Background

Probiotics are explicitly defined by the International Scientific Association for Probiotics and Prebiotics (ISAPP) consensus as "live microorganisms that, when administered in adequate amounts, confer a health benefit on the host." Based on this dual premise of strain and dosage, research has progressively focused on how their actions in the human body—via mucosal barrier function, immune homeostasis, and metabolic interactions—translate into observable clinical effects. Subsequently, the accumulation of randomized controlled trial (RCT) evidence targeting specific populations and outcomes has led to relatively consistent clinical application consensus for probiotics, initially in areas such as ulcerative colitis ^[1], antidepressant therapy ^[2], and alleviating PD symptoms ^[3].

However, most unmodified probiotic therapies face challenges such as poor colonization capacity and limited therapeutic efficacy. In recent years, with the rapid advancement of microbiomics and synthetic biology technologies, the use of synthetic biology approaches to genetically engineer probiotics—thereby endowing them with specific functions or enhancing their probiotic properties—has demonstrated significant application potential in the field of disease intervention and treatment ^[4-6]. Studies indicate that rationally designed engineered probiotics can precisely modulate the host microenvironment and have achieved preliminary success in treating various complex diseases, including metabolic disorders ^[7-8], cancer immunotherapy ^[9-11], and inflammatory bowel disease ^[12-13].

Fig.1 Mechanism of action of probiotics

From the perspective of microbial composition and design paradigms, probiotic interventions are primarily categorized into two types: single-strain and synthetic consortia:

Single-Strain Therapy

The advantages of single-strain probiotics lie in their clear origin, relatively straightforward mechanisms, and ease of standardization in production and quality control. However, they often exhibit only transient or highly individualized colonization in vivo ^[14], limited by the "niche occupation" of the resident microbiota and inter-host variability. Consequently, their effects tend to diminish after discontinuation, impacting clinical reproducibility and sustainability^[15], and often requiring frequent supplementation to maintain benefits.

Synthetic Consortia Therapy

Multi-strain consortia can enhance robustness and coverage through functional complementarity and division of labor ^[16], potentially improving synergistic colonization. In practice, however, such combinations are often designed empirically, lacking standardized modeling and decision-making frameworks. Variations in strain composition, ratios, and endpoints across studies lead to fragmented evidence, limiting comparability and generalizability^[17].

Fig.2 Comparison of Single-Strain and Multi-Strain Therapeutic Regimens

In summary:

1. Probiotics show therapeutic potential for multiple diseases, with a growing evidence base.
2. Clinical practice and bioengineering are gradually shifting from single-strain approaches to consortium-based therapies.
3. Consortium designs urgently require rational design and genetic engineering, supported by standardized modeling and evaluation frameworks.

Our Inspiration

Based on this, we propose a systematic solution: First, we collected strain information of probiotics and relevant research literature on disease relationships, establishing a comprehensive database that provides services including strain information queries, knowledge graphs, and AI-powered Q&A. Subsequently, we conducted in-depth investigation and clinical practice focused on specific diseases to identify clinical challenges, thereby guiding our further therapeutic design. Ultimately, informed by both our AI platform and clinical insights, we developed targeted module designs for specific diseases—Parkinson's disease—through synthetic biology approaches to engineer bacterial strains with specific therapeutic functions.

Fig.3 Framework of the TJUSX Project

PD

Disease Overview:

Parkinson's disease (PD) is a progressive neurodegenerative disorder. Its core pathology involves the loss of dopaminergic neurons in the substantia nigra and the formation of Lewy bodies. Clinically, PD manifests with motor symptoms such as tremor and bradykinesia, as well as non-motor burdens including hyposmia, constipation, and sleep/mood disturbances. As the disease progresses, patients often develop fluctuating responses to medication.

Fig.4 Pathology of Parkinson's Disease

Existing Therapeutic Approaches:

Pharmacotherapy, with Levodopa as the cornerstone and supplemented by neuroprotective agents, dopamine receptor agonists, and MAO-B/COMT inhibitors, can significantly alleviate motor symptoms. Deep brain stimulation (DBS) is effective for managing motor fluctuations and dyskinesias but is invasive and requires long-term management. Crucially, the majority of small molecules and almost all large-molecule/biologic drugs struggle to cross the blood-brain barrier (BBB). This fundamental limitation restricts the delivery of central nervous system-targeted therapeutics and hinders the development of disease-modifying strategies, representing a critical shortcoming of current treatment paradigms.

Fig.5 Therapeutic Approaches for PD

Nasal Administration:

Intranasal administration bypasses the blood-brain barrier via pathways along the olfactory epithelium and trigeminal nerve, enabling direct access to the brain. This approach offers faster, more localized, and potentially reversible drug exposure. Current research has explored the influence of nasal microbiota on PD and the potential of intranasal microbial administration for treating central nervous system diseases. By integrating multi-strain division of labor, controllable expression systems, and safety switches, it is possible to achieve on-demand, in situ drug production at the nasal mucosa. This strategy could provide an alternative delivery route and engineering framework for CNS disorders such as Parkinson's disease.

Our Solution

In developing our targeted approach for Parkinson's disease, we first leveraged our AI platform to identify potential probiotic strains. This was followed by flux balance analysis (FBA) modeling to optimize the microbial consortium. For the administration strategy, we innovatively adopted an intranasal delivery route to bypass the blood-brain barrier. Furthermore, through synthetic biology engineering, we endowed the selected strains with targeted colonization capability and enhanced therapeutic functions. Additionally, integrated safety modules were designed to address the potential risk of engineered strain escape.

AI platform

In our Human-Practice, we found that some researchers on probiotics like clinicians, translational researchers, product developers, and some consumers who pay attention to probiotics but have limited knowledge about them face persistent difficulties in obtaining accurate, interpretable, and context-specific evidence on probiotic strains. However, current information retrieval remains fragmented, time-intensive, poorly standardized, and limited in interpretability^[18-21]. These shortcomings hinder timely, responsible decision-making and delay downstream research, development, and application.

Under such circumstance, we build a AI Platform of Probiotia Design for Disease (ProbiEase), a comprehensive platform integrating strain information display, natural language Q&A powered by RAG technology, and knowledge graph visualization.

We first curated over 211 probiotic strains(with 27 Genus, 108 Species in total) with their potential or validated therapeutic/intervention roles against various diseases(241in total), supported by corresponding references, and thus building a dataset named BioMatch(More details can be found in the 'Software')

Fig.6 Overview of BioMatch2.0

Based on BioMatch 2.0, ProbiEase can provide three main functions- Strain information display, Q&A via RAG and Knowledge Graph, KG to serve a baseline of systematic, reproducible strain screening pipeline:

Users can explore the graph from the perspective of a strain, from the perspective of a disease, or gain a global overview of the 'probiotic--disease' network. helping users appreciate the complex relationships in a more intuitive and interactive way.

As a result, outputs provide cited, scoped answers and preliminary strains candidate lists that feed into the following modeling module, enabling risk-aware, evidence-backed recommendations aligned with Human Practices expectations.

In our project, we get three recommended strains[ Escherichia coli Nissle 1917, Lactobacillus plantarum WCFS1 and Enterococcus faecium] through asking "Could you recommend several bacterial strains that can treat Parkinson's disease? " on the ProbiEase.

Model - FBA

To further confirm the reliability, transparency, and safety of recommended probiotic strains in human-like environments, we introduce Flux Balance Analysis (FBA) ^[22-25]. This approach simulates metabolic behaviors and interactions, enabling a comprehensive evaluation of potential risks associated with probiotic candidates from the ProbiEase platform.

Fig.7 Schematic diagram of ODE and FBA

First, we employed metabolic flux analysis (FBA) to predict the metabolic flux distribution of the recommended strains under steady-state conditions in simulated human environments, such as an intestinal-like system. By constructing a stoichiometric matrix and optimizing objective functions including biomass growth, FBA successfully simulated the growth of all three strains in the human environment and predicted their associated metabolism.

Concurrently, experimental analysis revealed that Enterococcus faecium possesses the gene for tyrosine decarboxylase which can prematurely metabolize L-DOPA, the primary medication for Parkinson's disease, thereby reducing its therapeutic efficacy. Due to this significant risk of negative drug-microbe interaction, it was excluded from our final consortium.

Fig.8 Pipeline of the whole model

However, during our early experiment, several concerns(like Do these strains exhibit quorum sensing?) arose in our mindsand thus we would like to propose a two-tie modelling in the following work(more details in the 'Model'). Specifically, we desire to extend our static FBA to a dynamic one which introduces introduced ordinary differential equations to model the dynamically changing extracellular environment^[26-30], whilst incorporating timed nutrient supplementation ^[31-32]to simulate a more robust human environment.

Building upon FBA, dFBA extends the analysis to dynamic scenarios by integrating FBA with ordinary differential equations (ODEs) ^[33-34], thereby modelling time-dependent interactions between multiple recommended probiotic strains (such as E.coli Nissle 1917 and Lactobacillus plantarum WCFS1). By simulating nutrient competition, cross- feedings, and community dynamics between them in a human-like environment, dFBA successfully predicts whether interactions produce harmful substances (such as excessive organic acids causing pH imbalance or toxin accumulation) or whether the process itself induces adverse effects (such as inflammation) ^[35-38].

In dry-lab co-culture simulations, dFBA revealed that metabolite exchange between E.coli Nissle 1917 and Lactobacillus plantarum WCFS1 does not amplify risks unseen in single-strain cultures. Simultaneously, the co-culture of both strains significantly enhanced L-Dopa and glutathione, offering improved therapeutic efficacy for Parkinson's disease.

Result

In the simplified FBA model, we observed that the metabolites produced by Nissle 1917 and Lactobacillus plantarum WCFS1 are harmless to humans, whereas those from Enterococcus faecium are detrimental. Consequently, we retained Nissle 1917 and Lactobacillus plantarum WCFS1.

Concurrently, in our subsequent in dry experiments, we employed a static-dynamic dual-dimensional modelling approach—using static FBA to assess the safety of individual metabolites and dynamic FBA to analyse inter-community interaction risks. This enabled us to further simulate the growth behavior of ProbiEase-recommended strains within the human body.

The results were consistent with our static FBA findings:

Ingestion of Enterococcus faecium induces the production of harmful substances in the human body, whereas simultaneous ingestion of E.coli Nissle 1917 and Lactobacillus plantarum WCFS1 does not—they are retained as candidate strains for subsequent wet experiments.

Wet lab engineering:

Following further screening through our AI platform and FBA modeling, we have identified two bacterial strains with potential benefits for Parkinson's disease: Lactiplantibacillus plantarum WCFS1 and Escherichia coli Nissle 1917. In the next phase, we will implement customized engineering strategies tailored specifically for intranasal delivery.

Adhension Module

The olfactory epithelium (OE) is characterized by the apical distribution of N-acetyl heparan sulfate (NaHS) on the axon terminals of olfactory neurons, a distinctive feature that differentiates it from other nasal epithelial cells. Studies indicate that Lactobacillus strains exhibit significantly higher binding efficiency to OE NaHS compared to other bacteria, with Lactiplantibacillus plantarum WCFS1 demonstrating the strongest affinity ^[39].

Fig.9 Schematic representation of probiotic colonies adhering to the olfactory epithelium

In the adhesion module, we employed L. plantarum WCFS1 as a targeted delivery vector for OE colonization. By overexpressing the OppA protein in this strain, we enhanced its affinity for NaHS on epithelial surfaces, thereby improving its colonization capacity in the OE region.

Furthermore, leveraging the principle of specific antigen-antibody interaction, we designed and constructed a ternary bacterial adhesion system. Specifically, we engineered: L. plantarum WCFS1 to express antigen Ag2; One Escherichia coli Nissle 1917 strain to co-express nanobodies Nb2 and Nb3; Another Nissle 1917 strain to co-express nanobody Nb2 and antigen Ag3. Through this design, the three engineered strains can form interlinked pairs via specific Ag--Nb pairing, ultimately assembling into a stable multi-bacterial composite structure.

Therapeutic Module

Long-term monotherapy for Parkinson's disease often leads to challenges such as diminished efficacy and motor complications. Consequently, combination therapy has become a cornerstone of modern PD treatment. In our therapeutic module, we selected two treatment factors with synergistic potential: levodopa and glutathione. Within our proposed strategy, levodopa is responsible for immediate symptom relief, while glutathione theoretically functions to protect neurons from further damage. Through this multi-target, integrated "cocktail therapy" approach^[40], we aim to achieve dual therapeutic objectives: symptomatic improvement and neuroprotection.

Fig.10 Cocktail therapy

Levodopa:
As the gold standard for symptomatic treatment of Parkinson's disease, levodopa is a dopamine precursor molecule capable of crossing the blood-brain barrier and being converted into dopamine. This process compensates for the dopamine deficiency caused by the degeneration of dopaminergic neurons in the midbrain substantia nigra. The drug primarily improves motor dysfunction resulting from the loss of dopaminergic neurons and helps restore normal neurotransmission.

Glutathione:
Glutathione is an essential endogenous neuroprotective molecule that effectively scavenges free radicals and inhibits oxidative stress, thereby mitigating neuronal damage triggered by decreased dopamine levels. By reducing the accumulation of reactive oxygen species and lipid peroxidation, it protects brain cells from oxidative injury.

Control Module:

Achieving controlled drug release is a critical challenge in drug delivery. We aim to establish a tightly regulated system where therapeutic factor release is not only precisely controlled but also deeply coupled with the adhesion module—ensuring that release is triggered specifically upon adhesion, while simultaneously enhancing adhesive capability during the release process.

Fig.11 Schematic design of a control module gene circuit

To better realize functional coupling between the two bacterial strains, we further propose to construct an engineered bidirectional communication platform between Gram-negative and Gram-positive bacteria, applying it to coordinate task allocation within the consortium. Specifically, we engineered to secrete the SppIP peptide, which is sensed by L. plantarum via its SppK/R two-component system, leading to surface display of adhesion proteins. Concurrently, we introduced an AHL synthase into L. plantarum, enabling it to produce AHL that is perceived by the engineered E. coli, triggering the production of therapeutic factors. This design establishes mutual signaling and functional coupling between E. coli and L. plantarum.

Safety Module

Guided by the principles of responsible research, our team has proactively conducted a systematic assessment of potential biosafety risks associated with the practical application of engineered bacterial strains and established corresponding biocontainment strategies. To effectively prevent abnormal colonization of engineered bacteria in the gut and their leakage into the environment, we designed a dual suicide switch system. Additionally, an emergency containment system was implemented to enable rapid system shutdown in case of unexpected incidents. Specifically, for Escherichia coli Nissle 1917, we introduced an L-arabinose-inducible suicide mechanism, while for Lactiplantibacillus plantarum, a nisin-inducible peptide was utilized to trigger self-elimination. This inducible suicide system is designed to minimize potential hazards resulting from accidental release or unintended persistence.