CentraPharma: An End-to-End Multi-Agent System for Drug Design and Delivery

Tian Zihan
School of Computer Science and Technology, Xinjiang University, Xinjiang 830017, China

Abstract:
[Objective] To resolve the "technology silo" problem in drug research and development, where isolated models hinder integration across target screening, molecular design, efficacy evaluation, and clinical translation, lacking a lightweight end-to-end framework. [Methods] We constructed a multi-agent system based on the AutoGen+Ollama framework, integrating six specialized agents—input classification, disease-target mapping, gene translation, ligand generation, ADMET screening, and nanocarrier design—to form a closed-loop pipeline. [Results] Validation using the Alzheimer's BACE1 target generated 404 ligands, yielding 6 druggable molecules through six-dimensional ADMET screening, and outputting customized blood-brain barrier delivery schemes; the system operates on consumer-grade GPUs. [Limitations] The drug delivery design relies primarily on large model outputs and empirical rules, lacking constraints from physical mechanisms such as molecular dynamics simulation; workflow flexibility requires further improvement. [Conclusions] This study pioneers "CentraPharma," a multi-agent system integrating the complete drug design and delivery workflow, achieving end-to-end automated closed-loop from target identification to delivery scheme design, providing a new paradigm for AI-driven efficient drug research and development.

Keywords: Multi-Agent System; Drug Design; Drug Delivery; AI Drug Discovery
Classification Number: TP18

In recent years, the rapid advancement of artificial intelligence technology has reshaped the paradigm of drug research and development. From targeted molecular design to multi-modal data integration, AI-driven innovative tools have significantly enhanced the efficiency and precision of drug discovery. Transformer-based DrugGPT achieved 99.9% validity in molecular generation, while reinforcement learning-optimized DrugGen reduced binding energy by 24% [1]; lightweight deployment platforms like Ollama enable medical edge computing on consumer-grade GPUs through dynamic quantization; and large models such as BioGPT and BioMedGPT provide new perspectives for target mechanism analysis and drug-target relationship prediction through cross-modal knowledge fusion [2-3]. However, despite remarkable breakthroughs in individual stages, end-to-end drug R&D still faces systemic challenges—target screening, molecular design, efficacy evaluation, and clinical translation rely on isolated models or toolchains, lacking lightweight multi-agent collaboration frameworks for integration. This "technology silo" phenomenon restricts data flow, creates inefficient task handoffs, and prevents the formation of a closed-loop innovation chain from basic research to clinical application. How to construct scalable, highly secure multi-agent systems that can uniformly coordinate the entire drug R&D workflow has become a critical bottleneck that the AI pharmaceutical field must urgently address.

2 Related Work

The deep integration of artificial intelligence technology in drug research and development has significantly accelerated the transformation of innovation paradigms. Autoregressive generation models such as DrugGPT, which employ byte-pair encoding (BPE) algorithms to convert protein sequences and ligand SMILES into limited vocabularies and combine them with Transformer architectures to learn protein-ligand binding data, have achieved high-affinity molecule generation for specific targets [4]. Experiments demonstrate that its generated molecules reach 99.9% validity, supporting ligand prompt input and atom count control, providing an efficient tool for drug repositioning and novel pharmacophore design. Furthermore, the Iranian team's DrugGen enhances the DrugGPT architecture through reinforcement learning (PPO) and integrates a binding affinity prediction module, improving the binding energy of generated molecules from 7.22 to 5.81 and significantly enhancing clinical translation potential [1]. Such breakthrough progress marks that AI's precision capability in molecular design has approached experimental validation levels.

At the technical deployment level, the lightweight and local deployment requirements of large models have driven computational paradigm innovation. The Ollama platform reduces hardware barriers through quantization optimization; for example, the dynamically quantized 2.0 version of Qwen3-8B can achieve 45 tokens/second inference speed on 12GB VRAM devices (such as RTX 4070s 12G) while supporting 128K long context and multilingual generation, providing a feasible solution for medical edge computing scenarios. Its simplified configuration strategy (such as Temperature=0.6 to avoid cyclic generation) further enhances deployment efficiency, enabling real-time inference in resource-constrained environments. This progress provides flexible computational infrastructure support for AI-driven drug research and development.

Large model research in the biomedical field has also achieved important breakthroughs. Microsoft's BioGPT, pre-trained on large-scale biological literature, supports drug-target interaction extraction, document classification, and question-answering tasks, with its Hugging Face version achieving 98.6% accuracy in drug interaction prediction [2]. Meanwhile, multi-modal models such as BioMedGPT integrate text, sequence, and structural data, promoting cross-innovation in target mechanism research and drug design [3]; OpenAI's GPT-5 provides assisted diagnosis and treatment recommendations through certified medical knowledge bases (20 million documents) and has passed HIPAA compliance certification, demonstrating AI's compliance and practicality in clinical scenarios [5].

At the system architecture level, multi-agent collaboration frameworks (such as AutoGen) can construct closed-loop drug R&D chains of "target screening-molecule generation-efficacy evaluation" through task decomposition and cyclic feedback mechanisms. Their core advantages include: first, flexible networking capability, where agents can collaborate through natural language instructions without hard coding; second, tool integration capability, compatible with Python toolchains (such as RDKit, PLAPT) and supporting pharmaceutical tool calls; third, local security, ensuring data remains on-premise to meet medical privacy requirements. Such frameworks provide standardized solutions for the division of labor and collaboration in complex drug R&D tasks [6].

Specialized agents for complex biomedical tasks (such as Biomni) further expand AI's application boundaries. Biomni integrates retrieval-augmented planning with code execution capabilities, supporting tasks such as CRISPR experimental design, single-cell sequencing analysis (scRNA-seq), and ADMET prediction [7]. Its modular tool library integrates FDA drug databases and clinical guideline parsers, and connects to external tool servers through the Model Context Protocol (MCP) to achieve functional expansion. Notably, this agent has demonstrated autonomous research capabilities, such as generating verifiable hypotheses from natural language instructions (e.g., "predict 32 genes regulating T cell exhaustion"), providing new ideas for hypothesis-driven drug research and development.

Despite significant progress in specific stages, current research still faces key challenges: existing systems mostly focus on single-task optimization, lacking lightweight multi-agent collaboration frameworks for end-to-end targeted drug R&D. Specifically, stages such as target screening, molecule generation, efficacy evaluation, and clinical trials still rely on independent models or toolchains, making it difficult to form an end-to-end closed-loop system. Moreover, the deployment efficiency of agents and multi-modal data collaboration capabilities in resource-constrained scenarios have not been fully validated. Therefore, constructing lightweight, highly secure multi-agent architectures to integrate the entire drug R&D workflow remains a critical breakthrough for advancing AI pharmaceuticals from laboratory to industrialization.

3.1 Therapeutic Target Database (TTD)

The Therapeutic Target Database (TTD) is a global, free online database developed and maintained by the National University of Singapore, with the core objective of systematically collecting and providing target information related to drug research and development [8-9]. The database currently contains over 3,500 known and exploratory therapeutic protein and nucleic acid targets, along with detailed information on nearly 40,000 drug molecules associated with them. TTD not only provides basic target information but also integrates corresponding targeted diseases, involved pathways, and relevant drugs (including approved and in-development) for each target, clearly demonstrating the multiple association networks among "drug-target-disease." To more precisely guide research and evaluate target value, TTD proposes an efficacy-based target identification strategy that specifically distinguishes between "efficacious" and "non-efficacious" targets, and further subdivides "efficacious" targets with therapeutic potential into four tiers: successful targets (corresponding to approved drugs), clinical trial targets (corresponding to drugs in clinical trials), patent-recorded targets (documented in patent literature), and literature-reported targets (proposed in scientific literature). TTD attaches great importance to data traceability and reliability, with all included information supported by clear and verifiable references. Users can conveniently search by disease name, drug name, or target name to efficiently obtain critical information on potential therapeutic drugs and their targets for specific diseases. Additionally, TTD provides rich cross-links for seamless navigation to internationally renowned bioinformatics databases such as UniProtKB, PDB, KEGG, OMIM, and BRENDA for more in-depth target background knowledge. Since its launch in 2002, the database has undergone continuous content updates and functional optimization, with the latest data updated to January 10, 2024.

3.2 UniProt Knowledgebase

UniProt (Universal Protein) is the most comprehensive and extensive protein database internationally. It was constructed by integrating data from three previously independent databases—Swiss-Prot, TrEMBL, and PIR-PSD. Its core data primarily originates from subsequent protein sequence analysis of completed genome sequencing projects and aggregates extensive protein biological function information reported in literature. The core of the database is the UniProt Knowledgebase (UniProtKB) [10]. UniProtKB itself consists of two main components: UniProtKB/Swiss-Prot is a high-quality, non-redundant dataset that has undergone rigorous manual curation and annotation by experts, with entry information primarily based on published research and quality-controlled through computational analysis (such as E-value verification) to ensure reliability (for example, in the 2023 release, it contained over 560,000 entries); UniProtKB/TrEMBL mainly contains protein sequences annotated through automated pipelines, designed to effectively address the processing pressure from massive data generated by genome projects. It automatically collects and annotates coding sequence translation results from the three major nucleic acid databases (EMBL-Bank/GenBank/DDBJ) and gene prediction sequences from protein structure databases (PDB), Ensembl, Refeq, and CCDS (its entry count is huge, exceeding 220 million in 2023). Additionally, UniProt includes UniProt Archive (UniParc), a comprehensive non-redundant protein sequence archive designed to collect sequences from all major public protein databases. It addresses redundancy issues between different databases and within different versions of the same database by assigning a stable unique identifier (UPI) to each unique protein sequence. Notably, UniParc only stores sequence information itself without any annotation content. With its comprehensive sequence information, rich functional annotation (especially in the Swiss-Prot section), and ability to resolve sequence redundancy, the UniProt database has become an indispensable foundational resource for protein-related research.

3.3 MyGene

MyGene.info is a RESTful API web interface service developed with funding from the National Institute of General Medical Sciences (NIGMS) under the National Institutes of Health (NIH) [11]. Its core objective is to provide users with a simple and easy way to query and retrieve integrated gene annotation information from multiple authoritative sources. MyGene.info synchronizes and updates data weekly from over 20 bioinformatics databases (including NCBI Entrez Gene, Ensembl, UniProt, and UCSC), committed to providing a relatively comprehensive and up-to-date view of gene annotation. Although some of its integrated original data sources may have specific usage restrictions, the service provided by MyGene.info itself is free, with its source code hosted on GitHub. The service mainly provides two core functions: gene query service and annotation retrieval service, both returning structured result data in JSON format. The MyGene.info API continues to be updated and optimized; for example, its v3 version (as of the time of description) includes improvements and bug fixes in data representation (such as including RefSeq accession version information), relationship mapping (such as enhanced RNA-protein associations and improved mapping between Ensembl and Entrez Gene IDs), and data structure. To facilitate users in different programming environments, the MyGene.info community has developed and open-sourced official client libraries, including the MyGene R Client and MyGene Python Client (mygene module), allowing users to conveniently access MyGene.info web services in their own R or Python analysis pipelines. For example, in a Python environment, using the mygene module can easily achieve batch conversion between gene identifiers (such as Entrez ID, Ensembl ID, Symbol, etc.) or query detailed annotation information for specific genes, significantly simplifying the steps of gene data integration.

This study constructs a multi-agent collaborative system based on the AutoGen+Ollama framework, using the Qwen3:8B large model as the core reasoning engine for agents and integrating professional toolchains to achieve automation in biomedical R&D. The system implements an end-to-end automated workflow from target identification to molecule screening through the collaborative operation of six specialized agents: the input classification agent first parses user request types (disease/gene/protein/nanocarrier); the disease-target mapping agent establishes disease-target-protein sequence mapping relationships by integrating the TTD database and UniProt services; the gene translation agent performs bioinformatics conversion from gene identifiers to protein sequences using the MyGene API; the ligand generation agent drives the DrugGPT model for small molecule ligand generation; the ADMET screening agent integrates QED models and five-dimensional pharmacokinetic prediction modules for joint molecular property screening; and the nanocarrier design agent focuses on multi-parameter optimization of blood-brain barrier-penetrating carriers. Each agent exchanges data through strictly defined JSON protocols, adopts hierarchical storage structures for managing output files, and forms a closed-loop automated pipeline for biomedical R&D. The entire workflow architecture is shown in Figure 1 [FIGURE:1].

4.1 Data Processing Agents

Data processing agents constitute the foundational support layer of the system, comprising three core units: input classification, disease-target mapping, and gene translation. These agents collectively accomplish the standardized conversion of biomedical data, transforming unstructured biological entity information into machine-processable protein sequence data to provide structured input foundations for downstream molecular generation. Their design objective is to eliminate error risks from manual data conversion through automated processes and enhance the reliability of the overall R&D pipeline.

The input classification agent employs a Qwen3:8B large model-based semantic parsing engine to achieve initial request classification. This agent processes user input text through a technical route combining regular expression pattern matching and keyword feature extraction. The core processing workflow first applies the re.sub() function to perform text normalization, eliminating special character interference; then uses a predefined regular pattern library for entity recognition, including disease term patterns, UniProt ID patterns, and FASTA sequence feature patterns; and finally resolves semantic ambiguity through the contextual understanding capability of Qwen3:8B. Classification results strictly follow JSON Schema specifications for output, containing two required fields—category and content—to ensure downstream agents receive standardized, unambiguous task instructions.

The disease-target mapping agent constructs a cross-domain conversion channel based on biomedical databases. This agent implements core functionality through the disease_to_protein_sequences() function, with technical implementation comprising three key stages: first, loading the TTD disease-target database text file, using the csv.reader() module to parse the INDICATI field line by line, and applying regular expressions to validate and extract target IDs in compliant formats; then calling the UniProt service interface from the bioservices library to batch retrieve corresponding protein sequences through the UniProt.retrieve() method; and finally using BioPython's SeqIO.write() function to integrate multi-sequence FASTA files. The processing includes a term mapping fault-tolerance mechanism that automatically activates the Levenshtein distance algorithm for similar term expansion retrieval when exact matching fails, ensuring effective mapping of disease term variants (such as "Alzheimer" vs. "Alzheimers"). Output files follow the internationally accepted FASTA format standard, with each record containing complete target identifiers and amino acid sequence information.

The gene translation agent achieves bioinformatics conversion from gene identifiers to protein sequences through the gene_target_to_protein_sequence() function. This agent integrates the mygene public service and UniProt sequence retrieval service to form a two-layer parsing architecture: the first stage calls the mygene.MyGeneInfo.query() interface, setting species="human" and fields="uniprot.Swiss-Prot" parameters to obtain the UniProt ID set corresponding to gene symbols, using a priority sorting algorithm to prefer reviewed entries; the second stage retrieves FASTA format sequence data through the UniProt.retrieve() method from the bioservices library and applies RDKit's Chem.MolFromSequence() for amino acid composition validity verification. To address the challenge of gene naming complexity, a dynamic routing strategy is designed: when standard gene symbol parsing fails, the input is automatically attempted as a UniProt ID for direct sequence retrieval; when retrieval results are abnormal, a homologous gene expansion retrieval mechanism is activated. The final output file adopts standardized FASTA format, with file headers containing complete gene identifiers and UniProt ID mapping information.

This class of agents constructs the core infrastructure for biomedical data conversion, achieving reliable transformation from unstructured biological entities to machine-readable protein sequences through strictly defined function interfaces and standardized data processing workflows. Each agent adopts modular design principles, supporting functional expansion through parameterized configuration, and provides high-quality structured input data for downstream molecular generation.

4.2 Drug Design (Generation) Agent

The drug design agent constitutes the core computational unit of the system, responsible for converting protein target sequences into small molecule ligands with potential biological activity. This agent establishes an automated design workflow from target structure to candidate drugs by integrating the deep learning-driven molecular generation model DrugGPT, providing a data foundation for subsequent ADMET property screening. Its primary function is to transform standardized protein sequences output from the upstream data processing stage into candidate molecular structures in chemical space.

The drug design (generation) agent implements core functionality through the generate_ligands() function, constructing a molecular generation computational workflow based on the DrugGPT framework. The function first specifies GPU computing resources through environment configuration parameters to ensure necessary hardware acceleration support for the molecular generation process; then performs input preprocessing by writing protein sequences into temporary files compliant with FASTA 2.0 standards, with file headers containing target identifiers and generation timestamps as metadata; and finally calls the DrugGPT command-line interface to execute molecular generation tasks. Key runtime parameters include -p for specifying input protein sequence strings, -n for controlling the number of generated molecules (default 50), and -o for defining hierarchical output directory structures. The molecular generation model employs a conditional variational autoencoder architecture, extracting spatial geometric features and physicochemical property distributions of protein binding pockets through three-dimensional convolutional neural networks. These features serve as conditional vectors to guide molecular structure sampling in latent space, ensuring generated ligands match target binding sites in spatial complementarity and interaction patterns. Generation results are stored in structured CSV files, with each molecular record containing four key data fields: molecular hash identifier (128-bit MD5 encrypted string), canonical SMILES representation (following IUPAC standards), generation probability score (0-1 range confidence metric), and estimated binding free energy (ΔG approximation based on molecular mechanics optimization, in kcal/mol).

The drug design (generation) agent forms a tightly collaborative working mechanism with upstream data processing agents. When receiving multi-target FASTA files output from the disease-target mapping agent, it automatically activates the batch_generate_ligands() batch processing workflow, parsing file content through bioinformatics tools and creating independent computational tasks for each target, with parallel processing quantity controlled by the max_targets parameter (default limit of 3 targets). When collaborating with the gene translation agent, it directly accepts single-sequence FASTA file input, validates amino acid composition effectiveness through cheminformatics tools, and automatically extracts gene identifiers as target naming bases. The collaboration process follows strict data contract specifications: input must comply with FASTA 2.0 format requirements, with sequence lines containing no spaces or special characters; output adopts hierarchical directory structure organization; an exception status code mechanism is established, returning "invalid_sequence" codes when sequence validation fails and "resource_unavailable" status when GPU resources are insufficient; all error logs are automatically recorded in dedicated report text files in the output directory. This agent constructs an automated design bridge from biological targets to candidate drugs, with its deep learning model achieving integrated computation of structure generation and preliminary activity assessment. Modular design ensures seamless connection with upstream data processing stages, standardized output structures provide reliable input for downstream ADMET screening, and strictly defined data contracts and exception handling mechanisms significantly enhance the efficiency and reproducibility of early-stage drug R&D workflows.

4.3 ADMET Drug Screening Agent

The ADMET drug screening agent constitutes the core of the system's pharmacokinetic evaluation, responsible for comprehensive assessment of absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of candidate molecules output from the ligand generation stage. This agent establishes quantitative associations between molecular properties and druggability through integrated multi-parameter prediction models, providing critical decision-making bases for subsequent nanocarrier design and experimental validation. Its primary function is to systematically screen candidate molecules in chemical space to identify potential drug candidates with favorable pharmacokinetic characteristics.

The ADMET drug screening agent implements core functionality through the admet_filter_tool() function, constructing a comprehensive evaluation workflow based on RDKit and ADMET prediction models. The agent first locates molecular data files output from the ligand generation stage (hash_ligand_mapping.csv) through the find_csv_directories() function; then calls the ADMET prediction model to perform multi-dimensional property assessment for each molecule, with key computational metrics including blood-brain barrier permeability (BBB_Martins), drug-likeness score (QED), lipophilicity (logP), polar surface area (tpsa), cardiac toxicity risk (hERG), and mutagenicity (AMES). The screening process employs a composite conditional decision tree: first applying the QED.qed() function to calculate drug-likeness scores to ensure molecules meet basic Lipinski rule requirements; then constructing a druggability screening funnel through multi-parameter joint constraints (BBB≥0.70, QED≥0.50, 1.0≤logP≤4.5, tpsa≤90, hERG≤0.5, AMES≤0.5). Molecules passing the screening are stored in CSV format, with each record retaining original hash identifiers and SMILES structures to ensure data traceability. The technical implementation adopts modular design, encapsulating prediction algorithms through the ADMETModel() class to support batch processing and incremental updates.

The ADMET drug screening agent forms an end-to-end collaborative workflow with the ligand generation agent. When the ligand generation stage completes computation, it automatically triggers the screening task, precisely locating input files through hierarchical directory structures (output root directory/target ID/seq_min). The collaboration follows standardized data contracts: input files must contain hash_ligand_mapping.csv and comply with hash identifier-SMILES binary structure; output files adopt the _filtered.csv suffix naming convention, retaining original directory structures. Exception handling mechanisms include automatically marking invalid SMILES structures as "invalid_smiles" and recording molecules with prediction failures as "prediction_error." Screening results are fed back to the system hub through JSON format status reports, containing the number of successfully processed files, output path lists, and detailed screening statistics to provide decision support for subsequent nanocarrier design.

This agent constructs a pharmacokinetic evaluation system for candidate molecules, significantly improving the druggability potential of candidates through multi-parameter joint screening mechanisms. Its modular design and standardized interfaces ensure seamless connection with upstream ligand generation stages, and quantitative screening standards provide high-quality input for nanocarrier design and in vitro/in vivo experiments, substantially reducing late-stage R&D failure risks and optimizing drug R&D resource allocation efficiency.

4.4 Drug Delivery Agent Design

The drug delivery design agent addresses the critical technical challenge of candidate drug molecule delivery to the central nervous system. This module focuses on the blood-brain barrier penetration problem, designing nanocarrier systems with targeting functions through computation-driven material selection and structure optimization. Its core objective is to construct safe and efficient delivery solutions for active molecules optimized by the ADMET screening stage.

The drug delivery design agent employs a rational design approach based on biomaterial databases. The design workflow consists of three continuous stages: first, material compatibility analysis, where optimal carrier matrices are selected from a pre-validated material library (containing 12 pharmaceutical polymers including PLGA, PEG-PLGA, and chitosan) based on physicochemical parameters such as drug molecule logP values, molecular weight, and polar surface area; then carrier structure optimization, where key parameters including particle size range (80-120 nm), surface charge (-5 to +5 mV), and drug loading capacity are determined through hydrodynamic simulation; finally, targeting modification scheme design, where corresponding targeting ligands (such as transferrin, Angiopep-2 peptide) are selected based on target receptor expression characteristics (such as transferrin receptor, low-density lipoprotein receptor), and optimal modification density is calculated (2-4 ligand molecules per square nanometer). The design scheme output comprises three parts—carrier composition, preparation process, and critical quality attributes—forming a complete formulation scheme that can be experimentally validated.

This agent establishes a transformation pathway from active molecules to delivery systems, with its empirical data-based decision-making mechanism providing reliable guidance for nanoformulation development. By systematically integrating material property databases, physiological barrier models, and target expression data, it significantly enhances the design efficiency of central nervous system delivery schemes, laying a technical foundation for subsequent experimental research.

4.5 Web Interaction Interface Design

The system employs the Streamlit framework to construct a lightweight web interaction interface as an intuitive interaction channel between users and the agent cluster. This design enables complete visualization of the drug R&D workflow, allowing researchers to directly access system functions through browsers without programming knowledge or command-line operations. The interface adopts a conversational interaction mode that simulates natural research collaboration processes while providing real-time workflow status feedback. The actual web interaction interface is shown in Figure 2 [FIGURE:2].

The interaction system architecture comprises three core modules: 1) a conversation management module that maintains interaction history between users and agents, achieving state persistence through st.session_state; 2) a task distribution module that automatically routes to corresponding agents based on user input types; and 3) a result display module that transforms complex technical results into readable natural language descriptions. Security mechanisms employ session isolation design to ensure data from different users remains separate.

At the implementation level, the interface initializes through st.set_page_config() to set titles and layouts, creating a central control panel. The user input area implements real-time conversational interaction through st.chat_input(), with agent responses displayed in chat bubble formats. The core processing workflow is implemented through asynchronous task mechanisms: user input is first passed to the classification agent for task type identification, which parses text semantics based on the Qwen3:8B model to accurately distinguish between disease-target mapping, gene translation, ligand generation, or nanocarrier design tasks; according to classification results, the system automatically calls corresponding agent workflows—disease names trigger the disease-target mapping agent to access the TTD database, gene identifiers activate the gene translation agent to query the MyGene API, and protein sequences are passed to the ligand generation agent to drive DrugGPT computation. All processing results are encapsulated into InputInfo objects, whose data structures transform technical parameters into natural language descriptions through overridden str methods, and these structured results are ultimately displayed in the interaction interface's history area through Streamlit's st.chat_message() component in chat bubble format, enabling researchers to intuitively view complete workflow records.

The user operation process is intuitive and straightforward: researchers input natural language instructions in the chat box (such as "Convert disease Alzheimer to protein target sequences and develop related drugs"), and the system automatically parses requirements and triggers corresponding agent workflows. Status prompts are displayed in real-time during processing, with final results containing key information (such as target quantity, file paths) and subsequent operation suggestions. For multi-step tasks (such as disease→target→ligand generation→ADMET screening), the system automatically chains workflows, reducing manual user intervention.

This design achieves visual access to complex agent systems through the Streamlit framework, significantly lowering the usage threshold. The conversational interaction mode aligns with researchers' work habits, real-time status feedback enhances system transparency, provides an efficient collaboration platform for cross-disciplinary teams, and effectively promotes the engineering application of computational drug research and development.

5 Full Process Example: Alzheimer's Disease Targeted Drug R&D

This chapter demonstrates the complete system workflow from target to delivery scheme design using the Alzheimer's disease key target β-secretase 1 (BACE1) as an example. The target FASTA sequence has a length of 509 amino acids (sequence identifier: MAQALPWLL L...QHDDFADDISLLK), serving as the input foundation for the workflow.

5.1 From Target to Ligand

Using the Alzheimer's disease key target β-secretase 1 (BACE1) as the research object, we input its FASTA sequence composed of 509 amino acids. This sequence is processed by the ligand generation agent driving the DrugGPT model, which employs a three-dimensional convolutional neural network to extract spatial topological features of the BACE1 binding pocket and performs directed sampling in chemical space to generate ligand structures. The generation process strictly follows preset parameters: inputting the complete target sequence and setting the generation quantity to 404 candidate molecules. Output results are stored in hash identifier-SMILES key-value pair format, with each molecule assigned a 128-bit MD5 hash value as a unique identifier (such as 751c49fa3f906b94948f4ae22bea329004b24bae) and recording the corresponding canonical SMILES structure. All generated molecules are completely saved in csv files, achieving automated conversion from biological target structures to chemical ligands.

5.3 Corresponding Delivery Scheme

The 404 candidate molecules enter the ADMET screening workflow for multi-dimensional pharmacokinetic joint evaluation. Screening criteria require blood-brain barrier permeability (BBB_Martins) not less than 0.70, drug-likeness score (QED) not less than 0.50, lipophilicity (logP) controlled between 1.0 and 4.5, polar surface area (tpsa) not exceeding 90Ų, cardiac toxicity risk (hERG) not exceeding 0.5, and mutagenicity (AMES) not exceeding 0.5. The screening process is implemented through the admet_filter_tool function: first loading the pre-trained ADMET prediction model, then batch calculating physicochemical properties and biological activity parameters for each molecule, applying composite conditions for Boolean logic screening, and finally generating a subset of molecules meeting all criteria. The six candidate molecules passing screening (BACE1-001 to BACE1-006) and their complete ADMET parameters are recorded in the filtered_preds.csv file, where typical molecules such as BACE1-001 exhibit characteristic parameters including logP=3.66, BBB=0.59, and tpsa=77.15Ų.

Based on the ADMET characteristics of the six candidate molecules, the system generates customized delivery schemes. For high lipophilicity molecules BACE1-001 and BACE1-002, an LRP1-targeted PLGA-PEG polymeric micelle system is designed, with PLGA(50:50) as the core material, Angiopep-2 peptide as the surface targeting ligand, and co-loaded P-gp inhibitor Tariquidar to overcome efflux effects, with particle size strictly controlled in the 80-120nm range. For low blood-brain barrier permeability molecules BACE1-003 and BACE1-004, a dual-targeting responsive nanoparticle is constructed, with a DPPC/cholesterol lipid-polymer hybrid structure, surface-modified with transferrin and RVG29 peptide (1:1 ratio), achieving reactive oxygen species-responsive drug release through thioether bonds, with particle size range of 70-100nm. For balanced molecules BACE1-005 and BACE1-006, an enzyme-pH dual-responsive mesoporous silica carrier is developed, using mesoporous silica as the matrix (pore size 4nm), achieving drug loading through pH-sensitive hydrazone bonds, employing MMP-9 cleavable peptides as gate systems, and covering with chitosan-g-PEG coating to improve biocompatibility. All schemes detail material composition, modification density, and process parameters, providing clear guidance for experimental translation.

This case completely demonstrates the system's end-to-end R&D capability: starting from target sequence input, through ligand generation and multi-dimensional ADMET screening, to final customized delivery scheme output, forming a computation-driven drug R&D closed-loop. Each stage achieves seamless connection through standardized file formats, providing an efficient research paradigm for neurodegenerative disease drug development.

6 Summary and Outlook

This study constructs a multi-agent drug R&D system based on the AutoGen+Ollama framework, achieving for the first time a fully automated closed-loop from target identification to delivery scheme design. The system innovatively integrates six functional modules with Qwen3:8B large model as the agent reasoning core: an input classification agent for natural language task parsing; a disease-target mapping agent establishing disease-target association networks through the TTD database; a gene translation agent completing bioinformatics conversion from gene identifiers to protein sequences; a ligand generation agent driving DrugGPT for directed generation from target structures to ligands; an ADMET screening agent constructing a six-dimensional pharmacokinetic joint evaluation funnel; and a nanocarrier design agent outputting customized solutions for central nervous system delivery challenges. In the Alzheimer's BACE1 target validation case, the system successfully converted a protein target sequence of 509 amino acids into 404 candidate ligands, obtaining 6 druggable molecules through rigorous screening and generating matching delivery schemes. Core innovations include: pioneering an agent-based integrated architecture for the complete drug R&D workflow, achieving consumer-grade GPU deployment through lightweight design; deeply integrating professional toolchains (TTD/UniProt/DrugGPT) with large model capabilities to form a digital R&D chain; and transforming complex technical workflows into natural language operations through a Streamlit-based conversational interaction interface. Current limitations mainly manifest in delivery design relying on large model outputs without physical knowledge constraints such as molecular dynamics simulation, and workflow flexibility requiring improvement. Future research will focus on embedding physics-based validation mechanisms such as free energy perturbation calculations in the delivery stage, developing customizable workflow engines, and expanding to emerging paradigms such as antibody drugs, accelerating AI pharmaceuticals from computational design to clinical translation.

[1] Sheikholeslami Mahsa, Mazrouei Navid, Gheisari Yousof, et al. DrugGen enhances drug discovery with large language models and reinforcement learning. [J]. Scientific reports, 2025, (1): 13445.
[2] Luo Renqian, Sun Liai, Xia Yingce, et al. BioGPT: generative pre-trained transformer biomedical text generation and mining. [J]. Oxford University Press (OUP), (6).
[3] Kai Zhang, Rong Zhou, Eashan Adhikarla, et al. A generalist vision–language foundation model for diverse biomedical tasks. [J]. Nature Medicine, 2024, 30: 3129-3141.
[4] Li Yuesen, Gao Chengyi, Song Xin, et al. DrugGPT: A GPT-based Strategy for Designing Potential Ligands Targeting Specific Proteins. [Z]: Cold Spring Harbor Laboratory.
[5] Singhal Rahul-K. Arora-Jason-Wei-Rebecca-Soskin-Hicks-Preston-Bowman-Joaquin-Quin˜onero-Candela-Foivos-Tsimpourlas-Michael-Sharman-Meghan-Shah-Andrea-Vallone-Alex-Beutel-Johannes-Heidecke-Karan. HealthBench: Evaluating Large Language Models Towards Improved Human Health. [J],
[6] Chin Shuang-Ying, Why Dr.-Ng-Kok. Comparative of Multi-Agent System Frameworks: Crewai, Langchain, and Autogen. [Z]: Elsevier BV.
[7] Kexin Serena-Zhang-Hanchen-Wang-Yuanhao-Qu-Yingzhou-Lu-Yusuf-Roohani-Ryan-Li-Lin-Qiu-Gavin-Li-Junze-Zhang-Di-Yin-Shruti-Marwaha-Jennefer-N.-Carter-Xin-Zhou-Matthew-Wheeler-Jonathan-A.-Bernstein-Mengdi-Wang-Peng-He-Jingtian-Zhou-Michael-Snyder-Le-Cong-Aviv-Regev-Jure-Leskovec. Biomni: A General-Purpose Biomedical AI Agent. [EB/OL]: Cold Spring Harbor Laboratory, 2025. https://doi.org/10.1101/2025.05.30.656746. 10.1101/2025.05.30.656746.
[8] Zhu Feng, Han BuCong, Kumar Pankaj, et al. Update of TTD: Therapeutic Target Database. [J]. Oxford University Press (OUP), (suppl_1): D787-D791.
[9] Chen X.. TTD: Therapeutic Target Database. [J]. Oxford University Press (OUP), (1): 412-415.
[10] Uniprot consortium The. UniProt: the universal protein knowledgebase. [J]. Oxford University Press (OUP), (5): 2699-2699.
[11] Welcome to MyGene.py’s documentation! — MyGene.py v3.1.0 documentation. [EB/OL]. https://docs.mygene.info/projects/mygene-py/en/latest/.

(Corresponding author: Tian Zihan E-mail: 3208566786@qq.com)