Universiteit Leiden

nl en

Lecture | Center of Computational Life Sciences

CCLS Seminar

Tuesday 10 May 2022
After the seminar you're invited for drinks at the FooBar.
Niels Bohrweg 1
2333 CA Leiden

Data-driven Generation of Perturbation Networks for Relative Binding Free Energy Calculations

Early-stage drug discovery campaigns pose major challenges to the pharmaceutical industry, mostly owing to large costs and high rate of failure of high-throughput screening of candidate molecules to the therapeutic target in question. Recent advances in computational chemistry and molecular simulation offer promising techniques that allow researchers to conduct these screenings virtually; especially Alchemical Free Energy (AFE) calculations have increasingly gained traction in solving the ligand-optimisation problem in both academic and corporate drug discovery [1]. Free Energy Perturbation (FEP) - where candidate molecules are transformed into each other virtually to compute the relative free energy of binding - has been one of the most tractable AFE techniques.

A critical step in FEP workflows is the generation of effective perturbation networks to ensure transformations are being simulated with sufficient phase-space overlap maximise prediction reliability. Currently, state-of-the-art softwares use primarily molecular similarity metrics to estimate optimal edges in networks. This method often still requires the user to review and tweak the presented perturbation network, hindering development toward a fully-automated FEP workflow. The current study uses a machine-learning (ML) approach to train models on many solvation FEP simulations to predict the reliability of a given perturbation a priori. Such models can replace similarity-based metrics to plan perturbation networks (figure 1).

This open-source project has resulted in several advancements that further progress the field of FEP. Using the data-driven method, practicioners can generate more reliable FEP networks, are able to transfer-learn to alternative FEP software reliabilities and fine-tune the model to the molecule series that is being investigated. The training domain used in this work has been made publicly available to allow researchers to create alternative novel models in the context of FEP.

This website uses cookies.  More information.