Introduction
Our models establish a comprehensive computational framework that enables a closed-loop design for novel oral biotherapeutics—from macroscopic delivery to microscopic regulation. At the macroscopic level, a mathematical model integrates pharmacokinetics with an “effective-period” gene circuit, providing optimized strategies for drug delivery and safety. On the molecular regulation level, we not only designed BioPROTAC to efficiently degrade the key inflammatory factor HIF-1α, but also applied machine learning to direct the evolution of a VEGF antibody, endowing it with target-specific activation through a masking peptide to ensure safe administration. Additionally, we developed the Mol2Aptamer model, which can directly generate aptamer RNA sequences based on molecular features, aligning with our project's concept of synergistic medicine and nutrition. Together, these four interconnected models form an efficient, safe, and intelligent therapeutic system.
Mathematics Model
To evaluate the performance and safety of our oral therapeutic system, we developed a two-part Mathematics Model integrating pharmacokinetics and genetic regulation. Model 1 quantifies the in vivo transport, absorption, and distribution of OMVs to predict the concentration of therapeutic proteins at hemorrhoidal sites. Model 2 introduces an expiry-date circuit that controls the survival time of engineered bacteria. Together, these models bridge drug delivery efficiency with biological containment, providing a theoretical foundation for optimizing dosage, timing, and safety in our oral bio-therapeutic design.
BioPROTAC Design
We designed a novel BioPROTAC to selectively degrade the key transcription factor HIF-1α under hypoxic conditions, offering a new therapeutic strategy for hemorrhoids. The workflow involved structural prediction using AlphaFold 3.0, followed by molecular dynamics (MD) simulations to assess stability. Molecular docking was then applied to identify suitable ubiquitination sites, after which refined MD simulations and reaction energy calculations were performed to evaluate the structural and kinetic feasibility of the designed BioPROTAC.
Antibody Design
This project proposes an integrated framework for the design of camelid VHH antibodies targeting VEGF165. The workflow combines machine learning–based prediction of binding free energy changes (ΔΔG) using a low-rank bilinear model trained on the AbBind dataset, and structural quality assessment (Clashscore) via a random-forest model. These predictions are integrated through a weighted, rank-based module to prioritize VHH variants with improved affinity and structural stability.
To ensure safety, we incorporated the Probody strategy by introducing a conditionally activatable masking system. A de novo designed cyclic peptide was generated and optimized using RFdiffusion and ProteinMPNN to block the antibody’s active site. This peptide was linked through an MMP3-cleavable rigid linker, enabling activation specifically at pathological sites. The design was validated through molecular simulations, including Steered Molecular Dynamics (SMD) and Umbrella Sampling (US), to evaluate the peptide’s stability and controlled activation behavior.
Aptamer de novo Design
We designed Mol2Aptamer, a generative deep learning model to solve the challenge of designing aptamers for specific small molecules. Using a conditional Transformer with variational and multi-instance learning, it generates RNA sequences from molecular features. Validated on hippuric acid, Mol2Aptamer-produced aptamers show higher binding affinity and stability, supporting personalized riboswitch design under the “pharmaceutical–nutritional collaboration” concept.