AI for Science | Wenrui Fan

Classifying the stoichiometry of virus-like particles with interpretable machine learning

Tue, 18 Feb 2025 00:00:00 +0000

AI-driven anticancer peptide de novo design

Mon, 01 Jan 0001 00:00:00 +0000

Conventional cancer treatments can damage healthy tissues and cause substantial off-target toxicity, creating an urgent need for more precise and safer therapies. Anticancer peptides are promising because they can selectively disrupt cancer cells while potentially reducing harm to healthy cells. However, natural anticancer peptides are limited in number and diversity, and identifying effective candidates through large-scale experimental screening is costly and time-consuming. We therefore developed an AI-driven framework for the de novo design of novel anticancer peptides, enabling cancer-type-specific candidate generation and prioritisation followed by experimental validation.

Multistate Protein

Mon, 01 Jan 0001 00:00:00 +0000

Protein research spans multiple connected states and representations: understanding and designing proteins requires models to connect sequence, structure, function, interactions, and design. However, most current systems and benchmarks evaluate these problems independently. Strong performance in one state does not necessarily mean that a model can preserve biological constraints, transfer useful information, or support decisions in another.

MuSProt is motivated by this gap. It investigates protein AI as a connected multistate problem rather than a collection of isolated prediction tasks. The aim is to understand how knowledge can be transferred between states, how one prediction influences subsequent decisions, and how a model can remain biologically coherent across the complete workflow.

MusBench provides the evaluation setting for this research. Beyond measuring performance on individual tasks, it is intended to examine cross-state transfer, consistency, robustness, and error propagation. Together, MuSProt and MusBench aim to make multistate protein modelling measurable and to identify where current protein AI systems succeed or break down.

Self-Driving Lab for Metal-Binding Protein Design

Mon, 01 Jan 0001 00:00:00 +0000

Generative models can produce hundreds of candidate metal-binding proteins, but generation alone does not identify which designs will bind the intended metal, reject competing ions, remain structurally stable, and justify the cost of wet-lab synthesis. The central challenge is therefore selection and iterative improvement rather than simply producing more candidates.

This project develops a self-driving laboratory for metal-binding protein design. An agentic loop turns a scientific objective into candidate structures, ranks them with CCDC-informed selectivity and stability filters, verifies their coordination geometry and physical plausibility, records the evidence in durable memory, and uses diagnostic feedback to plan the next design cycle. The goal is to reduce a large generated pool to a small, auditable set of candidates worth experimental validation.