Deep Learning Molecular Properties
Binding Affinity Prediction With Machine Learning
The KDeep app provides ligand-based and structure-based machine learning predictors for multiple molecular properties. One example is binding affinity of a set of ligands against any given target with proven accuracy. Built using the TorchMD-net framework, they model the ligand-target complex as a 3D graph and apply state-of-the-art equivariant transformer layers to retrieve the binding affinities. Thanks to their computational efficiency they are able to screen multi-million datasets in just hours.

All of our predictors have been published in peer-reviewed journals and validated on data from pharmaceutical companies and in blind challenges. KDeep won two blind subchallenges of the D3R Grand Challenge 4 and it is actively used in the pharmaceutical industry. The ligand-based predictor reached the second highest rank with no significant statistical difference with the top-ranked predictor in the Sampl challenge.
The various predictors provide different prediction modalities and can be applied to any use case:
- KDeep: Absolute binding affinity predictor for regression of congeneric series
- KDeep-Bindscope: Binding affinity classifier for screening.
- KDeep-Trainer: Retraining on internal datasets drastically improves the reliability of prediction reaching correlations similar to FEP-based methods after retraining on approximately 50-100 molecules.

Use Cases
- Virtual Screening: Rank your library of docked compounds for binding affinity against a given target with KDeep or use together with AceDock. Use Bindscope to additionally detect non-binding ligands.
- Lead Optimization: Rank congeneric series of compounds with a common core with DeltaDelta.
- Molecule Generation: Use KDeep as a scoring function in AceGen to design compounds that meet a certain binding affinity threshold.
- Model Training and Finetuning: Train your own predictor or finetune one of Acellera´s predictors using your in-house data with KDeepTrainer. Our models are designed to use and predict different affinity metrics (e.g. pKd, pIC50,…) simultaneously.
Results
- A list of ranked compounds with their predicted affinity and probability of the compound being a binder.
- The newly trained or finetuned model file that can further be finetuned or used to generate predictions.