An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine

summary

The gist

An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine presents a Transformer-based autoencoder AI model trained…

In short

A Transformer-based AI model was trained on ancient Chinese herbal records to quantify how symptoms map to specific herbal treatments. This created an interpretable TCM embedding space (TCM-ES) that measures TCM practice and efficacy. The framework links these traditional principles to modern biology, revealing biological functions and latent disease relationships.

Key concepts

Transformer-based Autoencoder AI
This is a type of artificial intelligence model designed to learn complex patterns. It takes ancient Chinese formula records as input, compresses them into a simplified numerical representation (encoding), and then tries to reconstruct the original information (decoding). This process helps quantify the relationship between symptoms and herbs.
TCM Embedding Space (TCM-ES)
This is a low-dimensional mathematical space where different TCM entities—like symptoms, diseases, and herbal formulas—are represented as numerical points. Points that are close together in this space mean they share similar patterns or principles according to TCM.
Principal Components (PCs)
These are the main directions or axes within the TCM-ES that capture the most significant variation in the data. The paper found that these principal components strongly correlate with key biological functions like metabolism and immune response, linking traditional practice to modern biology.

Terminology used across episodes

This episode discusses

The paper

An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine · Read on arXiv

Department of Physics Hong Kong Baptist University · School of Chinese Medicine Hong Kong Baptist University · Vincent V.C. Woo Chinese Medicine Clinical Research Institute · Centre for Chinese Herbal Medicine Drug Development Hong Kong Baptist University Centre for Interdisciplinary Studies and Department of Physics Westlake University Centre for Interdisciplinary Studies and Department of Physics Westlake University Institute of Systems Medicine and Health Sciences Hong Kong Baptist University Institute of Computational and Theoretical Studies Hong Kong Baptist University

Traditional Chinese Medicine diagnosis and treatment principles, established through centuries of trial-and-error clinical practice, directly maps patient-specific symptom patterns to personalised herbal therapies. These empirical holistic mapping principles offer valuable strategies to address remaining challenges of reductionism methodologies in modern biomedicine. However, the lack of a quantitative framework and molecular-level evidence has limited their interpretability and reliability. Here, we present an AI framework trained on ancient and classical TCM formula records to quantify the symptom pattern-herbal therapy mappings. Interestingly, we find that empirical TCM diagnosis and treatment are consistent with the encoding-decoding processes in the AI model. This enables us to construct an interpretable TCM embedding space (TCM-ES) using the model's quantitative representation of TCM principles. Validated through broad and extensive TCM patient data, the TCM-ES offers universal quantification of the TCM practice and therapeutic efficacy. We further map biomedical entities into the TCM-ES through correspondence alignment. We find that the principal directions of the TCM-ES are significantly associated with key biological functions (such as metabolism, immune, and homeostasis), and that the disease and herb embedding proximity aligns with their genetic relationships in the human protein interactome, which demonstrate the biological significance of TCM principles. Moreover, the TCM-ES uncovers latent disease relationships, and provides alternative metric to assess clinical efficacy for modern disease-drug pairs. Finally, we construct a comprehensive and integrative TCM knowledge graph, which predicts potential associations between diseases and targets, drugs, herbal compounds, and herbal therapies, providing TCM-informed opportunities for disease analysis and drug development.

DOI: 10.1002/imt2.70185

Transcript

Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.

Ines: We're looking at the paper: "An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine".

Title and authors: Ines: So we're looking at this paper today, "An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine". This sounds like it’s trying to take the ancient ways of diagnosing illnesses and treating them with herbs and put them into a system that modern science can actually use.

Marcus: Exactly. It uses this Transformer-based autoencoder AI model, trained on thousands of ancient and classical Chinese medicine formula records, to quantify those symptom patterns and herbal therapy mappings we talked about before. The big idea is to build a quantitative foundation for traditional Chinese medicine principles so it can enhance modern biomedical science.

Yuki: I'm curious how they’ve managed to make that complex mapping into something interpretable. Because when you look at these systems, the interpretation part is often where it gets lost in the math.

Ines: That's what this paper focuses on, and they did a lot of work there. They designed the model with several schemes, including an autoencoder that reduces complex symptom patterns and herb combinations down to measurable vectors. Then they used a contrastive learning mechanism to make sure that symptom patterns and formulas live in the same space, forcing them to be close if they match.

Marcus: And they used transformer blocks with multi-head attention mechanisms in both the encoders and decoders. This lets them quantify self-attentions, which is basically mapping out how symptoms depend on each other and how herb combinations work together, plus cross-attentions for that bidirectional symptom-herb correspondence.

Yuki: It sounds like they are trying to capture the holistic nature of TCM—the idea that the whole pattern matters—and turn it into something mathematically measurable. What’s their main finding on this initial structure?

Ines: Their main finding is pretty neat, because they compared their encoding-decoding process against actual TCM syndromes and found it aligned really closely. This means the model’s way of taking symptom patterns and turning them into something else matches how TCM diagnoses those patterns into syndromes, which is a big deal for showing the principles hold up.

Marcus: And they made sure to exclude the syndrome information from training so that the AI learns objective mappings directly between symptoms and formulas, not just mimicking a pre-defined category. That's smart engineering because it keeps the learning focused on those core empirical relationships.

Yuki: That brings us to how they visualized this whole system. They used principal component analysis, and they found that the first three principal components capture over seventy-two point eight percent of the variance, and six principal components account for more than ninety-eight point seven percent of it. So, the core TCM structure is really concentrated in a relatively small set of dimensions.

Title and authors: Ines: It shows that even though TCM principles are incredibly complex, there's a strong underlying structure we can measure with just a few key dimensions. They also found that herbs with larger projections on the first principal component are mainly classified as hot or warm, which corresponds to the cold syndrome. That’s a direct link between the math and the traditional concept.

Marcus: Moving into what this means for clinical practice, they built this TCM embedding space, or TCM-ES, using that low-dimensional representation. This space is supposed to universally quantify TCM practice by looking at things like patient symptom co-occurrence and coprescriptions of herbs.

Yuki: So if I understand correctly, the proximity of two symptom patterns in the TCM-ES tells us they are more likely to co-occur in real patients, and similarly, if two herbs are close in this space, they tend to be prescribed together more often. That’s a concrete way to measure practice.

Ines: Right. And beyond just practice, they validated the efficacy of the herbal therapies too. They found that a formula will perform better for a patient's condition if their embeddings are more proximal within the TCM-ES, and that symptoms alleviated by the herbs are embedded closer to those prescribed formulas in this space.

Marcus: They even looked at synergistic or antagonistic effects, showing how close herb pairs are in the TCM-ES relates to how enhanced or reduced their therapeutic efficacy is. That’s moving beyond just simple correspondence and into quantifying interaction effects within the system.

Yuki: I’m interested in how they connected this to biology, because that seems like where the real integration with modern science happens. How does this TCM structure relate to actual biology?

Ines: They mapped biomedical entities, like diseases and herbal compounds, into this TCM-ES using correspondence alignment. And what they found was that the principal directions of the TCM-ES are significantly associated with key biological functions, such as metabolism, immune function, and homeostasis.

Marcus: Specifically, the first principal component showed dominant associations with metabolic processes, including both the reactive oxygen species metabolic process and the fatty acid metabolic process. That connects TCM principles to known biochemical pathways.

Yuki: And they didn't stop there; they assessed biological plausibility by comparing how diseases and herbs are spatially related in this TCM-ES against their genetic relationships on the human protein interaction network, or PPI network.

Ines: That’s a powerful check. It suggests that if a disease embedding is close to an herb embedding in the TCM-ES, those entities might also be connected on the actual protein interaction network. It gives molecular plausibility to the TCM structure.

Marcus: And this leads into uncovering latent relationships between diseases themselves. They built a K-nearest neighbour network for seventy nine diseases and found meaningful clusters where diseases with similar MeSH categories grouped together, even if they don't share genes or symptoms in the traditional sense.

Title and authors: Yuki: That’s really interesting for population genetics because it suggests there are underlying pathogenic mechanisms that TCM captures, which might be more general than just individual symptom matching. What about the future?

Ines: They constructed a comprehensive and integrative TCM knowledge graph based on this whole TCM-ES. This graph predicts potential associations between diseases and targets, drugs, herbal compounds, and herbal therapies. It’s a tool for TCMinformed drug development.

Marcus: And they used it to predict potential therapeutic targets for things like rheumatoid arthritis by distinguishing between two clinical subtypes of RA based on Cold and Heat associations with genes forming clusters around those respective subtypes. They even predicted specific targets like BTRC and compounds like Selenomethionine, which are supported by biological evidence or previous studies.

Yuki: So it’s moving from just quantifying practice to predicting novel therapeutic candidates. The future direction they point toward is extending this unified TCM-ES to decode more complex relationships, like direct disease-to-herbal compound links or drug synergistic effects.

Ines: And they also suggest that human diseases might be viewed as evolving trajectories within this multidimensional space, shaped by both biomedical factors and individual differences. This hints at end-to-end associations reflecting underlying complex molecular interactions that we can test experimentally.

Marcus: The authors themselves noted a limitation in that the framework is currently built on ancient and classical records, so extending it to include things like detailed dosage or preparation methods would be a next step for refinement.

Yuki: It’s clear this work provides a very rich quantitative lens for looking at traditional knowledge, linking clinical observation to measurable space and then pushing that into the realm of molecular biology and drug discovery.

Ines: To wrap up this discussion on "An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine", we’ve seen how this Transformer-based autoencoder AI model translates ancient formula records into a structured, interpretable embedding space called the TCM-ES.

Marcus: This TCM-ES allows for the universal quantification of TCM practice, linking symptom co-occurrence and herb co-prescription to therapeutic efficacy across various clinical datasets. We also see how biomedical entities map into this space, showing associations with key biological functions like metabolism and immune processes based on those principal components.

Yuki: And perhaps most exciting is that by mapping diseases onto this space, we get latent disease relationships that show connections between diseases without shared genes or symptoms, which gives us a new way to assess clinical efficacy for modern disease-drug pairs.

Ines: The final output is a knowledge graph that predicts associations between diseases, targets, drugs, herbal compounds, and therapies. This framework provides a quantitative structure to integrate holistic TCM principles with precision medicine approaches.

The paper's summary: Ines: So we’re looking at this paper today about using AI to quantify Traditional Chinese Medicine principles and bridge that gap with modern biomedicine.

Marcus: Yeah, it takes those ancient formula records and trains a Transformer based autoencoder AI model to turn them into measurable vectors for symptom patterns and herbal therapies.

Yuki: It sounds like they’re trying to build a universal language for TCM diagnosis so it can actually talk to modern science.

Ines: Exactly. They developed this system, which they call the TCM embedding space, or TCM-ES. It takes those complex symptoms and herb combinations and compresses them into a low-dimensional map where matched pairs are closer together than unmatched pairs do.

Marcus: And they’re not just looking at symptoms; they’re quantifying practice itself. They use this space to measure how often patients have certain symptoms together, or how frequently specific herbs are prescribed alongside each other.

Yuki: That’s the universal quantification part you mentioned before, but it sounds like it's doing a lot more than just counting prescriptions.

Ines: It does. They mapped modern biomedical stuff—diseases, drugs, proteins—into this TCM-ES through correspondence alignment. And they found that the directions of this space are strongly linked to real biological functions like metabolism and immune responses.

Marcus: The first principal component showed a big link to metabolic processes, including things like reactive oxygen species metabolism and fatty acid metabolism. That connects the ancient structure to actual biochemistry.

Yuki: So you're saying if two diseases or herbs are close in this TCM space, it might actually mean they’re connected on the human protein interaction network?

Ines: Yes. They compared their spatial relationships in the TCM-ES against genetic relationships on that protein interaction network, and it showed biological plausibility for many of these mappings.

Marcus: And they found latent disease relationships too by building a nearest neighbour network for seventy nine diseases. This uncovered connections between diseases that don't share genes or symptoms but still group together based on shared underlying mechanisms in this space.

Yuki: That’s a big thing for population genetics because it suggests there are deep, underlying pathogenic patterns that TCM might be capturing across different conditions.

Ines: And finally, they built a whole knowledge graph from this TCM-ES. This graph predicts potential connections between diseases and targets, drugs, herbal compounds, and therapies. It’s meant to guide personalized drug discovery based on these traditional principles.

Marcus: So the implication is that you can use this framework to look at complex human diseases through a TCM lens and see potential therapeutic targets or drug candidates that might have been overlooked by purely reductionist approaches.

Yuki: It moves the discussion from just individual patient matching to finding broader, latent disease relationships within a holistic system.

Ines: And the challenge now is taking this framework and extending it. They noted that they’re currently trained on ancient and classical records, so incorporating more detailed info like dosage or preparation methods would be a natural next step for refinement.

Marcus: So while this is a strong quantitative foundation, the next frontier is making it even more detailed with richer data sources to really unlock those complex interactions.

The paper's improvements: Ines: We’re talking about how the authors suggest they can make this framework even better going forward.

Marcus: So, they're looking at ways to take this TCM embedding space and add more layers of information beyond just symptom and herb pairs.

Yuki: I wonder what kind of extra data they mean when they talk about adding dosage or preparation methods.

Ines: They mentioned incorporating more detailed information about herbal therapies, specifically the dosage and the preparation methods used for those herbs.

Marcus: That would help move beyond just pairing up herbs to understanding how their specific application context affects the outcome in this space.

Yuki: If you can map that, it could potentially capture a lot more of that holistic view TCM is known for, moving closer to a full treatment picture.

Ines: Exactly. They think human diseases might be better conceptualized as evolving trajectories within this multi-dimensional space shaped by both biomedical factors and individual variability.

Marcus: That implies that the relationships we find aren't just static points but dynamic paths, which is a huge leap for predictive modeling in genomics.

Yuki: It suggests that individual patient outcomes could be seen as following a trajectory determined by the interaction of their specific biology and their TCM pattern within this model.

Ines: And they also suggest that end-to-end associations might reflect underlying complex molecular interactions, which gives us promising candidates for experimental validation in the lab.

Marcus: So, instead of just using the TCM-ES for diagnosis, we could use it to predict potential molecular interactions that lead to therapeutic effects.

Yuki: That brings us back around to the future direction—how this framework can be extended beyond simple distance-based inference to decode more complex relationships like disease-to-herbal compound links.

Conclusion: Ines: So we’ve covered how the authors built this AI framework to quantify Traditional Chinese Medicine principles towards modern biomedicine using a Transformer based autoencoder model trained on ancient records and embedding those concepts into a measurable space called the TCM embedding space, or TCM-ES.

Marcus: Basically, it takes complex symptom patterns and herb combinations and turns them into vectors so we can actually measure how close they are to each other in this new space.

Yuki: And that quantification allows us to see connections between symptoms, herbs, diseases, and even biological functions like metabolism or immune response.

Ines: The big implication is that this system provides a quantitative foundation for TCM principles, which used to be mostly empirical rules based on centuries of trial and error.

Marcus: It changes how we look at the bench-to-clinic gap because now we have a structured way to link traditional knowledge with actual measurable data from clinical practice across different datasets.

Yuki: For someone looking at this from a population genetics standpoint, it suggests that there are underlying mechanisms in TCM that might be more general than just matching symptoms to individual patients.

Ines: And the authors even show how this space uncovers latent disease relationships, meaning they find disease clusters that share underlying mechanisms even if they don't share genes or symptoms directly.

Marcus: The final output is a knowledge graph that predicts potential associations between diseases and targets and drugs, which opens up new avenues for TCMinformed drug development.

Yuki: That’s a pretty powerful tool for researchers trying to find novel therapeutic targets based on these integrated perspectives.

Ines: The paper shows how we can start integrating holistic systems with precision medicine approaches by using this AI structure to map and quantify those traditional principles.

Marcus: It’s a big step in making the ancient practice more scientifically rigorous and scalable.

Yuki: We’ll see how they extend this framework to incorporate more detailed clinical data, like specific dosages, which is what they suggested as future work.

More episodes

← Home