All research

Molecular Property Prediction via Fine-Tuned ML Models

Topics

Machine Learning
Transformers
Chemistry
Python

Molecular property prediction is a core challenge in computational chemistry — given a molecule's structure, predict a physical or biological property without running expensive experiments. My work focused on adapting transformer-based architectures to this domain using chemistry-specific pre-training.

Approach

I fine-tuned pre-trained language models on molecular SMILES representations, incorporating domain-specific tokenisation and pre-training objectives aligned with chemical structure. The goal was to improve downstream property prediction accuracy on benchmark datasets without requiring large amounts of labelled data.

Key techniques explored: - Chemistry-specific tokenisation for SMILES strings - Domain-adaptive pre-training on unlabelled chemical corpora - Transfer learning from general transformers to molecular property tasks - Evaluation on molecular benchmark datasets

Outcome

The fine-tuned models showed improved accuracy on downstream property tasks compared to general-purpose baselines. The project attracted an industrial lab role offer from IFFCO Dubai, validating the practical relevance of the approach.

Tools & Libraries

Python, PyTorch, HuggingFace Transformers, RDKit, Jupyter.