Revolutionizing Drug Discovery: The Role of AI in Lead Optimization

Lead optimization in the field of drug discovery poses a formidable challenge, predominantly hinging on conjectures and the expertise of medicinal chemists. Consequently, it frequently yields indeterminate results and operates with a marked degree of inefficiency.

Moreover, this endeavor is a protracted undertaking that necessitates substantial resources. Hence, the integration of artificial intelligence (AI) predictive tools to expedite this process holds significant promise in the realm of drug discovery.

In silico techniques like free energy perturbation (FEP) and molecular mechanics generalized born surface area (MM-GB/SA) have convincingly demonstrated their utility in enhancing lead optimization by calculating binding free energy.

However, their intricate preparation procedures, limited capacity for processing molecules, and restrictions on accommodating variations between molecules impose obstacles to their regular application. There is an immediate and pressing demand for the creation of a proficient and precise in silico predictive tool to steer lead optimization.

In a research study recently published in Nature Computational Science, a group of scientists, under the leadership of Professor Mingyue Zheng from the Shanghai Institute of Materia Medica (SIMM) at the Chinese Academy of Sciences, introduced a pairwise binding comparison network (PBCNet).

This network forecasts the relative binding affinities among closely related ligands by employing a physics-informed graph attention mechanism, utilizing a pair of protein pocket-ligand complexes as its input. PBCNet demonstrates its practical utility in facilitating structure-based drug lead optimization due to its speed, accuracy, and user-friendly interface.

To assess PBCNet's ranking capability and accuracy, the research team led by Zheng conducted validation using two withheld datasets provided by Schrodinger, Inc. and Merck KGaA. These datasets encompassed more than 460 ligands and 16 targets.

In their work, they employed transfer learning, a technique that entails training models initially on extensive datasets and then fine-tuning them for tasks with limited data. This approach notably enhanced the models' performance in these specific tasks.

The benchmarking results from the test data demonstrated that the pretrained PBCNet surpassed Schrodinger's Glide, MM-GB/SA, and four recently developed deep learning models (DeltaDelta, Default2018, Dense, and PIGNet).

Moreover, even when supplied with a limited amount of fine-tuning data (ranging from 2 to 10 ligands with known binding activity), PBCNet achieved a level of performance comparable to Schrodinger's FEP+. The latter is widely recognized as the industry standard for computational lead optimization in the pharmaceutical sector.

The researchers also assessed PBCNet's ability to efficiently pinpoint high-activity compounds in a practical lead optimization scenario. They employed a benchmark comprising nine recently published chemical series and compared the model's selection order with the experimental synthesis order.

The results of the evaluation revealed that the incorporation of PBCNet led to a remarkable acceleration of approximately 473% in the tested lead optimization projects, while simultaneously reducing resource investment by an average of 30%.

This study underscores the immediate practical value of PBCNet in guiding lead optimization endeavors. Additionally, there is a freely accessible academic version of PBCNet that aids in predicting ligand binding affinities.

The significance of AI in addressing scientific challenges has grown, thanks to its capacity to embed specialized domain knowledge into its models. PBCNet serves as an illustrative example of this methodology, as it seamlessly incorporates both physical and pre-existing knowledge into its modeling framework.