DNN-PP: A novel Deep Neural Network approach and its applicability in drug-related property prediction

Calculations of Absolute Solvation Free Energies with Transformato─Application to the FreeSolv Database Using the CGenFF Force Field

“…During each MD simulation, trajectories are saved; these are post-processed to compute the energy differences to all other intermediate states. Finally, the multistate Bennett’s acceptance ratio method (MBAR) as implemented in is used to obtain free energy differences from these data . For the full details, we refer the reader to refs and .…”

“…37,38 Lately, the FreeSolv database is also used in the field of machine learning to develop and validate models for predicting molecular properties related to solvation and hydration. 39,40 Large-scale free energy simulations require automated setups. We recently developed and presented a tool, called transformato, 41,42 for calculating relative solvation and relative binding free energies using the common-core/serialatom-insertion approach 43 in a semi-automated manner.…”

“…Computational approaches are not limited to free energy methods based on molecular dynamics (MD). Quantum chemical calculations combined with implicit solvent models are known to predict solvation free energies well. , Recently, excellent agreement between computed and experimental ASFE values was obtained using molecular density functional theory. , Lately, the FreeSolv database is also used in the field of machine learning to develop and validate models for predicting molecular properties related to solvation and hydration. , …”

Extended study on atomic featurization in graph neural networks for molecular property prediction

“…Currently, graph representations of molecules are displacing molecular fingerprints, as graph neural networks can learn a molecular representation that is tailored to the prediction task. Graph convolutional neural networks (GCNs) [ 44 – 49 ] and, more recently, also graph transformers [ 17 , 18 , 50 , 51 ] demonstrate outstanding results across numerous molecular property prediction tasks. Chuang et al [ 52 ] discuss the crucial importance of molecular representations for tasks such as property prediction or generation of novel compounds.…”

“…However, global information about molecules, such as partition coefficient or polar surface area, which is available when using molecular descriptors, is still missing from graph representations. Attempts to make this information available to graph neural networks are already known in the literature [ 7 , 49 ].…”