View our Online Documentation or try the AIMSim comprehensive tutorial in your browser.
Why Do We Need To Visualize Molecular Similarity / Diversity?
There are several contexts where it is helpful to visualize the diversity of a molecular dataset:
Exploratory Experimental Synthesis
For a chemist, synthesizing new molecules with targeted properties is often a laborious and time consuming task. In such a case, it becomes useful to check the similarity of a newly proposed (un-synthesized) molecule to the ones already synthesized. If the proposed molecule is too similar to the existing repertoire of molecules, it will probably not yield not enough new information / property and thus need not be synthesized. Thus, a chemist can avoid spending time and effort synthesizing molecules not useful for the project.
Lead Optimization and Virtual Screening
This application is the converse of exploratory synthesis where the interest is to find molecules in a database which are structurally similar to an "active" molecule. In this context, "active" might refer to pharmocological activity (drug discover campaigns) or desirable chemical properties (for example, to discover alternative chemicals and solvents for an application). In such a case, AIMSim helps to run virtual screenings over a molecular database and visualize the results.
Machine Learning Molecular Properties
In the context of machine learning, visualizing the diversity of the training set gives a good idea about its information quality. A more diverse training data-set yields a more robust model, which generalizes well to unseen data. Additionally, such a visualization can identify "clusters of similarity" indicating the need for separately trained models for each cluster.
Substrate Scope Robustness Verification
When proposing a novel reaction it is essential for the practicing chemist to evaluate the transformation's tolerance of diverse functional groups and substrates (Glorius, 2013). Using AIMSim, one can evaluate the structural and chemical similarity across an entire susbtrate scope to ensure that it avoids redundant species. Below is an example similarity heatmap generated to visualize the diversity of a three-component sulfonamide coupling reaction with a substantial number of substrates (Chen, 2018).
Many of the substrates appear similar to one another and thereby redundant, but in reality the core sulfone moiety and the use of the same coupling partner when evaluating functional group tolerance accounts for this apparent shortcoming. Also of note is the region of high similarity along the diagonal where the substrates often differ by a single halide heteratom or substitution pattern.
AIMSim can be installed with a single command:
pip install aimsim
This command also installs the required dependencies. It is recommended to install AIMSim in a separate virtual environment.
Optional: Previous versions of AIMSim provided direct support for the descriptors provided in the mordred package. Unforunately, mordred is no longer recieving updates and causes significant depdendency conflicts. Because of this, it is an optional add-on to AIMSim that is only compatible with Python 3.8. To install with mordred support, use pip install 'aimsim[mordred]' (note the single quotes, necessary in zsh).
Unit tests from previous versions of AIMSim have been kept but are not actively maintained due to the limitations above. Use mordred at your own risk.
AIMSim is compatible with Python 3.7 to 3.9.
Start AIMSim with a graphical user interface:
aimsim
Start AIMSim with a prepared configuration YAML file (config.yaml):
aimsim config.yaml
- Morgan Fingerprint (Equivalent to the ECFP fingerprints)
- RDKit Topological Fingerprint
- RDKit Daylight Fingerprint
The following are available via command line use (config.yaml) only:
- MinHash Fingerprint (see MHFP)
- All fingerprints available from the ccbmlib package (specify 'ccbmlib:descriptorname' for command line input).
- All descriptors and fingerprints available from PaDELPy, an interface to PaDEL-Descriptor. (specify 'padelpy:desciptorname' for command line input.).
- All descriptors available through the Mordred library (specify 'mordred:desciptorname' for command line input.). To enable this option, you must install with
pip install 'aimsim[mordred]'(see disclaimer in the Installation section above).
44 commonly used similarity scores are implemented in AIMSim. Additional L0, L1 and L2 norm based similarities are also implemented. View our Online Documentation for a complete list of implemented similarity scores.
-
Measure Search: Automate the search of fingerprint and similarity metric (called a "measure") using the following algorithm: Step 1: Select an arbitrary featurization scheme. Step 2: Featurize the molecule set using the selected scheme. Step 3: Choose an arbitrary similarity measure. Step 4: Select each molecule’s nearest and furthest neighbors in the set using the similarity measure. Step 5: Measure the correlation between a molecule’s QoI and its nearest neighbor’s QoI. Step 6: Measure the correlation between a molecule’s QoI and its further neighbor’s QoI. Step 7: Define a score which maximizes the value in Step 5 and minimizes the value in Step 6. Step 8: Iterate Steps 1 – 7 to select the featurization scheme and similarity measure to maximize the result of Step 7.
-
See Property Variation with Similarity: Visualize the correlation in the QoI between nearest neighbor molecules (most similar pairs in the molecule set) and between the furthest neighbor molecules (most dissimilar pairs in the molecule set). This is used to verify that the chosen measure is appropriate for the task.
-
Visualize Dataset: Visualize the diversity of the molecule set in the form of a pairwise similarity density and a similarity heatmap of the molecule set. Embed the molecule set in 2D space using using principal component analysis (PCA)[3], multi-dimensional scaling[4], t-SNE[5], Spectral Embedding[6], or Isomap[7].
-
Compare Target Molecule to Molecule Set: Run a similarity search of a molecule against a database of molecules (molecule set). This task can be used to identify the most similar (useful in virtual screening operations) or most dissimilar (useful in application that require high diversity such as training set design for machine learning models) molecules.
-
Cluster Data: Cluster the molecule set. The following algorithms are implemented:
For arbitrary molecular features or similarity metrics with defined Euclidean distances: K-Medoids[3] and Ward[8] (hierarchical clustering).
For binary fingerprints: Complete, single and average linkage hierarchical clustering[8].
The clustered data is plotted in two dimensions using principal component analysis (PCA)[3], multi-dimensional scaling[4], or TSNE[5].
- Outlier Detection: Using an isolation forest, check for which molecules are potentially novel or are outliers according to the selected descriptor. Output can be directly to the command line by specifiying
outputto beterminalor to a text file by instead providing a filename.
Developer: Himaghna Bhattacharjee, Vlachos Research Lab. (LinkedIn)
Developer: Jackson Burns, Don Watson Lab. (Personal Site)
Issues and Pull Requests are welcomed! To propose an addition to AIMSim open an issue and the developers will tag it as an enhancement and start discussion.
AIMSim includes an automated testing apparatus operated by Python's unittest built-in package. To execute tests related to the core functionality of AIMSim, run this command:
python -m unittest discover
Full multiprocessing speedup and efficiency tests take more than 10 hours to run due to the number of replicates required. To run these tests, create a file called .speedup-test in the AIMSim directory and execute the above command as shown.
To manually build the docs, execute the following with sphinx and m2r installed and from the /docs directory:
m2r ../README.md | mv ../README.rst . | sphinx-apidoc -f -o . .. | make html | cp _build/html/* .
Documentation should manually build on push to master branch via an automated GitHub action.
For packaging on PyPI:
python -m build; twine upload dist/*
Be sure to bump the version in __init__.py.
If you use this code for scientific publications, please cite the following paper.
Bhattacharjee, H.҂; Burns, J.҂; Vlachos, D.G. (2021): AIMSim: An Accessible Cheminformatics Platform for Similarity Operations on Chemicals Datasets. ChemRxiv. Preprint. https://doi.org/10.26434/chemrxiv-2022-nw6f5
This code is made available under the terms of the MIT Open License:
Copyright (c) 2020 Himaghna Bhattacharjee & Jackson Burns
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
[1] Collins, K. and Glorius, F., A robustness screen for the rapid assessment of chemical reactions. Nature Chem 5, 597–601 (2013). https://doi.org/10.1038/nchem.1669
[2] Chen, Y., Murray, P.R.D., Davies, A.T., and Willis M.C., J. Am. Chem. Soc. 140 (28), 8781-8787 (2018). https://doi.org/10.1021/jacs.8b04532
[3] Hastie, T., Tibshirani R. and Friedman J., The Elements of statistical Learning: Data Mining, Inference, and Prediction, 2nd Ed., Springer Series in Statistics (2009).
[4] Borg, I. and Groenen, P.J.F., Modern Multidimensional Scaling: Theory and Applications, Springer Series in Statistics (2005).
[5] van der Maaten, L.J.P. and Hinton, G.E., Visualizing High-Dimensional Data Using t-SNE. Journal of Machine Learning Research 9:2579-2605 (2008).
[6] Ng, A.Y., Jordan, M.I. and Weiss, Y., On Spectral Clustering: Analysis and an algorithm. ADVANCES IN NEURAL INFORMATION PROCESSING SYSTEMS, MIT Press (2001).
[7] Tenenbaum, J.B., De Silva, V. and Langford, J.C, A global geometric framework for nonlinear dimensionality reduction, Science 290 (5500), 2319-23 (2000). https://doi.org/10.1126/science.290.5500.2319.
[8] Murtagh, F. and Contreras, P., Algorithms for hierarchical clustering: an overview. WIREs Data Mining Knowl Discov (2011). https://doi.org/10.1002/widm.53