Executive Summary

This portfolio is the result of my decision to actively bridge the gap between my academic background and the skills expected in industry. While my research experience gave me a strong foundation in computational chemistry and molecular simulations, I recognized that I needed hands-on practice with cheminformatics, machine learning, and physics-based drug discovery tools. So I am building this collection of projects as a way to learn by doing.

The work progresses from ligand-based QSAR with descriptors and GNNs, to structure-based docking, to free energy perturbation (FEP) for relative binding free energies (RBFE), to foundational generative models. Each project is organized in one or more jupyter notebooks, and reflects a concrete step in my upskilling journey — designed to demonstrate initiative, adaptability, and the ability to quickly pick up new methods that are directly relevant to modern computational chemistry and drug discovery in an industry setting.

Computational Chemistry + Machine Learning Portfolio

Overview

This portfolio documents my transition from academic computational chemistry to industry-relevant workflows that integrate machine learning and physics-based modeling. The projects highlight a progression from descriptor-based QSAR, to graph neural networks (GNNs), and finally to structure-based approaches such as docking and free energy perturbation (FEP).

Projects are grouped into two complementary tracks:

• Ligand-based modeling

  • FoundationalML — Baseline models with molecular descriptors/fingerprints (toxicity classification, solubility regression).
  • GNN — Graph-based deep learning using PyTorch Geometric (solubility prediction and toxicity classification).

• Structure-based modeling

  • SBDD_docking — Virtual screening workflow with AutoDock Vina (protein prep, ligand docking, scoring, pose selection).
  • FreeEnergyPerturbation — Setup of relative binding free energy calculations with OpenFE/OpenMM.

• Generative modeling

  • RNN-based SMILES generation — Baseline model to generate new, valid SMILES strings after learning the grammar using a Recurrent Neural Network
  • VAE-based SMILES generation – Introduce Variational AutoEncoders
  • Diffusion models (in progress)

Through these projects I gained hands-on experience in:

• Molecular featurization and data preprocessing with RDKit
• Classical ML and deep learning with scikit-learn and PyTorch/PyG
• Graph representation learning and message passing for molecules
• Interpreting results in a chemical/biological context
• Setting up and running FEP workflows (mapping, alchemical transformations, thermodynamic cycles)
• Performing structure-based docking with flexible ligand placement in rigid protein pockets
• Undestanding the math/physics behind generative AI

This portfolio demonstrates how I combine data-driven ML and physics-based simulations into end-to-end workflows for molecular property prediction and drug discovery.

---

Repository Structure

• FoundationalML/ — Classical descriptor-based QSAR (toxicity, solubility)
• GNN/ — Graph neural network models and utilities
• SBDD_docking/ — Structure-based docking workflows
• FreeEnergyPerturbation/ — Free energy perturbation setup and examples
• GenerativeModels/ - RNN, VAE-based SMILES generators, Diffusion models
• README.md files in each folder describe project details.

---

How to Use This Repo

Each notebook is self-contained and runnable independently.

---

Notes

• The repository is actively maintained and will continue to expand (e.g., generative models)
• Any feedback/inquiry is greatly appreciated! Please write to lorenzobonimba@gmail.com. Thank you for your help with improving these materials!