Malaria-Instruct: A Low-Resource Benchmark for Instruction-Tuned LLMs in Antimalarial Drug Discovery

Overview:

This repository contains Malaria-Instruct, a specialized dataset designed to evaluate and fine-tune Large Language Models (LLMs) for molecular property prediction, specifically targeting antimalarial activity. While derived from ChEMBL, this dataset has been significantly restructured, filtered, and formatted for instruction-tuning and virtual screening tasks in resource-constrained research environments.

Key Features:

• Instruction-Response Format: Data is structured as natural language prompts (SMILES strings as input, binary activity as output) to align with the training objectives of modern LLMs (e.g., Gemma, LlaSMO, Mistral).

• Rigorous Dissimilarity Splitting: Unlike standard random or scaffold splitting, this dataset employs a "Hard Split" based on molecular dissimilarity (Tanimoto coefficient <0.4 for train-test and 0.5 for train-val). This ensures that models are tested on their ability to generalize to novel chemical spaces, simulating real-world lead optimization.

• Neglected Disease Focus: Specifically curated from assays involving various strains of Plasmodium falciparum, addressing the data gap in AI for the Global South. It contains assay on different strains on Plasmodium falciparum

• Benchmarking Metadata: Includes pre-calculated baseline results using Morgan Fingerprints (2048-bit) and classical models (Random Forest, XGBoost).

File Structure:

1. MalariaData_bioactivity_LLM_fewshot_dataset.zip: A zip file containing few shots (1-5 shorts) datasets in csv file with the splits.

2. MalariaData_bioactivity_selected_for_LLM_with_splits_fewshot.zip: A zip file containing the csv file of the dataset repeated splitted 5 times.

3. classical_results.csv: pre-computed classical ML baselines result summary.

4. README.md: Detailed documentation on filtering criteria and prompt templates.

Use Case: This dataset is intended for researchers developing chemistry-aware LLMs, MLOps engineers optimizing small-scale models for low-compute environments, and medicinal chemists working on malaria drug discovery.