Predicting How Molecules Interact with Proteins Using Two AI Streams

This patent describes a computer system that uses two separate artificial intelligence models, one for molecules and one for proteins, to predict how they will interact and what biological effects they might have.

What it covers

The system predicts how a target molecule (a ligand) will interact with a target protein segment and what biological effect that interaction will have. It does this by first taking the chemical notation of the target molecule and processing it through a trained graph-based neural network to get a "first vector result" (Claim 1). Separately, it takes the chemical notation of the target protein segment and processes it through a trained sequence-based neural network to get a "second vector result" (Claim 1). These two results are then combined, or "concatenated," and used to make a prediction about the bioactivity. For example, a drug researcher could input a new drug candidate molecule and a protein known to be involved in a disease, and the system would predict if the drug is likely to bind to the protein and what effect it might have.

What it doesn't cover

• —Does not cover systems that use only a single machine learning model to analyze both the ligand and protein simultaneously, rather than two separate streams.
• —Does not cover prediction systems that do not concatenate distinct vector results from separate ligand and protein processing streams.
• —Does not cover systems that predict interactions without using a graph-based neural network for the ligand analysis.
• —Does not cover systems that predict interactions without using a sequence-based neural network for the protein segment analysis.
• —Does not cover methods that do not receive chemical notation as input for both the target molecule and target protein segment.
• —Does not cover systems that predict interactions but do not specifically output a prediction as to the bioactivity.

The clever bit

The novelty lies in combining two specialized machine learning approaches: a graph-based neural network for molecules (which excel at understanding chemical structures) and a sequence-based neural network for proteins (which are good at understanding protein sequences), and then merging their distinct analytical outputs to make a unified prediction.

Why it matters

Predicting how molecules interact with proteins is a fundamental challenge in drug discovery and materials science. This system aims to speed up the early stages of drug development by computationally screening potential drug candidates. By accurately predicting bioactivity, researchers can prioritize promising compounds and avoid costly experiments on ineffective ones, potentially leading to new medicines or materials faster.

Real-world examples

• 1.Drug discovery platforms for new pharmaceuticals
• 2.Material science research for designing new polymers or catalysts
• 3.Toxicology screening for environmental safety assessments
• 4.Biotechnology applications for enzyme engineering