JXIONG

Posted on May 11

Can an Algorithm Find What Your Body Actually Needs? Introducing SEMO for Longevity Technology

#ai #algorithms #datascience #science

Everyone in longevity technology is asking a familiar question:

What should I take?

A supplement?

A metabolite?

A senolytic candidate?

A lifestyle intervention?

A drug-repurposing lead?

But maybe that is not the best first question.

A better one might be:

Where, inside my biological network, is there a measurable mismatch that an intervention could plausibly reshape?

That is the problem the SEMO algorithm is designed to address.

SEMO is not just another recommendation engine. It is a network-medicine algorithmic framework developed by DeepoMe to connect individual omics signals, compound target networks, and personalized intervention hypotheses.

The algorithm was introduced by Jianghui Xiong in a bioRxiv preprint titled “Utilizing Pre-trained Network Medicine Models for Generating Biomarkers, Targets, Re-purposing Drugs, and Personalized Therapeutic Regimes: COVID-19 Applications.” In that paper, SEMO stands for Selective Remodeling of Protein Networks by Chemicals.

The SEMO framework has also moved beyond a conceptual proposal. A related Chinese invention patent, “Method, system and application for generating compound intervention schemes based on a pre-trained model”, has been granted under publication number CN117766054B.

In the broader vision of steerable biomedical AI, SEMO can be understood as one possible algorithmic layer beneath a larger question raised by Steerable World:

Can we move from predicting biological decline to steering biological state?

The Longevity Problem: Too Many Signals, Too Little Direction

Longevity science has no shortage of measurements.

We can measure:

DNA methylation age
inflammatory markers
metabolic biomarkers
microbiome composition
gene variants
wearable signals
proteomic and metabolomic profiles
sleep, glucose, HRV, and exercise response

The problem is no longer simply “not enough data.”

The problem is that most data do not automatically tell us what to do next.

A biological age clock may say that someone is aging faster than expected.

A blood test may show a few abnormal markers.

A wearable may show poor recovery.

A supplement database may list hundreds of potentially beneficial compounds.

But how do we connect these pieces into an individualized intervention hypothesis?

Most current systems still rely heavily on population-level logic:

people with marker X often benefit from nutrient Y
compound A has been associated with pathway B
supplement C is popular for aging-related mechanism D
risk score E is high, so generic intervention F is recommended

This can be useful, but it is not enough for true precision longevity.

Longevity is not a one-marker problem.

It is a network-state problem.

From Deficiency Thinking to Network Gap Thinking

Traditional health recommendations often begin with a deficiency model:

What nutrient is low?

SEMO points toward a different model:

What network region shows a compound-relevant state gap in this individual?

This distinction matters.

A person may not be “deficient” in a simple nutritional sense. Yet their biological network may still show a local mismatch: a compound's known target region may differ from its surrounding molecular background in a way that is visible through omics data.

That difference can be treated as a network gap.

In simple terms, SEMO asks:

Which proteins or genes are targeted by a compound?
Where do those targets sit inside the human protein–protein interaction network?
What is the individual’s omics state around those targets?
Is the target region different from the nearby non-target background?
Could that difference suggest a personalized intervention hypothesis?

This is a very different logic from “this supplement is good for everyone.”

It is closer to:

This compound maps to a network region that appears unusually relevant to this person’s current biological state.

What SEMO Does Algorithmically

At a high level, SEMO combines several ideas from network medicine and representation learning.

It can be described as a pre-trained network-medicine framework that:

maps compounds to known or predicted biological targets
embeds those targets into protein–protein interaction networks
constructs reusable compound–network representations
compares target-associated regions with local network backgrounds
integrates individual omics signals, such as DNA methylation-derived features
generates ranked hypotheses for biomarkers, targets, drug repurposing, or personalized intervention candidates

The key idea is that a compound should not be treated only as a chemical name.

A compound is also a network perturbation hypothesis.

It may influence a set of targets.

Those targets sit inside biological modules.

Those modules may correspond to aging-related functions such as inflammation, metabolism, mitochondrial adaptation, immune regulation, stress response, repair, or cellular resilience.

When an individual’s omics data are mapped onto these same network structures, the algorithm can ask whether a compound-relevant region appears meaningfully different from the local background.

That is where SEMO becomes interesting for longevity.

From Preprint to Patent: SEMO Has Already Been Demonstrated

The original SEMO paper did not present the algorithm only as a theoretical idea. It used COVID-19 as a demonstration case.

In the preprint, Xiong described SEMO as a pre-trained network medicine model that divides the global human protein–protein interaction network into smaller sub-networks, then quantifies the potential effects of chemicals by statistically comparing target and non-target gene sets.

The study combined 9,607 PPI gene sets with 2,658 chemicals to create a pre-trained pool of SEMO features. These features were then applied to DNA methylation profiling data from two clinical COVID-19 cohorts to identify SEMO patterns associated with COVID-19 severity.

One important result was that nutraceutical-derived SEMO features could be used to predict COVID-19 severity, with reported AUC values of approximately 81% in the training data and 80% in independent validation data.

That COVID-19 demo matters because it shows SEMO’s intended use case: not simply describing compounds, but linking compound-associated network effects with individual molecular states and clinically relevant outcomes.

The later Chinese invention patent further signals that SEMO-related methods have been formalized as an applied technical system for generating compound intervention schemes from pre-trained models. For longevity technology, this is important because it suggests that SEMO can be viewed not only as a research algorithm, but also as an IP-backed computational infrastructure for personalized intervention hypothesis generation.

Why This Matters for Longevity Technology

Longevity interventions are difficult because aging is not one disease and not one pathway.

Aging involves many interacting processes:

mitochondrial decline
chronic inflammation
immune remodeling
epigenetic drift
stem-cell exhaustion
proteostasis stress
metabolic inflexibility
cellular senescence
impaired stress adaptation
reduced repair capacity

If we treat aging as a list of hallmarks, we still face a practical problem:

Which hallmark matters most for this person, now?

SEMO offers a possible computational route.

Instead of asking whether a compound is generally anti-aging, SEMO can help ask whether a compound’s network region is specifically relevant to an individual’s current molecular state.

That turns longevity intervention from a generic recommendation problem into a structured hypothesis-generation problem.

For example:

A compound associated with mitochondrial targets may not be equally relevant to every older adult.
A polyphenol with inflammatory and metabolic targets may matter more in one network state than another.
A repurposed drug may appear promising only when its target region aligns with an individual's molecular mismatch.
A lifestyle or nutritional intervention may need to be evaluated by the network response it induces, not by its label.

This is the real promise: not “the best supplement,” but the best next hypothesis for this biological network state.

SEMO as a Bridge Between Network Medicine and Steerable AI

The SEWO framework introduced at Steerable World argues that biomedical AI should become steerable, not merely predictive.

A steerable biomedical model should be able to represent state, simulate intervention-induced transitions, inspect failure, and revise the next hypothesis.

SEMO can be viewed as a more concrete algorithmic component inside this broader vision.

If SEWO asks:

How do we steer biological trajectories?

SEMO asks:

Which compound-linked network regions may be worth steering first?

This makes SEMO complementary to a steerable medicine world model.

A world model needs:

a state representation
candidate interventions
intervention-response semantics
counterfactual transition logic
feedback and quality control

SEMO contributes to the candidate-intervention layer by converting compound information and individual omics data into ranked, network-aware hypotheses.

In other words, SEMO helps transform a massive intervention search space into a smaller, more biologically interpretable set of possibilities.

From Recommendation Lists to Personal Science

Many precision-health products still generate static recommendation lists.

You take a test.

You receive a report.

The report suggests supplements, foods, lifestyle changes, or risk categories.

But longevity technology should not stop there.

A more powerful model is longitudinal personal science:

Measure an individual’s biological state.
Identify network gaps or state mismatches.
Generate intervention hypotheses.
Apply a safe, clinically appropriate intervention.
Re-measure the state.
Ask whether the expected network gap changed.
Keep, revise, or discard the hypothesis.

SEMO is valuable because it fits into this iterative loop.

It does not have to claim that an intervention will definitely work.

Instead, it can generate a testable network hypothesis:

This compound-related network region appears relevant. If the hypothesis is correct, a suitable intervention should move the corresponding molecular state in a measurable direction.

That is a much more scientific formulation than a one-time recommendation.

It also aligns with the future of N-of-1 longevity studies, where the goal is not to prove that one intervention works for everyone, but to understand which intervention changes which state in which individual.

Why Network Gaps Are Better Than Generic Rankings

A generic ranking might say:

compound A is popular
compound B has strong literature support
compound C affects many aging pathways
compound D has antioxidant activity

A SEMO-style ranking asks something more specific:

does compound A map to this person’s relevant network region?
does the target region show a measurable omics difference?
is the signal local, interpretable, and potentially trackable?
can we re-measure the same network region after intervention?

This is important because longevity science is full of interventions that look promising in general but fail to translate consistently across individuals.

The reason may not be that the intervention has no biological effect.

It may be that the intervention is applied to the wrong state.

SEMO provides a way to make state matching more explicit.

A Practical Example

Imagine two people with similar biological age scores.

Person A has a network pattern suggesting mitochondrial-adaptation stress.

Person B has a network pattern suggesting inflammation-resolution imbalance.

A generic longevity report might recommend similar “anti-aging” supplements to both.

A SEMO-style algorithm would instead ask:

Which compound target networks align with Person A’s mitochondrial-related mismatch?
Which compound target networks align with Person B’s inflammatory-resolution mismatch?
Are these differences visible in the individual omics layer?
Can future measurements test whether the predicted network state changed?

This is not clinical treatment advice.

It is a computational hypothesis-generation process.

But that is exactly what longevity technology needs at this stage: better hypotheses, better measurement loops, and better ways to connect interventions with individual biological states.

What SEMO Does Not Claim

It is important to be clear about the boundary.

SEMO is not a validated clinical decision system.

It does not prove that a compound is effective for a specific person.

It does not replace clinical trials, safety assessment, medical supervision, or regulatory evaluation.

It does not mean that network association equals therapeutic benefit.

Instead, SEMO should be understood as an algorithmic framework for organizing intervention hypotheses.

Its value is not that it gives a final answer.

Its value is that it makes the question more computable:

Given this individual’s molecular network state, which compound-linked network hypotheses deserve attention, testing, and longitudinal follow-up?

That is already a major step beyond generic supplement logic.

The Potential Contribution to Longevity Science

SEMO could contribute to longevity technology in at least five ways.

1. More individualized intervention hypotheses

It can help move from population-average recommendations to individual network-state matching.

2. Better prioritization of compounds

Instead of ranking compounds only by literature popularity or general mechanism, SEMO can prioritize candidates by their relationship to a person’s omics-mapped network state.

3. Mechanistic traceability

Because the algorithm uses target networks and omics features, hypotheses can be inspected and challenged rather than hidden inside a black box.

4. Longitudinal feedback

A network gap can potentially be re-measured after intervention, allowing the hypothesis to be updated.

5. Integration with steerable biomedical AI

SEMO can provide candidate intervention hypotheses for broader steerable world-model systems, such as the SEWO framework introduced at Steerable World.

Why DeepoMe’s Approach Is Worth Watching

DeepoMe has been developing computational approaches around DNA methylation, aging, capability measurement, and network-based intervention reasoning.

SEMO fits naturally into that direction.

If DNA methylation and other omics layers provide a way to observe durable biological state, and SEWO provides a framework for steerable biomedical world models, then SEMO helps answer a practical intermediate question:

Which compound-linked network interventions might be worth testing for this state?

That makes SEMO less like a conventional supplement recommender and more like a hypothesis engine for precision longevity.

Final Thought: The Future Is Not “What Should I Take?”

The future of longevity technology should not be reduced to the question:

What should I take?

A more mature question is:

What is my current biological network state, what mismatch is most actionable, which intervention could plausibly move it, and how will we know whether it worked?

SEMO is interesting because it tries to make that question computational.

It does not promise a shortcut to immortality.

It does not turn longevity into a one-click recommendation system.

It does not eliminate the need for validation.

But it may help build the algorithmic foundation for a more rigorous form of personalized longevity science:

network-aware
omics-informed
hypothesis-driven
longitudinally testable
compatible with steerable biomedical AI

That is the potential contribution of SEMO.

Not just recommending interventions.

Helping longevity technology learn where to steer next.

Suggested hashtags

#Longevity #Bioinformatics #BiomedicalAI #NetworkMedicine #PrecisionHealth #SEMO #SEWO #AI

DEV Community