}\hfill & h_G=\sum _{h\in \left(H^{\left(T\right)}\right)}h.\hfill \end{array}$$

(8)

The readout phase was then defined as $\widehat{y}=f\left(h_G\right)$ where $f(·)$ was a feed-forward neural network. The D-MPNN consisted of Equations (1)–(5) but then used the hidden feature vectors for each node directly by applying Equation (5) and then immediately Equation (7) to encode the whole graph as $h_G$.

GIN on the other hand was initialized and trained, as shown in Equation (6) in order to update the hidden feature vectors of each node. After l update step, the hidden feature vector of each node served as the input of Equation (7) to achieve the aggregated representation $h_G$ for the whole graph. D-GIN used all of these functions in a combined way described above Equations (1)–(8). The main principle behind this approach was to first use the key aspect of directed-edge message passing to propagate information via directed-edges to form messages Equations (1)–(4), which then updated the hidden node features Equations (5). These updated hidden node features were then used in the GIN message passing to further propagate information Equations (6) and (7) while also learning $ϵ$. These two information propagation phases are the key aspects of the two different sub-architectures.

2.6. Graph Neural Network Implementation, Training, and Hyper-Parameter Search

All GNNs have been implemented and trained using python version 3.7.8 and TensorFlow 2.3.0 [31]. We used TensorFlow’s keras models as our super-class and then transferred Equations (1)–(8) into the “fit” method of the keras model. A hyper-parameter search was conducted to find the best parameters which were further used to train all models. Further details on the hyper-parameters are given in the corresponding model’s configuration files accessible via the graph_networks Github repository [28].

Each GNN model type was trained twice with either 24 different settings when training on the logD or logS property or 12 on the logP property-in total 48 or 24 training runs per model type were performed. Each non-GNN model type was trained with 8 different settings. For training, evaluation, and testing we split each of the datasets as described in Section Experimental Data. Each of the GNNs were trained for 1600 epochs and the model with the best performance was identified using RMSE as the evaluation metric on the validation set. To evaluate the model type performance, we used the model with the best RMSE of the two runs performed for each model setting. When evaluating the average model type performance, the average RMSE of the different model settings was used for the calculation. To evaluate models with several properties, we summed all RMSEs. For example, when using logD and logP for training, we summed the RMSE of the logD and logP prediction on the evaluation set to receive a combined RMSE. When the combined RMSE was below the last best combined RMSE, the model weights were saved. We used these models to test the model on the test set. Each model was run two times and the results with the best test set performance were taken.

Additionally, the 95% confidence interval range was calculated by applying bootstrapping 100 times while leaving out 10% of the test dataset.

To generate consensus models between GNN and non-GNN models, we combined the best GNN model for each physicochemical property with the best non-GNN model. We did this by adding the predicted log values of one model with the other and then divided it by two. These hybrid models are then called according to their GNN model type plus consensus (e.g., D-GIN cons.).

2.7. Other Machine Learning Approaches

We used the random forest (RF), support vector machine (SVM), and k-nearest neighbor (K-NN) implementations of scikit-learn (Version 0.23.2 [32]). Default hyperparameters were used. The featurization is described in Table A7. When using descriptors as input, we standardized them with the scikit-learn StandardScaler. For the fingerprints and descriptors we used version 2020.09.2 of the RDKit [29] python package. Each of the models were trained in a single-task manner for each of the property values.

2.8. Hardware and Run-Time

Calculations were performed on machines within the Department of Pharmaceutical Sciences at the University of Vienna, Austria. We ran each model on a single CPU (Intel(R) Core(TM) i7-8700K CPU @ 3.70GHz). The run-time to fit the used RF, SVM, and KNN models with 3380 compounds on logD property values is approximately 50 s (RF), 25 s (SVM), and 0.5 s (KNN). When training the GNN model types on the 3380 logD compounds it takes for each epoch approximately 56 s (D-GIN), 35 s (D-MPNN), and 28 s (GIN).

3. Results and Discussion

For clarity, we define certain terms used throughout this publication that might have ambiguous meanings. The term “model type” refers to different kinds of machine learning algorithms. For example, a model type can be RF, SVM, KNN, D-GIN, GIN, or D-MPNN. The term “model” refers to a trained model instance with particular training and featurization strategies. The term “training strategy” is used to distinguish between different single- and multi-task training approaches trained with a combination of molecular properties. For example, logD/S/P is used to show that logD, logS, and logP values were used during training. The term “featurization strategy” is used to describe the different node and edge features utilized for the models to train on (Table A1, Table A2, Table A3, Table A4, Table A5 and Table A6 in the Appendix A). In addition, we distinguish between consensus (cons.) and non-consensus models. These hybrid models are a combination of the best GNN and best non-GNN models (SVM and D-GIN for logD and logS and RF and D-GIN for logP). To obtain consensus predictions, the predicted property values of the two models were combined and averaged. The averaged values were used as “new” predictions for the RMSE calculation and referred to their GNN model type plus cons (e.g., D-GIN cons).

Overall, 6 different machine learning model types were used in this study. The three GNN model types were D-MPNN, GIN, and D-GIN. The three non-GNN model types were random forest (RF) regression, support vector machines (SVM), and the k-nearest-neighbor (KNN) algorithm. Each model type was trained with the same hyperparameters, but 7 different learning strategies and 6 different node/edge featurization strategies. We trained each GNN model type for each physico-chemical property with all possible strategies twice. Subsequently, the best performing model from each of the two runs (measured on the evaluation set) was selected resulting in 24 models for the logD and logS property and 12 for the logP property, which were then used on the test set and their performance was reported.

The results of this approach are reported and discussed in two parts. First, we discuss different GNNs and non-GNN methods used in this work to identify the best performing model type according to its average performance across all used strategies (discussed in Section General Model Performance). Subsequently, we investigate the impact of the 6 different training strategies (i.e., multi-task vs. single task learning), as well as different featurizations on the performance (discussed in Sections Impact of Molecular Featurization and Impact of Training Strategies).

A dataset of 10,617 molecular structures with annotations for one of the three physico-chemical properties was assembled for model training, evaluation, and testing. It included 4174 logD, 6443 logS, and 10,617 logP experimentally measured values. The same training, evaluation, test set was used for all GNN and non-GNN model types.

3.1. General Model Performance

In the following, the reported results vary by the used model type. Each combination of featurization and training strategy was used to calculate a total of 24 RMSE values for the logD and logS property, and 12 for the logP property per model type. This resulted in a RMSE distribution shown in

Table 2. Overview of the best performing machine learning model types independent of training and featurization strategy for prediction of logD, logS, and logP. The performance was calculated as the distribution average over all used model root mean squared error (RMSE) values. In total 24 models were used for the loD and logS property, and 12 for the logP property. RMSE values highlighted in dark and light gray show the best and next best models. Red asterisks mark the lowest RMSE for the non-consensus models for each property prediction.

Molecular	Model Type	Mean	Min	Max
Property		RMSE	RMSE	RMSE
logD	D-GIN	0.615 ± 0.039 *	0.553	0.704
	D-MPNN	0.762 ± 0.065	0.686	0.911
	GIN	0.804 ± 0.061	0.738	0.911
	RF	0.780 ± 0.084	0.699	0.890
	SVM	0.740 ± 0.068	0.639	0.814
	KNN	0.951 ± 0.067	0.801	1.003
	D-GIN cons.	0.575 ± 0.019	0.548	0.622
	D-MPNN cons.	0.647 ± 0.028	0.613	0.710
	GIN cons.	0.666 ± 0.029	0.627	0.719
logS	D-GIN	0.867 ± 0.070 *	0.795	1.061
	D-MPNN	0.896 ± 0.030	0.857	0.961
	GIN	1.210 ± 0.102	1.088	1.400
	RF	0.997 ± 0.253	0.760	1.284
	SVM	1.006 ± 0.154	0.729	1.162
	KNN	1.500 ± 0.217	1.057	1.676
	D-GIN cons.	0.738 ± 0.028	0.705	0.820
	D-MPNN cons.	0.762 ± 0.012	0.743	0.785
	GIN cons.	0.881 ± 0.045	0.825	0.969
logP	D-GIN	0.529 ± 0.064*	0.472	0.662
	D-MPNN	0.600 ± 0.063	0.540	0.734
	GIN	0.784 ± 0.077	0.716	0.901
	RF	0.681 ± 0.224	0.470	0.928
	SVM	0.693 ± 0.134	0.493	0.833
	KNN	1.014 ± 0.123	0.743	1.102
	D-GIN cons.	0.455 ± 0.028	0.428	0.515
	D-MPNN cons.	0.475 ± 0.027	0.443	0.532
	GIN cons.	0.566 ± 0.034	0.533	0.618

and Figure 2. For each of these distributions, the average, minimum, and maximum RMSE was calculated and will be reported and discussed subsequently.

Table 2 shows the RMSE distribution average of the different machine learning model types regardless of their training and featurization strategy on the hold-out test set. For each value the standard error of the mean was calculated and added.

For logD property prediction, the D-GIN model type performed with mean, minimum, and maximum logD RMSE of 0.615 ± 0.039, 0.553, and 0.7048, and the corresponding consensus model with 0.575 ± 0.0192, 0.548, and 0.622, making it the best performing model type (results shown in Table 2, and Figure 2). The consensus GIN performed on average (distribution mean of logD RMSE values of 0.666 ± 0.029) better than the best non-GNN method (distribution mean logD RMSE of 0.740 ± 0.068).

For the logS prediction, the best model type was the D-GIN consensus model with a average RMSE value of 0.738 ± 0.028 (shown in Table 2 and Figure 2). It performed on average better than the best performing non-GNN model type (SVM), which performed with an average RMSE value of 1.006 ± 0.154 (but it also had a single run with a RMSE value of 0.729 making it the model type with the best single run performance and highlighting the importance of multiple repetitions for reporting model type performances). The consensus D-MPNN also outperformed the D-GIN.

The consensus D-GIN (average RMSE value of 0.455 ± 0.028) and consensus D-MPNN (average RMSE value of 0.475 ± 0.027) showed the best average performance for logP prediction (Table 2, and Figure 2). The RF and SVM model types also performed with low minimum RMSE values of 0.470 and 0.493, respectively. However, their average RMSE values (RF: 0.681 ± 0.224 and SVM: 0.693 ± 0.134) were higher than the D-GIN and D-MPNN model types.

Consensus models are often used in deep learning applications typically combining either different models that were trained on slightly different training data or multiple model architectures with different strengths and weaknesses. Nevertheless, further investigations are required to give a rationale of why in all our invested cases, the consensus models performed better than their individual counterparts. Furthermore, it should be noted that a direct comparison between the average performance of the GNNs and non-GNN models (RF, SVM, and KNN) can be difficult since the amount of information about a single molecule fed to each of the different model classes is quite different. For example, the non-GNN methods used a wide range of different descriptors and fingerprints shown in Table A7.

Figure 3, Figure 4 and Figure 5 show the best performing model architectures for prediction of each physicochemical property. Each plot shows the RMSE values for each GNN model applying all training and featurization strategies. It should be noted that the performance of many model types with different training or features do not significantly differ from each other and their CI overlap. Some trends are still visible: in Figure 3, Figure 4 and Figure 5, regardless of the physicochemical property, the D-GIN model type (shown in blue) performs overall better than the D-MPNN (shown in orange) or the GIN (shown in green).

The reason why the D-GIN outperforms the GIN and D-MPNN could be its higher complexity and network depth. It uses the key aspects of both sub-models and might be able to better abstract higher-order features. This could be facilitated by including skip connections between edge feature extraction mainly performed in the first (D-MPNN) and node feature extraction while learning $ϵ$ in the second (GIN) part. This increased complexity could have helped to perform better than its individual parts.

3.2. Impact of Molecular Featurization

The average performance of each featurization strategy across all model types and training strategies is shown in

Table 3. Impact of the featurization and training strategy on the different molecular properties independent of what GNN model type was used. For each endpoint the mean, minimum, and maximum RMSE can be seen. The dark gray boxes show the best RMSE for the particular property, the light gray the second best.

Featurization		logD RMSE			logS RMSE			logP RMSE
Strategy	Mean	Min	Max	Mean	Min	Max	Mean	Min	Max
3.0	0.689 ± 0.079	0.575	0.781	0.954 ± 0.146	0.831	1.177	0.596 ± 0.120	0.472	0.751
4.0	0.694 ± 0.072	0.587	0.791	0.948 ± 0.142	0.807	1.160	0.591 ± 0.105	0.487	0.725
5.0	0.813 ± 0.099	0.660	0.911	1.099 ± 0.180	0.938	1.361	0.760 ± 0.110	0.652	0.901
6.0	0.727 ± 0.132	0.553	0.900	1.015 ± 0.250	0.818	1.400	0.641 ± 0.183	0.493	0.883
7.0	0.732 ± 0.075	0.615	0.823	0.961 ± 0.113	0.851	1.144	0.629 ± 0.077	0.541	0.727
8.0	0.706 ± 0.083	0.579	0.814	0.970 ± 0.155	0.795	1.216	0.609 ± 0.114	0.477	0.750
Training		logD RMSE			logS RMSE			logP RMSE
Strategy	Mean	Min	Max	Mean	Min	Max	Mean	Min	Max
logD/P/S	0.730 ± 0.102	0.579	0.900	0.979 ± 0.166	0.795	1.325	0.639 ± 0.129	0.487	0.901
logD/P	0.719 ± 0.105	0.553	0.911	-	-	-	-	-	-
logD/S	0.737 ± 0.106	0.582	0.911	0.993 ± 0.176	0.821	1.359	-	-	-
logS/P	-	-	-	0.988 ± 0.173	0.812	1.333	-	-	-
logD	0.722 ± 0.087	0.596	0.881	-	-	-	-	-	-
logS	-	-	-	1.004 ± 0.187	0.818	1.400	-	-	-
logP	-	-	-	-	-	-	0.637 ± 0.130	0.472	0.894

. Considering the performance for all physicochemical properties, featurization strategy 5 showed the highest RMSE (mean logD/logS/logP RMSE of 0.813 ± 0.099, 1.099 ± 0.180, and 0.760 ± 0.110). This trend was also observed when separating according to the model type (shown in

Table 4. Impact of the featurization on the different molecular properties. For each property and model type, the mean, minimum, and maximum RMSE are shown. The dark gray boxes represent the best RMSE for the particular property and model type, the light gray the second best. The red asterisk is the overall best RMSE for the particular property.

Model Type	Featurization		logD RMSE			logS RMSE			logP RMSE
	Strategy	Mean	Min	Max	Mean	Min	Max	Mean	Min	Max
D-GIN	3.0	0.585 ± 0.011	0.575	0.601	0.835 ± 0.007	0.831	0.847	0.484 ± 0.017 *	0.472	0.496
	4.0	0.603 ± 0.014	0.587	0.622	0.827 ± 0.018	0.807	0.851	0.490 ± 0.004	0.487	0.493
	5.0	0.682 ± 0.018	0.660	0.704	1.007 ± 0.039	0.969	1.061	0.657 ± 0.006	0.652	0.662
	6.0	0.580 ± 0.019 *	0.553	0.596	0.825 ± 0.008 *	0.818	0.836	0.502 ± 0.012	0.493	0.511
	7.0	0.637 ± 0.020	0.615	0.661	0.880 ± 0.025	0.851	0.906	0.550 ± 0.013	0.541	0.560
	8.0	0.602 ± 0.020	0.579	0.629	0.826 ± 0.027	0.795	0.854	0.493 ± 0.023	0.477	0.509
D-MPNN	3.0	0.728 ± 0.023	0.703	0.759	0.835 ± 0.007	0.831	0.847	0.561 ± 0.012	0.552	0.570
	4.0	0.715 ± 0.016	0.692	0.728	0.879 ± 0.017	0.862	0.896	0.563 ± 0.010	0.556	0.570
	5.0	0.878 ± 0.023	0.857	0.911	0.951 ± 0.011	0.938	0.961	0.725 ± 0.012	0.716	0.734
	6.0	0.712 ± 0.017	0.686	0.724	0.868 ± 0.011	0.857	0.879	0.545 ± 0.007	0.540	0.551
	7.0	0.805 ± 0.016	0.783	0.823	0.892 ± 0.018	0.865	0.905	0.616 ± 0.017	0.604	0.629
	8.0	0.734 ± 0.036	0.712	0.788	0.911 ± 0.007	0.905	0.921	0.588 ± 0.007	0.581	0.593
GIN	3.0	0.755 ± 0.018	0.741	0.781	1.149 ± 0.022	1.122	1.177	0.745 ± 0.009	0.738	0.751
	4.0	0.765 ± 0.019	0.743	0.791	1.137 ± 0.020	1.116	1.160	0.720 ± 0.006	0.716	0.725
	5.0	0.880 ± 0.026	0.856	0.911	1.339 ± 0.017	1.325	1.361	0.897 ± 0.005	0.894	0.901
	6.0	0.889 ± 0.019	0.860	0.900	1.351 ± 0.037	1.314	1.400	0.877 ± 0.008	0.871	0.883
	7.0	0.755 ± 0.007	0.747	0.765	1.112 ± 0.028	1.088	1.144	0.722 ± 0.007	0.716	0.727
	8.0	0.781 ± 0.033	0.734	0.814	1.172 ± 0.033	1.135	1.216	0.745 ± 0.006	0.741	0.750

and Figure 6). The reason for the relatively bad performance of featurization 5 might be that it only included two node properties (chemical element and formal charge), as well as only a single edge feature (bond order-Table A3 in the Appendix A).

Featurization 6 (Table A4 in the Appendix A) also displayed considerably worse performance than other strategies when used in combination with the GIN architecture, for which the mean RMSE performance for logD and logS properties were worse than using featurization strategy 5. One explanation could be that the GIN utilizes node features quite extensively and featurization 6 only included two node feature types similar to featurization 5. The additional edge features in strategy 6 without the appropriate architecture to deal with them could push the optimizer of the GIN network into the wrong direction rather than help with the property prediction.

Although it is easy to identify bad featurization strategies, it is difficult to come up with an unambiguous recommendation for the best performing featurization strategy. The mean RMSE across all training strategies and model types in Table 3 show that featurization 3 and 4 (Table A1 and Table A2 in the Appendix A) achieved very good results for logD with a RMSE value of 0.689 ± 0.079 and 0.694 ± 0.072, for logS with a RMSE 0.954 ± 0.146 and 0.948 ± 0.142 and for logP with a RMSE 0.596 ± 0.120 and 0.591 ± 0.105, respectively. Both featurization strategies utilize the maximum number of node and edge features used in this work. They only differ in the way molecular ring sizes are described. Featurization 3 used a float value calculated by 1 divided by the size of the ring system whereas featurization 4 used a one-hot encoding of ten instances (0,3,4,5,6,7,8,9,10,11).

Table 4 shows the mean RMSE values concerning featurization and model type. As performance criteria for featurization strategies we used the sum of model ranks in Table 4. Applying this approach, featurization 3 with two models as best and three models as second-best performers achieved a better ranking than featurization 4 with one model ranked best and two models as second best. Both strategies perform similarly well. Featurization 8 (shown in Table A6 in the Appendix A) used a set of optimized node and edge features. Node and edge features were optimized by masking single edge and node features at a time and evaluating their impact on the test set RMSE. The five node features and the three edge features that had the biggest impact on the RMSE were subsequently used. This approach also revealed that the size of ring systems for the node features appears to be of importance and was, therefore, included in 8. Using featurization 8, we were able to achieve two times the second-best performance. It shows an average good performance, but not as good as featurizations 3 or 4, even though its edge and node features were selected for maximum impact on the final prediction. The mean RMSE of featurization 6 and 7 (Table A5 in the Appendix A) in Table 3 show diminished results compared to featurization 3 and 4.

When evaluating the rank score, the featurization strategy that performs either best or second-best for each physicochemical property, the best featurization strategy was number 6. It was used in four of the best performing runs and once in a second-best run. However, it only performed well in combination with two GNN architectures (D-GIN and D-MPNN) and strongly underperformed with the GIN. The D-GIN and D-MPNN architecture types use primarily edge features for their information propagation and featurization strategy 6 provided these. It utilized only two-node feature types, potentially reducing the noise for the feature extraction to a minimum in this setting.

On average, featurization strategies 6 and 7 performed similarly well. However, when separating the results at a model type level, it became evident that there was a strong model architecture dependency, so it seems important to choose the features according to the architecture at hand (Figure 3, Figure 4 and Figure 5). Furthermore, featurization 3 might perform worse than featurization 6 or 7. Nevertheless, when unsure which features to use, simply adding more features could be the safer option rather than using less. This observation is also supported by comparing featurization 3 or 4 to, e.g., 6, 7, or 8.

When analyzing the results for the non-GNN models and their different featurizations, the mean RMSE variance was large in comparison to the GNN models. Moreover, in similar deep-learning benchmark studies that predicted molecular properties, predominantly fingerprints have been used. From

Table A13. Shows the logS RMSE and $r^2$ results for the D-MPNN model type used during this survey. The last column consist of the unique identify. Training strategy 0 represents a training strategy combining logD, logS, and logP, 1 represents a strategy combining logD and logP, 2 stands for a combination of logD and logS, 3 represents a strategy using logS and logP, 4 represents a strategy using logD, 5 represents a strategy using logS, and 6 represents a strategy using logP.

Model	Training	Featurization	logD	${\mathit{r}}^2$	Unique
Type	Strategy	Strategy	RMSE		ID
D-MPNN	3	8	0.913 ± 0.052	0.756 ± 0.026	dmp832
D-MPNN	2	6	0.857 ± 0.035	0.782 ± 0.015	dmp621
D-MPNN	2	3	0.865 ± 0.038	0.784 ± 0.016	dmp322
D-MPNN	0	5	0.962 ± 0.038	0.718 ± 0.018	dmp501
D-MPNN	0	8	0.907 ± 0.049	0.757 ± 0.021	dmp802
D-MPNN	5	8	0.921 ± 0.046	0.751 ± 0.025	dmp851
D-MPNN	2	8	0.906 ± 0.041	0.763 ± 0.020	dmp821
D-MPNN	0	7	0.906 ± 0.049	0.764 ± 0.027	dmp701
D-MPNN	2	7	0.902 ± 0.040	0.772 ± 0.021	dmp722
D-MPNN	3	5	0.938 ± 0.044	0.733 ± 0.029	dmp532
D-MPNN	5	5	0.945 ± 0.041	0.734 ± 0.030	dmp552
D-MPNN	3	4	0.896 ± 0.062	0.782 ± 0.028	dmp432
D-MPNN	0	4	0.893 ± 0.040	0.762 ± 0.022	dmp401
D-MPNN	3	3	0.893 ± 0.054	0.775 ± 0.026	dmp332
D-MPNN	5	6	0.879 ± 0.046	0.777 ± 0.021	dmp651
D-MPNN	5	3	0.878 ± 0.053	0.772 ± 0.027	dmp351
D-MPNN	0	6	0.877 ± 0.039	0.780 ± 0.021	dmp601
D-MPNN	0	3	0.874 ± 0.046	0.779 ± 0.024	dmp302
D-MPNN	2	5	0.961 ± 0.036	0.721 ± 0.021	dmp521
D-MPNN	5	4	0.866 ± 0.054	0.778 ± 0.022	dmp452
D-MPNN	3	7	0.865 ± 0.049	0.773 ± 0.027	dmp731
D-MPNN	2	4	0.863 ± 0.046	0.800 ± 0.020	dmp422
D-MPNN	5	7	0.897 ± 0.050	0.763 ± 0.024	dmp751
D-MPNN	3	6	0.861 ± 0.047	0.780 ± 0.024	dmp632
D-MPNN cons.	5	4	0.748 ± 0.055	0.830 ± 0.020	dmp452_cons
D-MPNN cons.	0	4	0.768 ± 0.040	0.819 ± 0.015	dmp401_cons
D-MPNN cons.	2	8	0.769 ± 0.039	0.819 ± 0.014	dmp821_cons
D-MPNN cons.	3	8	0.762 ± 0.052	0.823 ± 0.022	dmp832_cons
D-MPNN cons.	5	6	0.749 ± 0.053	0.829 ± 0.021	dmp652_cons
D-MPNN cons.	0	8	0.770 ± 0.044	0.818 ± 0.016	dmp802_cons
D-MPNN cons.	3	3	0.762 ± 0.052	0.823 ± 0.021	dmp332_cons
D-MPNN cons.	0	7	0.771 ± 0.044	0.818 ± 0.016	dmp701_cons
D-MPNN cons.	2	7	0.771 ± 0.040	0.818 ± 0.016	dmp722_cons
D-MPNN cons.	5	7	0.761 ± 0.051	0.824 ± 0.021	dmp751_cons
D-MPNN cons.	5	8	0.774 ± 0.048	0.817 ± 0.022	dmp851_cons
D-MPNN cons.	3	5	0.777 ± 0.048	0.816 ± 0.021	dmp532_cons
D-MPNN cons.	5	5	0.779 ± 0.045	0.815 ± 0.021	dmp552_cons
D-MPNN cons.	3	4	0.765 ± 0.056	0.822 ± 0.021	dmp431_cons
D-MPNN cons.	2	5	0.784 ± 0.038	0.811 ± 0.016	dmp521_cons
D-MPNN cons.	0	3	0.760 ± 0.042	0.823 ± 0.016	dmp302_cons
D-MPNN cons.	0	6	0.759 ± 0.039	0.823 ± 0.014	dmp601_cons
D-MPNN cons.	2	3	0.756 ± 0.039	0.825 ± 0.014	dmp322_cons
D-MPNN cons.	2	6	0.744 ± 0.038	0.831 ± 0.013	dmp621_cons
D-MPNN cons.	5	3	0.752 ± 0.055	0.828 ± 0.021	dmp351_cons
D-MPNN cons.	3	7	0.744 ± 0.053	0.831 ± 0.021	dmp731_cons
D-MPNN cons.	2	4	0.750 ± 0.044	0.828 ± 0.016	dmp422_cons
D-MPNN cons.	3	6	0.748 ± 0.055	0.830 ± 0.021	dmp632_cons
D-MPNN cons.	0	5	0.785 ± 0.040	0.811 ± 0.015	dmp501_cons

,

Table A14. Shows the logS RMSE and $r^2$ results for the GIN model type used during this survey. The last column consists of the unique identifier. Training strategy 0 represents a training strategy combining logD, logS, and logP, 1 represents a strategy combining logD and logP, 2 stands for a combination of logD and logS, 3 represents a strategy using logS and logP, 4 represents a strategy using logD, 5 represents a strategy using logS, and 6 represents a strategy using logP.

Model	Training	Featurization	logD	${\mathit{r}}^2$	Unique
Type	Strategy	Strategy	RMSE		ID
GIN	2	4	1.126 ± 0.047	0.612 ± 0.029	g421
GIN	0	3	1.122 ± 0.049	0.615 ± 0.028	g301
GIN	0	4	1.116 ± 0.053	0.624 ± 0.037	g402
GIN	0	7	1.089 ± 0.044	0.650 ± 0.031	g701
GIN	3	7	1.088 ± 0.066	0.643 ± 0.036	g731
GIN	2	7	1.127 ± 0.049	0.613 ± 0.028	g721
GIN	3	4	1.147 ± 0.084	0.602 ± 0.052	g431
GIN	5	7	1.145 ± 0.060	0.602 ± 0.038	g752
GIN	5	6	1.400 ± 0.064	0.407 ± 0.045	g652
GIN	5	5	1.362 ± 0.061	0.437 ± 0.043	g552
GIN	2	6	1.359 ± 0.045	0.441 ± 0.036	g621
GIN	2	5	1.345 ± 0.044	0.447 ± 0.030	g522
GIN	3	6	1.334 ± 0.058	0.463 ± 0.047	g631
GIN	0	5	1.326 ± 0.049	0.474 ± 0.031	g502
GIN	0	8	1.136 ± 0.058	0.605 ± 0.034	g801
GIN	3	5	1.326 ± 0.066	0.472 ± 0.049	g532
GIN	5	8	1.217 ± 0.062	0.554 ± 0.040	g851
GIN	5	3	1.178 ± 0.070	0.579 ± 0.042	g352
GIN	2	8	1.173 ± 0.050	0.580 ± 0.027	g821
GIN	3	8	1.166 ± 0.059	0.603 ± 0.036	g832
GIN	5	4	1.161 ± 0.089	0.598 ± 0.057	g452
GIN	2	3	1.151 ± 0.047	0.595 ± 0.025	g321
GIN	3	3	1.148 ± 0.064	0.610 ± 0.036	g331
GIN	0	6	1.314 ± 0.053	0.474 ± 0.034	g602
GINcons.	5	6	0.969 ± 0.057	0.714 ± 0.029	g652_cons
GINcons.	2	7	0.840 ± 0.036	0.784 ± 0.014	g721_cons
GINcons.	5	5	0.945 ± 0.045	0.728 ± 0.023	g552_cons
GINcons.	0	3	0.842 ± 0.041	0.783 ± 0.016	g301_cons
GINcons.	5	4	0.861 ± 0.065	0.774 ± 0.031	g452_cons
GINcons.	2	8	0.867 ± 0.043	0.770 ± 0.017	g821_cons
GINcons.	5	3	0.872 ± 0.056	0.769 ± 0.026	g352_cons
GINcons.	3	3	0.857 ± 0.054	0.776 ± 0.024	g331_cons
GINcons.	0	7	0.825 ± 0.038	0.791 ± 0.016	g701_cons
GINcons.	5	8	0.887 ± 0.056	0.760 ± 0.026	g851_cons
GINcons.	2	3	0.855 ± 0.043	0.776 ± 0.015	g321_cons
GINcons.	5	7	0.854 ± 0.055	0.778 ± 0.026	g752_cons
GINcons.	3	7	0.828 ± 0.052	0.791 ± 0.023	g731_cons
GINcons.	0	4	0.835 ± 0.040	0.786 ± 0.016	g402_cons
GINcons.	3	4	0.860 ± 0.061	0.775 ± 0.029	g431_cons
GINcons.	2	4	0.835 ± 0.041	0.786 ± 0.014	g421_cons
GINcons.	0	6	0.926 ± 0.045	0.737 ± 0.017	g602_cons
GINcons.	0	5	0.933 ± 0.038	0.733 ± 0.015	g502_cons
GINcons.	3	5	0.934 ± 0.053	0.734 ± 0.025	g532_cons
GINcons.	2	5	0.938 ± 0.040	0.730 ± 0.015	g522_cons
GINcons.	3	6	0.940 ± 0.048	0.731 ± 0.025	g631_cons
GINcons.	2	6	0.945 ± 0.041	0.726 ± 0.017	g621_cons
GINcons.	0	8	0.850 ± 0.042	0.779 ± 0.014	g802_cons
GINcons.	3	8	0.861 ± 0.053	0.774 ± 0.018	g832_cons

and

Table A15. Shows the logS RMSE and $r^2$ results for the non-GNN model types used during this survey. The last column consists of the unique identifier. Training strategy 0 represents a training strategy combining logD, logS, and logP, 1 represents a strategy combining logD and logP, 2 stands for a combination of logD and logS, 3 represents a strategy using logS and logP, 4 represents a strategy using logD, 5 represents a strategy using logS, and 6 represents a strategy using logP.

Model	Training	Featurization	logD	${\mathit{r}}^2$	Unique
Type	Strategy	Strategy	RMSE		ID
KNN	logS	14	1.547 ± 0.063	0.268 ± 0.054	KNN14
KNN	logS	11	1.587 ± 0.063	0.230 ± 0.054	KNN11
KNN	logS	10	1.587 ± 0.064	0.229 ± 0.056	KNN10
KNN	logS	12	1.600 ± 0.066	0.217 ± 0.055	KNN12
KNN	logS	13	1.280 ± 0.066	0.499 ± 0.036	KNN13
KNN	logS	15	1.676 ± 0.066	0.140 ± 0.068	KNN15
KNN	logS	17	1.058 ± 0.055	0.658 ± 0.032	KNN17
KNN	logS	16	1.670 ± 0.065	0.147 ± 0.061	KNN16
RF	logS	12	1.239 ± 0.056	0.530 ± 0.032	rf12
RF	logS	14	0.760 ± 0.043	0.823 ± 0.020	rf14
RF	logS	15	0.764 ± 0.044	0.821 ± 0.022	rf15
RF	logS	13	1.128 ± 0.061	0.611 ± 0.039	rf13
RF	logS	16	0.765 ± 0.045	0.821 ± 0.024	rf16
RF	logS	17	0.770 ± 0.049	0.818 ± 0.025	rf17
RF	logS	10	1.284 ± 0.055	0.495 ± 0.033	rf10
RF	logS	11	1.271 ± 0.057	0.506 ± 0.035	rf11
SVM	logS	12	1.142 ± 0.055	0.601 ± 0.035	svm12
SVM	logS	11	1.149 ± 0.057	0.596 ± 0.035	svm11
SVM	logS	16	0.966 ± 0.049	0.715 ± 0.024	svm16
SVM	logS	15	0.930 ± 0.046	0.735 ± 0.023	svm15
SVM	logS	13	1.086 ± 0.060	0.639 ± 0.034	svm13
SVM	logS	17	0.730 ± 0.041	0.837 ± 0.020	svm17
SVM	logS	14	0.891 ± 0.046	0.757 ± 0.023	svm14
SVM	logS	10	1.162 ± 0.056	0.587 ± 0.036	svm10

in the Appendix A, one can see that especially featurizations that include descriptors in addition to fingerprints perform exceptionally well. We think that when comparing GNN with non-GNN models, differences in used features should be taken into consideration when trying to identify and understand (deep-learning) method performance.

3.3. Impact of Training Strategies

The impact of different training strategies are shown in Table 3. The lowest mean logD RMSE can be obtained by a multi-task strategy that involves learning on both logD and logP values. This is similar to the best training strategy for the logS property, which is a multi-task approach including logS and logP properties. As for the logP property, the best approach is a single-task strategy including logP values, however the multi-task approach which combines all physicochemical properties achieves similarly good performance.

When analyzing the logD/S/P RMSE predictions with respect to training strategy and model type,

Table 5. Impact of the training strategies on the different molecular properties. Each model type is evaluated separately. For each property and model type, the mean, minimum, and maximum RMSE are shown. The dark gray boxes represent the best RMSE for the particular property and model type, the light gray the second best. The red asterisk highlights the overall best RMSE for the particular property.

Model Type	Training		logD RMSE			logS RMSE			logP RMSE
	Strategy	Mean	Min	Max	Mean	Min	Max	Mean	Min	Max
D-GIN	logD/P/S	0.617 ± 0.052	0.579	0.704	0.851 ± 0.071 *	0.795	0.987	0.534 ± 0.067	0.487	0.662
	logD/P	0.607 ± 0.044 *	0.553	0.685	-	-	-	-	-	-
	logD/S	0.610 ± 0.036	0.582	0.679	0.875 ± 0.072	0.821	1.010	-	-	-
	logS/P	-	-	-	0.868 ± 0.095	0.812	1.061	-	-	-
	logD	0.625 ± 0.024	0.596	0.660	-	-	-	-	-	-
	logS	-	-	-	0.872 ± 0.053	0.818	0.969	-	-	-
	logP	-	-	-	-	-	-	0.524 ± 0.067 *	0.472	0.652
D-MPNN	logD/P/S	0.759 ± 0.063	0.712	0.864	0.903 ± 0.031	0.873	0.961	0.596 ± 0.062	0.551	0.716
	logD/P	0.741 ± 0.066	0.686	0.857	-	-	-	-	-	-
	logD/S	0.788 ± 0.070	0.721	0.911	0.892 ± 0.039	0.857	0.960	-	-	-
	logS/P	-	-	-	0.894 ± 0.029	0.860	0.938	-	-	-
	logD	0.761 ± 0.067	0.713	0.881	-	-	-	-	-	-
	logS	-	-	-	0.897 ± 0.030	0.865	0.944	-	-	-
	logP	-	-	-	-	-	-	0.603 ± 0.071	0.540	0.734
GIN	logD/P/S	0.815 ± 0.063	0.756	0.901	1.183 ± 0.106	1.089	1.325	0.786 ± 0.083	0.716	0.901
	logD/P	0.809 ± 0.075	0.742	0.911	-	-	-	-	-	-
	logD/S	0.811 ± 0.056	0.751	0.896	1.213 ± 0.108	1.126	1.359	-	-	-
	logS/P	-	-	-	1.201 ± 0.102	1.088	1.333	-	-	-
	logD	0.780 ± 0.060	0.734	0.860	-	-	-	-	-	-
	logS	-	-	-	1.243 ± 0.109	1.144	1.400	-	-	-
	logP	-	-	-	-	-	-	0.782 ± 0.077	0.725	0.894

and Figure 7, Figure 8 and Figure 9 show that there is no particularly favorable learning strategy for any of the model types. The datasets used in this study are specific for one particular physicochemical property. When comparing different learning strategies we thus focused on one particular physicochemical property for each model type. Starting with the results for the prediction of the logD property in Table 5, we can see that the overall best model (red asterisk), as well as the two best models for each model type (dark gray), are multi-task models. In particular, the models with a combination of logD and logP properties perform well.

Considering all combinations of training and featurizations strategies for each model, the learning strategy with the best average, as well as the best minimum logD RMSE was obtained using the logD/P multi-task training approach resulting in RMSE values of 0.719 ± 0.105 and 0.553, respectively (Table 3. Yet, using this multi-task learning strategy we also obtained single run performance worse than using a single-task learning strategy with only logD values, showcasing once more the importance of validating multiple learning and featurization strategies. The results are similar for the prediction of logS values: again, the multi-task learning strategy performs better than its single task counterpart. The best model for logS prediction was obtained by training on logD, logS, and logP values. Considering all combinations of training and featurization strategies for each model, the best average, minimum, and maximum logS RMSE of 0.979 ± 0.166, 0.795, and 1.325, respectively, was observed during the multi-task training with all properties. We should note here that while it seems that the average performance is improved by multi-task learning, the variance of model performance is also increased.

4. Conclusions

We introduced the directed-edge graph isomorphism network (D-GIN), a novel graph neural network that showed improved average performance for aqueous solubility and lipophilicity prediction compared with other baseline models. We showed that by combining different models with distinct key characteristics, we can increase the depth of the model while also improving its predictive power. Furthermore, applying different training strategies and featurizations constraints enables to obtain more information regarding general, average model performance. This strategy showed that the D-GIN model outperforms other machine-learning models on average and argued that comparing the mean performance rather than single metric values of the best performing model type gives more insight into the general behavior and ultimately facilitates a better understanding and higher robustness of deep-learning models.

In concurrence with previous publications [33,34,35], we showed that there is a tendency towards multi-task learning approaches for the GNNs utilized in this survey. On average they performed better than their single-task counterpart for the corresponding physicochemical property. We could not find clear evidence that more than two properties increase the model’s performance.

Furthermore, we highlighted that the usage of additional features did not improve the GNN model performance. However, we also conclude that very little featurization led to the worst performance. In general it is necessary to be aware of the type of GNN that is used and whether its architecture focuses more on edge or node features. When trying to obtain the best performing model it can be advisable to do feature engineering, but when in doubt which features to use, it can be safer to use more than less. We showed that this awareness can help improve the GNNs predictive power at hand.

For the non-GNN models, we could conclude that by excessively adding descriptors to the molecular fingerprint the performance of these models increases substantially. We further argued that for future comparisons it would be advisable to include not only fingerprints but also descriptors to the non-GNN baseline models to be more competitive.

By combining the best GNN model with the best non-GNN model we could see a slight improvement in the overall performance in all cases. Consensus models have often shown to improve performance. However, in this case, further investigations are needed to attain to a conclusion on why this is the case.

We showed that advanced deep-learning methods such as GNNs do have great potential in the physicochemical property prediction area and, when applied properly, can serve as a promising and robust method for any computer-aided drug discovery pipeline, especially for chemical property prediction.