Use ESPsim Online

ESP-Sim: 3D Shape and Electrostatic Potential Similarity for Scaffold Hopping and Generative Drug Design

ESP-Sim is an open-source Python package for comparing the electrostatic potential and 3D molecular shape of molecules and fragments. It was introduced by Giovanni Bolcato, Esther Heid and Jonas Bostrom (University of Padova, MIT and AstraZeneca) in the Journal of Chemical Information and Modeling (2022), alongside DeepFMPO v3D, an extension of their fragment-based reinforcement learning generator that uses ESP-Sim to reward fragment swaps with similar 3D shape and electrostatics. The paper states that the code is open-source and freely available.

Shape and electrostatics are primary determinants of molecular recognition, yet the authors note they have been little explored in de novo generative methods. One reason is the notoriously hard problem of obtaining bioactive conformations. In addition, small edits to a SMILES string can change 3D structure a lot, so most generative methods that work on 2D strings cannot optimize smoothly in the space of 3D properties. The authors also point out that virtual screening is limited to whatever is in the queried database.

How ESP-Sim Computes Electrostatic and Shape Similarity

ESP-Sim uses RDKit to generate conformations, optionally with a constrained anchor or core group, and then computes shape and electrostatic similarity for pairs of molecules. Pre-aligned molecules can also be compared directly.

1. Generate conformers: an ensemble of low-energy conformers is built with ETKDG in RDKit.

2. Assign partial charges: choose Gasteiger, MMFF94, a bundled machine learning model, or RESP charges from the Psi4 quantum chemistry program (for example B3LYP/6-31G**).

3. Compute the potential: the Coulomb potential is either fitted with three Gaussian functions and integrated analytically, or integrated by Monte Carlo in the space outside the van der Waals radii and inside a user-defined margin.

4. Score overlap: the overlap integral and self-overlap integrals give a Tanimoto or Carbo similarity, and shape Tanimoto describes volume overlap.

Fragment Alignment Inside DeepFMPO v3D

DeepFMPO breaks single bonds extending from ring atoms to create fragments, encodes similar fragments with similar binary strings using a balanced tree, and learns replacements with an actor-critic reinforcement learning model. To align fragments, an anchor group (a hexazine ring on a methylene linker by default) is attached at the attachment atom and fixed in space. The conformer pair with the best shape overlay is kept, the anchor is replaced by hydrogen, and the ESP similarity is then computed. This avoids having to generate bioactive conformations for whole molecules, and the authors report that results were essentially independent of the anchor chosen (hexazine, carboxylic acid, piperidine or iodine).

Benchmark Results for ESP-Sim and DeepFMPO v3D

Partial Charge Quality

On about 3000 neutral molecules, the authors compared each charge model with QM RESP charges (MP2 with a polarizable basis set). AM1-BCC reproduced the quantum mechanical potentials best, followed by the machine learning model, MMFF and Gasteiger.

The authors conclude that even Gasteiger charges give a fair depiction of the potential for most molecules. ESP-Sim scores correlated strongly with the EON tool on about 450 fragments (Spearman about 0.8), and the paper reports good performance on the dopamine D4 receptor and the 102 DUD-E targets in its Supporting Information.

Scaffold Hopping with CDK2 Inhibitors

In a retrospective study, ESP-Sim (B3LYP/6-31G**) ranked a known equipotent pair of CDK2 bicyclic scaffolds (imidazopyridine and pyrazolopyridine) first among 30 scaffolds, while 2D methods ranked it lower.

DeepFMPO v3D then generated 6359 unique molecules after 1000 epochs, about two-thirds of which met the molecular weight, clogP and polar surface area targets. Many contained the pyrazolopyridine scaffold, while runs using Morgan, MACCS or topological fingerprints produced none.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with ESP-Sim on Tamarind Bio

• Scaffold hopping: find bioisosteric replacements for a lead's central ring system, as in the CDK2 case study.

• Ligand-based virtual screening: rank candidates by 3D shape and electrostatic similarity to known actives, as benchmarked on DUD-E and D4 data.

• Lead optimization triage: score generated or purchased analogs by how closely they match a lead's electrostatic surface.

• Biasing generative design: use similarity scores as a reward so that generators stay close to known active chemistry.

How to Use ESP-Sim on Tamarind Bio

1. Open the tool: log in to tamarind.bio and select ESP-Sim.

2. Provide the molecules: enter a reference molecule and the molecules or fragments to compare (for example as SMILES or structure files), as the form defines.

3. Pick the partial charge method: choose among options such as Gasteiger, MMFF94, ML or RESP, if exposed. The paper recommends RESP or AM1-BCC charges when accuracy matters.

4. Set alignment options: decide whether molecules are pre-aligned or conformers are generated, and set the number of conformers, if exposed.

5. Choose the score: select ESP Tanimoto or Carbo similarity, shape similarity, or the combined score, if exposed.

6. Run the job: submit and wait for completion.

7. Download and interpret: rank molecules by similarity and review the top candidates.

8. Use downstream: take the best matches into docking, further property filtering or synthesis planning.

Things to Keep in Mind

• RESP charges via Psi4 are computationally demanding, and the RESP/Psi4 setup is not parametrized beyond argon, so larger atoms such as bromine need their van der Waals radius set separately.

• Cheaper charge models gave similar trends in the benchmarks, but the authors recommend a higher level of theory when affordable.

• The case study is retrospective and covers one target family, and the authors stress that generative methods should augment, not replace, medicinal chemistry judgment.

• Similarity-based design intentionally stays near known leads, so it is not meant to explore the whole of chemical space.

Source: Bolcato G, Heid E, Bostrom J. On the Value of Using 3D Shape and Electrostatic Similarities in Deep Generative Methods. J. Chem. Inf. Model. 2022, 62, 1388-1398. DOI: 10.1021/acs.jcim.1c01535. Code: https://github.com/hesther/espsim.