---
title: A machine, not software: how Biomiv finds what the research has not yet connected | Biomiv Blog
url: https://biomiv.com/blog/a-machine-not-software
---
[All articles](https://biomiv.com/blog)

November 16, 2026 · 11 min read · Cymetica Biosciences

# A machine, not software: how Biomiv finds what the research has not yet connected

Most health AI repeats what is already known. Biomiv is built to find relationships no paper has stated yet, using vector embeddings, hidden relationships and a design borrowed from how biology remembers, selects and adapts. Here is how it works, where the idea comes from, and what a prediction from it does and does not mean.

Ask most health AI about a condition and you get a summary of what has been written about it most often. That is useful, and it is also the one thing it cannot go beyond. A system trained to produce the most likely next sentence returns the consensus, because the consensus is what is most likely.

Biomiv is built for a different job: to find relationships between foods, nutrients, compounds, genes and conditions that *no paper has stated yet*, and to show them to you labelled as what they are. This post explains how, where the idea comes from, and what a prediction from it does and does not mean.

## What is already known is not a discovery

In 1986 Don Swanson, an information scientist at the University of Chicago, read two sets of medical papers that never cited each other. One set reported that fish oil lowers blood viscosity and platelet aggregation. The other reported that patients with Raynaud's syndrome have high blood viscosity and platelet aggregation. Nobody had put the two together. Swanson did, and proposed that fish oil might help Raynaud's patients [1]. Three years later a double-blind controlled trial tested fish oil in patients with Raynaud's phenomenon and reported a benefit in those with the primary form [2].

He did it again in 1988, connecting magnesium and migraine through eleven lines of evidence that sat in separate literatures [3]. He called what he had found **undiscovered public knowledge**: facts that are each published, each true, and never joined, because no single person reads both halves [1], [4].

That was two literatures and one careful reader. PubMed now holds tens of millions of papers. The number of unjoined halves is far beyond what people can read, which is why the field that grew from Swanson's work, literature-based discovery, moved from reading to computing [5].

## Hidden relationships

We use one definition throughout Biomiv.

- A **known relationship** is explicit: two things have appeared together in the same paper.
- A **hidden relationship** is implicit: two things have *never* appeared together in any paper, yet the things written around each of them keep converging.

A search engine can only return the first kind, because it matches what is on the page. A language model trained on the same pages leans the same way. A hidden relationship has no page to match. It has to be measured.

There is published evidence that this measurement can run ahead of the literature. In 2019 a group at Berkeley trained word embeddings on 3.3 million materials science abstracts and showed that embeddings built only from papers up to a given year ranked, among their top predictions, materials that were reported as thermoelectrics in later years [6]. The knowledge was latent in the text before anyone wrote it down.

Our own group reported the same effect in biology. In work presented to the NASA GeneLab Animal Analysis Working Group in 2022, 13 genes involved in muscle atrophy under microgravity were measured against 12,903 human proteins across 206,233 PubMed papers published from 2010 to 2015. The analysis found 88 implicit interactions: pairs with no direct connection in that literature whose vectors were moving closer year by year. One example shown is PARP1 and the atrophy gene NFAT5, converging through 2010 to 2015, ahead of the later papers the presentation cites for the connection [7].

## Vector embeddings: meaning as position

The measurement rests on an old observation in linguistics: a word is characterised by the company it keeps [8], [9]. If two words keep the same company, they mean related things, even if they never meet.

A vector embedding turns that into geometry. Every object the system reads about, a gene, a compound, a food, a condition, becomes a point whose position is set by the contexts it appears in. Two points are close when their contexts are alike. Distance can then be computed between any two objects, including two that have never shared a sentence.

Cymetica's founding team built this for science before it had its current name. A provisional patent filed in June 2005 at Lawrence Berkeley National Laboratory, later granted as US 7,987,191, describes a context window slid over a corpus, relationship scores that decay with distance, one vector per object held in an associative memory, and the recovery of hidden relationships through shared associates [10]. The Laboratory described the system that year as measuring "the proximity of every word to every other word in millions of documents" to find "indirect connections … that haven't been noticed before" [11]. Word2vec, the method that made embeddings a household term in AI, was published in 2013 [12]. The full dated record is in a separate essay [20].

The model behind Biomiv, which we call the Tuatara vector model, descends from that work.

## Why we call it a machine

Software does what its instructions say. Give it the same input and it returns the same output until a programmer changes a line. That is the right design for a calculator and the wrong one for discovery, because the thing being studied, the world's research, changes every day.

Biomiv is built as a machine in the older sense of the word: a set of parts that take in material, do work on it and hand the result to the next part, continuously.

- **Intake.** A research pipeline reads new papers from global scientific research databases every day and records which foods, nutrients, compounds, genes and conditions each one names.
- **Memory.** Those readings feed the vector model. As the literature moves, positions in the space move. Nobody rewrites a rule.
- **Reasoning.** For a condition you name, the BIOMIC-42 algorithm finds the body processes behind it and the items reported to act on each one, weighs benefit against reported harm, and assembles a combination.
- **Expression.** BioAI answers in plain language and draws the cards and charts, each with its sources.
- **Correction.** Each part is measured and adjusted by another part, described below.

The practical difference: ask the same question next month and the answer may have changed, because the research did. That is the machine working, not a defect.

## What it borrows from biology

We did not arrive at this design by analogy after the fact. The 2005 system was described at the time as "mimicking the process of auto-association, which is a cognitive principle that describes how a human stores and recollects information" [11]. The parts of Biomiv map onto biological processes one by one.

| Biological process | What it does in a living system | What it does in Biomiv |
| Associative memory | Things experienced together become linked, and repetition strengthens the link [13] | Objects that share contexts are stored as associates, with a strength that grows as more papers repeat the context |
| Recall from a cue | Part of a memory brings back the whole [15], [16] | A condition you name recalls its associates from the memory, whatever words you used for it |
| Spreading activation | A recalled idea activates its neighbours, and theirs [14] | The query is widened through its strongest associates. This is the step that reaches an item no paper has linked to your condition |
| Selection | Variants compete and the fitter ones persist [18] | Candidate settings and candidate combinations are scored against the evidence, and the better ones are kept and varied |
| Learning from error | A mistake changes later behaviour | BioAI is examined, and every question it misses becomes a correction before it is examined again |

None of this was fashionable in 2005. It has since become the mainstream account of how modern AI works: in 2020 the attention mechanism at the centre of today's language models was shown to be the update rule of a modern associative memory network [17].

The last two rows are what let the machine improve its own ability instead of waiting for a release. Where a setting can be tuned, such as how wide a context window is, how fast a score decays with distance or where a threshold sits, our optimization follows evolution rather than a single hand-picked value: a population of candidates, a score, selection, variation, repeat [18]. The test we hold a hidden-relationship method to is the one in the GeneLab work: build from the literature up to a date, then check the predictions against what was published after it [7].

BioAI itself is taught the way a student is. It follows a curriculum of courses on what Biomiv shows and how to explain it honestly, and a tutor examines it on each course at a fixed pass mark. A missed question is corrected and the examination is run again.

## What is ours

The papers are not ours. They belong to their authors and publishers, and Biomiv links to them. What is proprietary is everything built on top of reading them:

- **The data we derive.** A daily record of which items each new paper names, the history of how that attention moves, and the relationship vectors computed from it. That history cannot be bought or reconstructed quickly, because it accumulates one day at a time.
- **The algorithms.** The Tuatara vector model, and BIOMIC-42, which turns a condition into body processes and an engineered combination with a score from 0 to 100.
- **The optimizations and tuning.** The model is tuned for its domain. A space built for genes, compounds and foods is not the space built for another field, and the windows, weights and thresholds differ between them.

Alongside our own pipeline, Biomiv uses public and curated reference sources and names them on the page: PubMed for papers, ChEMBL for approved indications, the Comparative Toxicogenomics Database and DrugBank for curated records, ClinicalTrials.gov for trials and the Protein Data Bank for molecular structures [19].

## What you see on Biomiv

When you ask BioAI about a condition on [Discoveries](https://biomiv.com/discoveries), each part of the machine shows up as something on the page.

- **The engineered diet and its score** come from BIOMIC-42: the body processes behind the condition, the foods, plants and nutrients reported to act on each, and a combination scored after subtracting reported harm.
- **Related items** come from the names in your message and from the vector model. Some of them will not appear in any paper about your condition. That is the point.
- **"Similar compounds by Tuatara vector embedding"**, on the Repurposed drug compounds card, are compounds the model places closest to drugs with evidence. They are predictions.
- **"Already on record"** shows the opposite: what curated databases already hold. Known and predicted are kept in separate places so you can tell them apart.
- **The charts and paper counts** show where research attention is moving. They measure attention, not whether something works.

## What a prediction is not

A hidden relationship is a hypothesis. Swanson's fish oil idea became knowledge only when someone ran a trial [2]. Most hypotheses, from people or machines, do not survive that step. An item the model places near your condition may never have been tested for it in any study, and may not work.

So Biomiv labels its output. A model prediction is called a prediction. A count of papers is called a count. A score describes a combination against the research on record and says nothing about you. Drugs, genes and proteins are shown as research context, not as something to take.

Everything on Biomiv is research information, not medical advice. Talk to a doctor or pharmacist before changing a medicine, a supplement or your diet.

What the machine offers is a better list of questions: the connections worth testing that nobody has written down yet. Discoveries have never come from repeating what everyone already knows.

## References

- Swanson DR. Fish oil, Raynaud's syndrome, and undiscovered public knowledge. *Perspectives in Biology and Medicine*. 1986;30(1):7-18. [doi:10.1353/pbm.1986.0087](https://doi.org/10.1353/pbm.1986.0087). [PMID 3797213](https://pubmed.ncbi.nlm.nih.gov/3797213/).
- DiGiacomo RA, Kremer JM, Shah DM. Fish-oil dietary supplementation in patients with Raynaud's phenomenon: a double-blind, controlled, prospective study. *American Journal of Medicine*. 1989;86(2):158-164. [doi:10.1016/0002-9343(89)90261-1](https://doi.org/10.1016/0002-9343(89)90261-1). [PMID 2536517](https://pubmed.ncbi.nlm.nih.gov/2536517/).
- Swanson DR. Migraine and magnesium: eleven neglected connections. *Perspectives in Biology and Medicine*. 1988;31(4):526-557. [doi:10.1353/pbm.1988.0009](https://doi.org/10.1353/pbm.1988.0009). [PMID 3075738](https://pubmed.ncbi.nlm.nih.gov/3075738/).
- Smalheiser NR. Rediscovering Don Swanson: the past, present and future of literature-based discovery. *Journal of Data and Information Science*. 2017;2(4):43-64. [doi:10.1515/jdis-2017-0019](https://doi.org/10.1515/jdis-2017-0019). [PMID 29355246](https://pubmed.ncbi.nlm.nih.gov/29355246/).
- Henry S, McInnes BT. Literature based discovery: models, methods, and trends. *Journal of Biomedical Informatics*. 2017;74:20-32. [doi:10.1016/j.jbi.2017.08.011](https://doi.org/10.1016/j.jbi.2017.08.011). [PMID 28838802](https://pubmed.ncbi.nlm.nih.gov/28838802/).
- Tshitoyan V, Dagdelen J, Weston L, Dunn A, et al. Unsupervised word embeddings capture latent knowledge from materials science literature. *Nature*. 2019;571(7763):95-98. [doi:10.1038/s41586-019-1335-8](https://doi.org/10.1038/s41586-019-1335-8). [PMID 31270483](https://pubmed.ncbi.nlm.nih.gov/31270483/).
- Vector Space Biosciences. Using scientific data engineering pipelines to accelerate novel discoveries in space biosciences. Presentation to the NASA GeneLab Animal Analysis Working Group, 27 July 2022. [PDF](https://vectorspacebio.science/nasa/Vector-Space-Biosciences-NASA-GeneLab-AWG-presentation-v12.pdf) (the atrophy gene analysis is on pages 23 and 24, the definition of explicit and hidden relationships on page 27).
- Harris ZS. Distributional structure. *Word*. 1954;10(2-3):146-162. [doi:10.1080/00437956.1954.11659520](https://doi.org/10.1080/00437956.1954.11659520).
- Firth JR. A synopsis of linguistic theory, 1930-1955. In: *Studies in Linguistic Analysis*. Oxford: Blackwell; 1957:1-32.
- US Patent 7,987,191 B2. System and method for generating a relationship network. Priority 6 June 2005, granted 26 July 2011. Assignee: The Regents of the University of California. [Google Patents](https://patents.google.com/patent/US7987191B2/en).
- Lawrence Berkeley National Laboratory. A search engine that thinks, almost. Science@Berkeley Lab, 31 March 2005 (archived on the Internet Archive).
- Mikolov T, Chen K, Corrado G, Dean J. Efficient estimation of word representations in vector space. 2013. [arXiv:1301.3781](https://arxiv.org/abs/1301.3781).
- Hebb DO. *The Organization of Behavior: A Neuropsychological Theory*. New York: Wiley; 1949.
- Collins AM, Loftus EF. A spreading-activation theory of semantic processing. *Psychological Review*. 1975;82(6):407-428. [doi:10.1037/0033-295X.82.6.407](https://doi.org/10.1037/0033-295X.82.6.407).
- Hopfield JJ. Neural networks and physical systems with emergent collective computational abilities. *Proceedings of the National Academy of Sciences*. 1982;79(8):2554-2558. [doi:10.1073/pnas.79.8.2554](https://doi.org/10.1073/pnas.79.8.2554). [PMID 6953413](https://pubmed.ncbi.nlm.nih.gov/6953413/).
- Kanerva P. *Sparse Distributed Memory*. Cambridge, MA: MIT Press; 1988.
- Ramsauer H, Schäfl B, Lehner J, et al. Hopfield networks is all you need. 2020. [arXiv:2008.02217](https://arxiv.org/abs/2008.02217).
- Holland JH. *Adaptation in Natural and Artificial Systems*. Ann Arbor: University of Michigan Press; 1975.
- Data sources named on Biomiv: [PubMed](https://pubmed.ncbi.nlm.nih.gov/) (U.S. National Library of Medicine); [ChEMBL](https://www.ebi.ac.uk/chembl/) (EMBL-EBI); the [Comparative Toxicogenomics Database](https://ctdbase.org/); [DrugBank](https://go.drugbank.com/); [ClinicalTrials.gov](https://clinicaltrials.gov/); the [RCSB Protein Data Bank](https://www.rcsb.org/).
- Cymetica. Eight years before word2vec: the relationship vectors Berkeley Lab patented in 2005. [cymetica.com/blog](https://cymetica.com/blog/relationship-vectors-before-word2vec).

**More from the blog**

- [Discoveries are hidden relationships: what an enzyme map for a rare disease shows](https://biomiv.com/blog/discoveries-are-hidden-relationships)

[All articles](https://biomiv.com/blog) · [Ask BioAI about a condition](https://biomiv.com/discoveries)
