HFHosein Fooladi

Hosein Fooladi

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Hosein Fooladi has a total of 12 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Top Cited Papers
    Open Access

    Evaluating Machine Learning Models for Molecular Property Prediction: Performance and Robustness on Out-of-Distribution Data

    by Hosein Fooladi, Thi Ngoc Lan Vu, Miriam Mathea, Johannes Kirchmair

    Today, machine learning models are employed extensively to predict the physicochemical and biological properties of molecules. Their performance is typically evaluated on in-distribution (ID) data, i.e., data originating from the same distribution as the training data. However, the real-world applications of such models often involve molecules that are more distant from the training data, necessitating the assessment of their performance on out-of-distribution (OOD) data. In this work, we investigate and evaluate the performance of 14 machine learning models, including classical approaches like random forests, as well as graph neural network (GNN) methods, such as message-passing graph neural networks, across eight data sets using ten splitting strategies for OOD data generation. First, we investigate what constitutes OOD data in the molecular domain for bioactivity and ADMET prediction tasks. In contrast to the common point of view, we show that both classical machine learning and GNN models work well (not substantially different from random splitting) on data split based on Bemis-Murcko scaffolds. Splitting based on chemical similarity clustering (UMAP-based clustering using ECFP4 fingerprints) poses the most challenging task for both types of models. Second, we investigate the extent to which ID and OOD performance have a positive linear relationship. If a positive correlation holds, models with the best performance on the ID data can be selected with the promise of having the best performance on OOD data. We show that the strength of this linear relationship is strongly related to how the OOD data is generated, i.e., which splitting strategies are used for generating OOD data. While the correlation between ID and OOD performance for scaffold splitting is strong (Pearson's r ∼ 0.9), this correlation decreases significantly for all the cluster-based splitting (Pearson's r ∼ 0.4). Therefore, the relationship can be more nuanced, and a strong positive correlation is not guaranteed for all OOD scenarios. These findings suggest that OOD performance evaluation and model selection should be carefully aligned with the intended application domain.


  • Open Access

    Integrating Machine Learning-Based Pose Sampling with Established Scoring Functions for Virtual Screening

    by Thi Ngoc Lan Vu, Hosein Fooladi, Johannes Kirchmair

    Physics-based docking methods have long been the cornerstone of structure-based virtual screening (VS). However, the emergence of machine learning (ML)-based docking approaches has opened new possibilities for enhancing VS technologies. In this study, we explore the integration of DiffDock-L, a leading ML-based pose sampling method, into VS workflows by combining it with the Vina, Gnina, and RTMScore scoring functions. We assess this integrated approach in terms of its VS effectiveness, pose sampling quality, and complementarity to traditional physics-based docking methods, such as AutoDock Vina. Our findings from the DUDE-Z benchmark dataset show that DiffDock-L performs competitively in both VS performance and pose sampling in cross-docking settings. In most cases, it generates physically plausible and biologically relevant poses, establishing itself as a viable alternative to physics-based docking algorithms. Additionally, we found that the choice of scoring function significantly influences VS success.


  • Open Access

    Evaluating Machine Learning Models for Molecular Property Prediction: Performance and Robustness on Out-of-Distribution Data

    by Hosein Fooladi, Thi Ngoc Lan Vu, Johannes Kirchmair

    Today, machine learning models are employed extensively to predict the physicochemical and biological properties of molecules. Their performance is typically evaluated on in-distribution (ID) data, i.e., data originating from the same distribution as the training data. However, the real-world applications of such models often involve molecules that are more distant from the training data, which necessitates assessing their performance on out-of-distribution (OOD) data. In this work, we investigate and evaluate the performance of twelve machine learning models, including classical approaches like random forests, as well as graph neural network (GNN) methods, such as message-passing graph neural networks, across eight data sets using seven splitting strategies for OOD data generation. First, we investigate what constitutes OOD data in the molecular domain for bioactivity and ADMET prediction tasks. In contrast to the common point of view, we show that both classical machine learning and GNN models work well (not substantially different from random splitting) on data split based on Bemis-Murcko scaffolds. Splitting based on chemical similarity clustering (K-means clustering using ECFP4 fingerprints) poses the hardest challenge for both types of models. Second, we investigate the extent to which ID and OOD performance have a positive linear relationship. If a positive correlation holds, models with the best performance on the ID data can be selected with the promise of having the best performance on OOD data. We show that the strength of this linear relationship is strongly related to how the OOD data is generated, i.e., which splitting strategies are used for generating OOD data. While the correlation between ID and OOD performance for scaffold splitting is strong (Pearson r ∼ 0.9), this correlation decreases significantly for cluster-based splitting (Pearson r ∼ 0.4). Therefore, the relationship can be more nuanced, and a strong positive correlation is not guaranteed for all OOD scenarios. These findings suggest that OOD performance evaluation and model selection should be carefully aligned with the intended application domain.


  • Open Access

    Integrating Machine Learning-Based Pose Sampling with Established Scoring Functions for Virtual Screening

    by Thi Ngoc Lan Vu, Hosein Fooladi, Johannes Kirchmair

    Physics-based docking methods have long been the cornerstone of structure-based virtual screening (VS). However, the emergence of machine learning (ML)-based docking approaches has opened up new possibilities for enhancing VS technologies. In this study, we explore the integration of DiffDock-L, a leading ML-based pose sampling method, into VS workflows by combining it with the well-established Vina and Gnina scoring functions. We assess this integrated approach in terms of its VS effectiveness, pose sampling quality, and complementarity to traditional physics-based docking methods, such as AutoDock Vina. Our findings from the DUDEZ benchmark dataset show that DiffDock-L performs competitively in both VS performance and pose sampling in cross-docking settings. In most cases, it generates physically plausible and biologically relevant poses, establishing itself as a viable alternative to physics-based docking algorithms. Additionally, we found that the choice of scoring function significantly influences VS success.

  • Open Access

    Quantifying the hardness of bioactivity prediction tasks for transfer learning

    by Hosein Fooladi, Steffen Hirte, Johannes Kirchmair

    Today, machine learning methods are widely employed in drug discovery. However, the chronic lack of data continues to hamper their further development, validation, and application. Several modern strategies aim to mitigate the challenges associated with data scarcity by learning from data on related tasks. These knowledge-sharing approaches encompass transfer learning, multi-task learning, and meta-learning. A key question remaining to be answered for these approaches is about the extent to which their performance can benefit from the relatedness of available source (training) tasks, in other words, how difficult (“hard”) a test task is to a model, given the available source tasks. This study introduces a new method for quantifying and predicting the hardness of a bioactivity prediction task based on its relation to the available training tasks. The approach involves the generation of protein and chemical representations and the calculation of distances between the bioactivity prediction task and the available training tasks. In the example of meta-learning, we demonstrate that the proposed task hardness metric is inversely correlated with performance. The metric will be useful in estimating the task-specific gain in performance that can be achieved through meta-learning.

  • Open Access

    Encoding Protein-Ligand Interactions: Binding Affinity Prediction with Multigraph-based Modeling and Graph Convolutional Network

    by Nedra Mekni, Hosein Fooladi, Ugo Perricone, Thierry Langer

    Machine learning models are employed to enhance the speed and provide novel insights in drug discovery due to their demonstrated effectiveness in predicting properties of small molecules like pKa, solubility, and binding affinity. These approaches accelerate drug discovery by helping researchers efficiently identify, prioritize, and optimize compounds. Nonetheless, when investigating properties that depend on the interaction between a ligand and its corresponding protein, a compelling need arises to incorporate the protein counterpart information within the models. Recently, graph neural networks (GNNs) have been developed to incorporate 3D structural information to improve our understanding of the underlying protein-ligand interactions. However, incorporating 3D information into GNNs is not always straightforward. To address the challenge, we introduce a model called InterGraph, which models the protein-ligand interaction as topological multigraphs. By leveraging a topological representation, InterGraph offers a comprehensive approach to a graph representation of the intricate spatial organization and connectivity patterns within protein-ligand systems. We introduce interaction spheres that assign varying edge densities, capturing the proximity-based influence of interactions. This approach enables us to capture the characteristics of the interaction network, filtering out the ones that are beyond 9 Å from the ligand since they are not considered relevant or established. Finally, we trained the model using a ligand binding dataset from PDBbind and tested it on a hold-out test set, achieving an RMSE value of 1.34. Our findings have demonstrated the power of the multigraph to encode the importance of close interactions, a factor that is relevant in the context of binding affinity. On average, our model accurately predicts binding affinity values for several protein-ligand complexes and exhibits higher accuracy for hydrolase, lyase, and families of proteins involved in mediating protein-protein interactions. Additionally, the Intergraph method displayed sensitivity to the binding mode when compared to a set of complexes that had undergone redocking calculations


  • Open Access

    Bayesian Optimization for Ternary Complex Prediction (BOTCP)

    by Arjun Rao, Tin M. Tunjic, Michael Brunsteiner, Michael Müller, Hosein Fooladi, Noah Weber

    Proximity-inducing compounds (PICs) are an emergent drug technology through which a protein of interest (POI), often a drug target, is brought into the vicinity of a second protein which modifies the POI’s function, abundance or localisation, giving rise to a therapeutic effect. One of the best-known examples for such compounds are heterobifunctional molecules known as proteolysis targeting chimeras (PROTACs). PROTACs reduce the abundance of the target protein by establishing proximity to an E3 ligase which targets the protein towards degradation via the ubiquitin-proteasomal pathway. Design of PROTACs in silico requires the computational prediction of the ternary complex consisting of POI, PROTAC molecule, and the E3 ligase. Here, we present a novel machine learning-based method for predicting PROTAC-mediated ternary complex structures using Bayesian optimization. We show how a fitness score combining an estimation of protein-protein interactions with PROTAC binding energy calculations enables the sample-efficient exploration of candidate structures. Furthermore, our method presents two novel scores for filtering and reranking which take PROTAC stability (Autodock-Vina based PROTAC stability score) and protein interaction restraints (the TCP-AIR score) into account. We evaluate our method using DockQ scores and demonstrate, that even with a clustering that require members to have a high similarity, i.e. with smaller clusters, we can assign high ranks to those clusters that contain poses close to the experimentally determined native structure of the ternary complexes. We also demonstrate the resultant improved yeild of near-native poses in these clusters.