AIC50: An Innovative Machine Learning Approach to Precise IC50 Prediction

The half-maximal inhibitory concentration (IC50) is a critical metric in pharmacology, used to evaluate a compound's potency by indicating the concentration needed to inhibit a biological target by fifty percent. Accurate IC50 measurement is essential for drug discovery, yet traditional experimental assays are costly and time-consuming, limiting their scalability for early-stage compound screening. Although high-throughput screening (HTS) technologies streamline IC50 testing, they remain expensive and resource-intensive. Consequently, computational methods have gained attention as efficient alternatives for predicting IC50 values and prioritizing compounds. In this study, we propose an advanced deep learning model for IC50 prediction that leverages bidirectional Long Short-Term Memory (LSTM) networks and multi-head attention mechanisms to address the complexity of protein-ligand interactions. Byte Pair Encoding (BPE) was applied to tokenize protein sequences and SMILES strings, providing efficient sub-word representations. Molecular properties such as molecular weight and hydrogen bond characteristics were directly incorporated as features, adding critical biochemical context. Using a comprehensive dataset from the ChEMBL database, containing 94,000 protein-ligand pairs annotated with IC50 values, our model achieved a maximum validation accuracy of 83%, outperforming baseline models that lack attention mechanisms. While promising, the model has limitations, including data variability from diverse ChEMBL assay sources and constraints on sequence length, which limits its applicability to large protein sequences and complex ligands. Broader IC50 ranges for higher classes also reduce prediction precision. Future improvements include refining IC50 class structures, training with multi-source datasets, and increasing model complexity to accommodate longer sequences and structural nuances better. This work advances computational drug discovery by presenting a robust IC50 prediction model that integrates sophisticated neural architectures, BPE tokenization, and molecular features, showing potential as an efficient tool for early drug screening. With future enhancements, this model could set a new standard for predictive accuracy, supporting accelerated and cost-effective drug discovery. Keywords: IC50 prediction, LSTM networks, Computational Drug discovery, Protein-Ligand interactions, Byte Pair Encoding Obi, E.D., Mbatuegwu, D., Yentumi, J.A. & Ajayi, O.O. (2024): AIC50: An Innovative Machine Learning Approach to Precise IC50 Prediction. Journal of Advances in Mathematical & Computational Science. Vol. 12, No. 4. Pp 31-44 Available online at www.isteams.net/mathematics-computationaljournal. dx.doi.org/10.22624/AIMS/MATHS/V12N4P3