Advancing Model Interpretability for Drug Discovery using VSA Explainer

1. Problem Statement & Hypothesis

The quantitative description of molecular properties through numerical descriptors represents a cornerstone of AI-driven drug discovery research. Molecular descriptors are widely used in cheminformatics workflows, including quantitative structure analysis relationships (QSAR), machine learning, and chemical AI.¹Among the commonly used numerical descriptors, Van der Waals Surface Area (VSA) represents three-dimensional molecular geometry with one-dimensional numerical values that partition molecular surface area according to atomic physicochemical properties, thereby encoding structural and electronic characteristics within one descriptor. VSA descriptors are widely accessed through descriptor calculators including RDkit² and Mordred Community.³These VSA descriptors are extremely information-rich, frequently appearing as useful in a variety of molecular and material property prediction tasks. For example, feature importance analysis has revealed VSA descriptors to be strongly correlated with properties ranging from pharmacokinetic parameters,^4,5 protein binding,^6-8 blood brain barrier (BBB) permeability,⁹ and lipid nanoparticle transfection efficiency,¹⁰ among a number of other applications.

However, despite their universality and information-density, VSA descriptors are challenging to interpret and attribute.¹¹This is in stark contrast to substantial recent efforts to interpret models based on other representations. Similarity maps are a widely utilized tool to interpret molecular fingerprint representations.¹² For sequence-based models trained directly on SMILES strings, XSMILES¹³ and a module developed by Zhao et al. can both attribute character-level contributions to properties.¹⁴Transformer-CNN¹⁵ and other graph interpretation tools have also been developed for molecular graph representations.¹⁶Additionally, existing tools can decompose atom-level Crippen contributions.¹⁷ These tools have become useful instruments in iterative molecular design workflows, helping chemists visualize how atom-level substitutions impact target properties. Yet, for VSA descriptors, feature importance ranking is typically the last step performed when interpreting results from an ML model trained on VSA descriptors. While feature ranking can offer useful insight, it is generally un-actionable for molecular design, since the connection between differences in VSA descriptor values to differences between molecular structures in non-intuitive. Until now, no tools exist for visualizing VSA descriptors.

Recently, our team has started to develop VSA Explainer, an open-source tool that enables the interpretation of individual descriptors by highlighting their atomic contributions within a molecule. This tools enables the precise attribution of VSA descriptor differences to differences in molecular structures.

Hypothesis
Developing open-sourced, well-documented case studies using VSA Explainer will demonstrate how atom-level interpretability of VSA descriptors can directly guide molecular design decisions for blood brain barrier (BBB) permeability and toxicity, critical properties in neurodegenerative drug design.

2. Technical Approach

We demonstrate the use of VSA Explainer in cheminformatics workflows for molecular property prediction. We will train models predicting BBB permeability using data from the B3DB dataset,^18,19 and models predicting toxicity from ToxValDB.²⁰ For each case study, we will focus on example molecular series relevant to neurodegenerative disease drawn from the literature. We will then use VSA Explainer to rationalize our model predictions, deepening our ability to attribute atomic differences. These case studies will show how the modular VSA Explainer can be integrated into existing cheminformatics toolkits (RDKit and Mordred) and model interpretability tools (SHAPley values). We will highlight how VSA Explainer visualizations explain different effects of scaffold decorations during lead optimization, bridging the gap between ML predictions and actionable insight in molecular design.

3. Expected Outputs

Overall, this project will help develop VSA Explainer as a reusable tool that drug hunters can readily use for drug design and ML model interpretability. We will publish our findings as easy-to-follow reproducible Jupyter notebooks in the GitHub repository for the community to use. From each case study, we also expect that new insights from VSA Explainer will deepen our understanding of BBB and toxicity optimization, since the tool provides new capabilities for model interpretation that the field has not yet had. These results will be compiled and published in a computational drug discovery journal. Finally, we will continue to develop the open source VSA Explainer toolkit, including a command-line interface (CLI) and a grid-view for comparing multiple analogs around a shared scaffold. We expect this project to take between 2-3 months.

4. Relevance to AI-Driven Drug Discovery

This project directly advances AI-driven discovery through developing a tool to model interpretability that addresses a crucial gap. Additionally, the molecular properties we have targeted, BBB penetration and toxicity, are central design challenges in developing therapies for Alzheimer’s, Parkinson’s, ALS, and other neurodegenerative diseases. This project directly supports three of the areas of interest for this CFP:

AI for Molecular Design (interpretable ADMET optimization)
Open Tools, Data, & Infrastructure (open-source interpretability tool)
DeSci workflows (transparent, reusable, low-cost research artifacts)

This work will multiply the usability of existing AI models for drug discovery, helping making AI outputs actionable for chemists.

5. Budget Allocation

Research time (case study development & analysis): $3,500
Compute & software infrastructure: $500
Dissemination: $1,500

6. Team

This project will be conducted by Lucas Attia, a graduating Chemical Engineering PhD candidate from MIT. Lucas’ research expertise spans molecular property prediction using ML, molecular dynamics simulations, and nanoparticle drug delivery. His research has resulted in several high impact publications in venues including Nature Communications, Advanced Materials, Nature Computational Science, and NeurIPS 2025. Lucas’ deep expertise in cheminformatics positions him to successfully conduct and lead this project.

References

Khan, A. U.; others Descriptors and their selection methods in QSAR analysis: paradigm for drug design. Drug discovery today 2016, 21, 1291–1302.
Landrum, G.; others RDKit: Open-source cheminformatics. https://www.rdkit.org, 2023; Version 2023.09.5.
Moriwaki, H.; Tian, Y.-S.; Kawashita, N.; Takagi, T. Mordred: a molecular descriptor calculator. Journal of cheminformatics 2018, 10, 4.
Wang, Y.; Liu, H.; Fan, Y.; Chen, X.; Yang, Y.; Zhu, L.; Zhao, J.; Chen, Y.; Zhang, Y. In silico prediction of human intravenous pharmacokinetic parameters with improved accuracy. Journal of chemical information and modeling 2019, 59, 3968–3980. 15
Chen, Y.; Cameron, K.; Guzman-Perez, A.; Perry, D.; Li, D.; Gao, H. Structure– pharmacokinetic relationship of in vivo rat biliary excretion. Biopharmaceutics & Drug Disposition 2010, 31, 82–90.
Yoshida, K.; Niwa, T. Quantitative structure- activity relationship studies on inhibition of hERG potassium channels. Journal of chemical information and modeling 2006, 46, 1371–1378.
Cui, Z.; Zhang, N.; Zhou, T.; Zhou, X.; Meng, H.; Yu, Y.; Zhang, Z.; Zhang, Y.; Wang, W.; Liu, Y. Conserved sites and recognition mechanisms of T1R1 and T2R14 receptors revealed by ensemble docking and molecular descriptors and fingerprints combined with machine learning. Journal of Agricultural and Food Chemistry 2023, 71, 5630–5645.
Smits, R. A.; Adami, M.; Istyastono, E. P.; Zuiderveld, O. P.; van Dam, C. M.; de Kanter, F. J.; Jongejan, A.; Coruzzi, G.; Leurs, R.; de Esch, I. J. Synthesis and QSAR of quinazoline sulfonamides as highly potent human histamine H4 receptor inverse agonists. Journal of Medicinal Chemistry 2010, 53, 2390–2400.
Zhang, L.; Zhu, H.; Oprea, T. I.; Golbraikh, A.; Tropsha, A. QSAR modeling of the blood–brain barrier permeability for diverse organic compounds. Pharmaceutical research 2008, 25, 1902–1914.
Xu, Y.; Ma, S.; Cui, H.; Chen, J.; Xu, S.; Gong, F.; Golubovic, A.; Zhou, M.; Wang, K. C.; Varley, A.; others AGILE platform: a deep learning powered approach 16 to accelerate LNP development for mRNA delivery. Nature communications 2024, 15, 6305.
Cherkasov, A.; Muratov, E. N.; Fourches, D.; Varnek, A.; Baskin, I. I.; Cronin, M.; Dearden, J.; Gramatica, P.; Martin, Y. C.; Todeschini, R.; others QSAR modeling: where have you been? Where are you going to? Journal of medicinal chemistry 2014, 57, 4977–5010.
Riniker, S.; Landrum, G. A. Similarity maps-a visualization strategy for molecular fingerprints and machine-learning methods. Journal of cheminformatics 2013, 5, 43.
Heberle, H.; Zhao, L.; Schmidt, S.; Wolf, T.; Heinrich, J. XSMILES: interactive visualization for molecules, SMILES and XAI attribution scores. Journal of Cheminformatics 2023, 15, 2.
Zhao, L.; Montanari, F.; Heberle, H.; Schmidt, S. Modeling bioconcentration factors in fish with explainable deep learning. Artificial Intelligence in the Life Sciences 2022, 2, 100047.
Karpov, P.; Godin, G.; Tetko, I. V. Transformer-CNN: Swiss knife for QSAR modeling and interpretation. Journal of cheminformatics 2020, 12, 17.
Hochuli, J.; Helbling, A.; Skaist, T.; Ragoza, M.; Koes, D. R. Visualizing convolutional neural network protein-ligand scoring. Journal of Molecular Graphics and Modelling 2018, 84, 96–108.
Rasmussen, M. H.; Christensen, D. S.; Jensen, J. H. Do machines dream of atoms? Crippen’s logP as a quantitative molecular benchmark for explainable AI heatmaps. SciPost Chemistry 2023, 2, 002.
Meng, F.; Xi, Y.; Huang, J.; Ayers, P. W. A curated diverse molecular database of
blood-brain barrier permeability with chemical descriptors. Scientific Data 2021, 8 .
Meng, F.; Chen, J.; Collins-Ramirez, J. S.; Ayers, P. W. B3clf: A Resampling-
Integrated Machine Learning Framework to Predict Blood-Brain Barrier Permeabil-
ity. ChemRxiv 2025.
Wall, J. T.; Sayre, R. R.; Smith, D., Winter, S.; Groover, M.; Hope, J.; ... & Weitekamp, C. A. Development of the toxicity values database, ToxValDB: A curated resource for experimental and derived human health-relevant toxicity data. Computational Toxicology 2025 35, 100365.

Advancing Model Interpretability for Drug Discovery using VSA Explainer

Advancing Model Interpretability for Drug Discovery using VSA Explainer

Funding Progress

Peer Reviews

Related Funding

CerebrumDAO Microgift CFP: AI for Drug Discovery ($5,000)

Nonprofit Recipient

Topics

DOI