For Medical Professionals

The numbers of PubMed studies on the human microbiome have exploded in the last few years as shown in the chart below. In the early 1990’s we had around 10 studies a year. In 2017 we reached 4421 studies — a massive increase and expect at least 4000 every year going forward.

This creates a challenge for medical professionals. Most have never received anything beyond the simplest introduction. Even recent graduates are unlikely to receive information. 74% of the papers have been published in the last 5 years – their contents were likely not incorporated into any medical curriculum.

I am by training and job experience, a statistician, a knowledge engineer, a data scientist and an artificial intelligence practitioner. Early in my careers, I taught high school general science, chemistry and physics at colleges. I have become very interested in this area due to family health challenges.

Microbiome having a signature for many conditions

There are many conditions that have found statistically significant bacteria patterns, I have extracted their data (with links to studies) here.

Microbiomes can predict some conditions accurately

This area is under researched — likely because getting good data scientists taking a major pay cut doing medical data is hard with current business offers. Typical statistics taught in medical school are insufficient.

Microbiome bacteria produce signal chemicals/metabolites

See this list of TED Talks for background. In short, the abundance or shortage of the appropriate mixtures impact cells. The same chemical may be produced by many different bacteria; this means that the suspect list is often larger than expected.

Hypothesis: Microbiome determine Symptoms

My main interest has been myalgic encephalomyelitis – a condition infamous over 100 symptoms with very few in common between patients. From over 750 donated microbiome samples, some 400 samples were annotated with their symptoms. I tried a variety of statistical approaches and got no where. I then switched to some non-parametric approaches and in May 2019, I shouted Eureka! (but not running thru the streets naked). From these 400 annotated microbiomes, I ended up with 1000’s of statistically significant relationships. For example for the vague “brain fog” we have a few listed below. Some of the relationships were found at the species level.

Why your skill sets need to be complemented

Time for Inference

If a microbiome shift of a specific group of bacteria results in a symptom, then we should test the hypothesis that normalizing this group of bacteria should cause the symptom to reduce or disappear.

No Time for Probiotics!

The urban myth that all that you need is to take probiotics (any type) will fix matters. TOTALLY BOGUS – show me the PubMed studies!!!

Use Antibiotics with caution (if at all)

Experimental protocols for myalgic encephalomyelitis found that certain combinations of long term antibiotics could result in remission for a significant percentage, other antibiotics would make things much worse. We know that antibiotics causes significant changes in some bacteria.

Problem #1:Time to read recent studies

Recent studies, especially veterinary studies, use 16s analysis of the whole microbiome to see what the impacts of various foods, minerals, supplements and spices have on the microbiome. We should create a spread sheet of what food impacts which bacteria and then figure out the optimal combination…. At this point reality hits: we are dealing with 2000+ bacteria taxonomies and 2000+ modifiers… so we may need a spreadsheet with 2000 x 2000 cells. Looking deeper into the studies, we find that different studies on the same combination produced results that disagree, so we may have multiple items in each of these 4,000,000 cells.

Problem #2: What needs to be corrected?

This is actually where most of the current studies fall down badly. Typically the average and standard deviation are computed and no one looks below the cover to see if we are dealing with a bell-curve. The data is rarely ‘normal’. Let us look at values for two taxonomy from the 750 donated microbiome:

Almost 75% of the population has below AVERAGE levels…. think about it
Over 75% of the population has below AVERAGE levels

The Problem

We have too many moving parts, too many bacteria, too many modifiers, disagreement between studies, studies results often vague using non quantified expressions “Increased” (by 1% or 500%?)

The Solution

All of the above would be overwhelming to most people. Fortunately, I happen to have the technical quantitative skills and the emotional motivation to tackle this issue – often spending 40+ hr/week working on the solution. The solution is evolving as more data comes in (both studies and samples), and I get creative (I do have multiple patents granted)

In short:

  • I have created a 58,000 record database of how various items modify the microbiome (each linked to the source study). The Modifiers
    • This is done by applying text data mining to studies. At present, 14.1 GB of data from PubMed is in the mine, a lot of it is waste.
  • I have created the ability for people to upload their microbiome and open source the code used. Including the basics of parametric analysis that I used.
  • I have implemented several algorithms (which are enhanced over time) to determine which bacteria should be the focus. At present the three available are:
    • Original: based on averages and standard deviations (IHMO a poor choice). This gives the biggest list with likely a lot of false positives. It is still the most common approach that I see used in studies.
    • BoxPlot: named after a data science method. This gives a much shorter list, it identifies outliers with a non-parametric method.
    • Symptom-BoxPlot: This matches the statistically significant patterns of symptoms to the microbiome and only attempts to modify those. The goal is to reduce the symptoms. This is the smallest list, and occasionally has nothing.
  • From the database we then proceed to optimize the choice of modifiers to correct the shift of the selected bacteria. At this point, we are using fuzzy logic and bayesian methods to create weights for items:
    • Clear Positive – helps with all identified bacteria and no known negative impact on these bacteria
    • Clear Negative – hurts all identified bacteria and no known positive impact on these bacteria
    • Mixed – helps some and hurts some. The weight can be deceptive because it is a weight of the probability(and not the impact). Care needs to be taken.
  • Many of the modifiers are specific species of probiotics. Getting single species or strain probiotics is often not possible for most people. As a result, users have entered the contents of probiotics mixtures which allows the same rating as above: Clear Positive/Negative and Mixed
  • Many of the studies cite flavonoids etc –Often these are greek to the common person. When flavonoids are clear positives, the US Dept of Agriculture resources are used to identify the foods that contains them, and the amount

All of the above is free and open. The donated data is made available at for others to use. The source code is being cleaned up, systematically released, and made available at

The Challenge of Using it for Studies

A typical medical study is done by having two groups, one receives something and the other a placebo. With this system, suggestions being produced are different for each person (and changes over time). One could intentionally tell someone to take the Clear Negative to see if symptoms worsen – but this is immorale (Cause no harm).

As a classically trained scientist, a key factor of any model is the ability to predict. This approach is rich in predictive ability. There have been multiple positive individual reports from users, often with the microbiome test results to validate subjective reports.

Personally, I have seen in family members, Crohn’s Disease with Fistulas improve greatly and not require any prescription for CD for 3 years, nor any more fistulas. Not a remission, but very happy to do regular testing and updates of diets and supplements. There are still some black holes with not clear bacteria associations yet (the number of people reporting conditions are too few and/or comorbid with other conditions) for example, mast cell issues.

This is a difficult path because it is ultra personalized medicine. Explicit to a person and valid for 1-3 months usually. Off label prescription drugs can come into play because they do modify bacteria (examples here).

How long before it is accepted by Main Stream Medicine?

Given that it took decades for H. Pylori to be accepted as the cause of ulcers, I do not expect it to be soon. The usual promoters for medical change, pharmaceutical companies, have no profit in this approach. If I (a seasoned software developer/architect who worked for startups), asked how much do I need to properly commercialize this (including double checking data, increasing samples and meta data), I would say between $5 million and $20 million dollars for starters.

What we have is a citizen science implementation that uses a lot of duct tape and bailing wire. Some medical professionals are using it as one of several inputs for their patients.

Some Snapshots of features:

The chart below shows interactions obtained from the 750 samples using non-parametric methods. The thickness of the line increases the strength of the interactions, the size of the oval indicates the amount of each bacteria in an individual sample (this sample had 6% being Bifidobacterium — placing it at the 79%ile in the samples) .

This next diagram is based on medical literature summarized by Peter D’Adamo. The lines are all the same thickness because the degree of impact is not known.

The next chart shows percentile levels and is far better to see if a shift is significant. Note that average is intentionally left off. Selecting what to inspect as a chart was simple, look at the box-plot outlier bacteria list.

Bottom Line

The site starts with a hypothesis and runs with it logically. There are some assumptions made in the processing, often the choice is driven by a bayesian computation to pick the most probable option. It attempts to be totally open on how it comes to conclusions (and readers have caught a couple of data entry errors because of this openness).

From the literature on risk evaluation and professionals, I believe that the suggestions generated will perform better then suggestions from most medical professionals. The system is still evolving (and with open sourcing it, being refactored and minor issues corrected).