EffiChem : Efficient Adaptation of Chemical Language Models for Molecular Property Prediction

Chemical language models (CLMs) such as ChemBERTa and Molformer enable compound property prediction, albeit using an over-parameterized model, and the computational demands of full finetuning such models limit their adoption in resource-constrained settings. Parameter-efficient finetuning techniques are critical to democratize AI-driven drug discovery and design. In this paper, we integrate Low-Rank Adapters (LoRA) with chemical language models, significantly reducing the number of trainable parameters required for finetuning. We observe improved performance of 3-5% for AUC metric across classification tasks, for molecule toxicity, blood- brain barrier permeability, and flavor prediction tasks over state-of-the-art like Molformer-XL and MAMMAL. We attain optimal performance by comprehensive benchmarking of embeddings generated from zero-shot and finetuned CLMs combined with molecular physicochemical properties across three diverse datasets. Our approach achieves MCC scores ranging from 0.80 to 0.90 across three tasks, while reducing model parameters by 62% to 96% in the top-performing models. This is achieved via lightweight adaptation of CLMs, which retains their performance efficiency while reducing the over-parameterization burden. This work provides a practical framework for deploying a scalable and sustainable paradigm for high-performance molecular AI tools in resource-constrained settings.