GigaScreen

GigaScreen is a deep learning-powered virtual screening solution developed by Molsoft L.L.C., designed to tackle the computational challenges of screening ultra-large chemical databases containing billions of compounds. By combining machine learning and deep learning tools, GigaScreen enables drug discovery teams and computational chemists to efficiently identify top hits from giga-sized libraries without requiring expensive infrastructure or additional software.

The GigaScreen protocol integrates three core components: Molsoft's RIDGE structure-based docking engine, the neural network scoring function RTCNN, and a Chemical Graph Convolutional Neural Network (GCNN) model. Together, these tools form an iterative screening pipeline that progressively refines compound selection from massive databases in a practical and resource-efficient manner.

Core Methodology and Workflow

• The process begins by selecting small random subsets or batches from the full database — typically 0.1% of the total library size, equating to approximately one to two million entries per batch.
• Two initial random sets are selected: one for testing and one for the first training iteration.
• During each iteration, the selected subset is trained using the GCNN model, which utilises fingerprint ECFP and Chemical Graph Convolutional approaches.
• The trained model is then applied to the entire database to identify top hits using a score cutoff, recommended to capture 80% of the top 0.1% hits by RTCNN and Squad scores.
• The cutoff value can be adjusted between 75% and 90% depending on specific project requirements.
• If the resulting number of compounds remains too large after applying the cutoff, an additional random batch of approximately one million compounds is selected and the iteration is repeated, with the expectation that model performance improves with each cycle.

Model Performance Monitoring

• Two key parameters are tracked across iterations to monitor model improvement.
• The first is the 25th percentile RTCNN score from each iteration after docking, with the goal of seeing this score improve progressively throughout training.
• The second is the total number of compounds remaining after applying the score cutoff, which should decrease with each iteration to a manageable level.
• Reduction can be further managed by randomly selecting a percentage of the remaining compounds for subsequent rounds.

Hardware Requirements and Performance Benchmarks

• GigaScreen is designed to run entirely on a single machine, eliminating the need for costly distributed computing infrastructure or additional third-party software.
• On a system equipped with an RTX 4090 GPU and an AMD processor, each training iteration against a database of two billion compounds takes approximately 15 hours.
• Accounting for five full iterations, docking of the test set, and back-end processing, the complete screening workflow can be finished in approximately three to four days.
• The entire implementation is delivered as a single script with adjustable parameters, making it straightforward to configure and deploy for different database sizes and screening objectives.