Siftlink runs AI-accelerated programs to discover novel bioactive applications and compositions towards performance, quality of life and preventing disease across life stages, conditions and needs.
One trial met its primary endpoint. One did not. Our model has seen neither result. It picks the one that worked 89 times out of 100.
correct calls on the outcome of clinical trials, based on temporal benchmarking
Correct calls, per 100 head-to-head pairs
ROC-AUC 0.893, 95% CI 0.838 to 0.939. Temporal hold-out: trained on trials reading out before 2020, tested on 2020 and later. Comparison: HINT, Cell Patterns 2022, 0.645.
Our Virtual Cell model connects molecular and clinical data to identify associations between bioactives and clinical endpoints and biomarkers. Point the model at a clinical trial that is already recruiting, or one still on paper, and it predicts whether that trial will meet its primary endpoint, before any result exists. It forecasts only from evidence that was available when the trial began; nothing published afterwards is visible to it.
Model can prioritize clinical endpoints and biomarkers for known bioactives, novel ones or even new structures or bioactives for a given endpoint/biomarker.
The evidenceBotanical extracts are hard to protect, so the clinical evidence built around a species ends up shared with everyone selling it. Differentiation depends on holding something exclusive, and internal discovery pipelines are thin. Lets work together in developing clinically supported bioactive applications.
Category leadership needs an active nobody else can put on a label. Buying a branded ingredient means competing with everyone else who bought it. Join our Precision Innovation sprint to design products and portfolios in 6 weeks.
Innosuisse Innovation Booster 2023, won for resource efficiency
Biopôle, Lausanne
We build products and portfolios.
We explore innovation from single bioactives, to synergistic compositions and to product portfolio lines to address the needs of targeted market segments through well designed proprietary positions.
We run our own discovery programs, while we are looking for partnerships to co-develop towards the market.
This category makes claims faster than it can substantiate them, and the gap is not evenly distributed. We measured it.
Across 216,350 consumer reports and 13,875 clinical trials mapped onto a shared vocabulary, consumer demand and clinical research point in different directions. Sleep is 39% of consumer discussion and 3% of trials. Skin, hair and ageing, 21% against 5%. Cardiovascular endpoints take 10% of trials and 2% of discussion.
Demand concentrated where evidence is thinnest is where a category leader can still be made. It is also where most entrants fail, because they launch the claim without building the evidence underneath it.
Our programs are selected against that map and built to close the gap rather than to exploit it.
Published on medRxiv, June 2026Share of 216,350 consumer comments against share of 13,875 clinical trials, by application area, on a common vocabulary. Bars scaled to a 40% axis. Sorted by gap size.
| Area | Consumer | Trials |
|---|---|---|
| Sleep and insomnia | 39.2 | 3.0 |
| Skin, hair and ageing | 21.3 | 4.8 |
| Anxiety and stress | 8.2 | 5.9 |
| Cardiovascular | 1.6 | 10.1 |
| Pain | 1.7 | 6.3 |
| Mood | 1.5 | 5.6 |
Metabolic health is the most contested space in nutraceuticals and the least owned. Most entrants are commodity botanicals with no protectable position, which is why they compete on marketing rather than on mechanism.
A portfolio here spans glycaemic response, satiety, lipid handling and lean mass on one claim architecture, so the evidence built for one composition supports the others and the filings overlap rather than standing alone.
Rapid weight loss takes lean tissue with it, and every kilogram of muscle lost lowers resting energy expenditure roughly three times as much as a kilogram of fat. The people who most need to protect lean mass are also eating least, which makes the standard answer, more protein, the one thing they cannot do. Protein targets of over a gram per kilogram of body weight are not reachable on a suppressed appetite.
That makes it a density and signalling problem rather than a volume problem, and the current commercial answer is commodity: protein, creatine, a beta-hydroxy metabolite, none of it protectable and all of it sold by everyone. A composition designed to work at low intake volume is both the unmet need and the ownable position.
Discovered through computational screening against metabolic targets, then assessed structurally and dynamically in house before anything was disclosed. The filing covers a compound class and its analogue space rather than a single molecule, which is the difference between an asset and a lucky hit.
Confirmation at the receptor level is the next step and the natural point for a partner to join. The compound enters wet-lab work already filtered, so the validation program is narrow rather than exploratory.
Ageing is not one endpoint. That is precisely why single-active products underperform in it, and why composition work matters more here than anywhere else in the category.
A portfolio spans cognitive, skin and metabolic ageing on a shared evidence base.
The ageing field has arrived at the same conclusion by a different route: single molecules do not move a multi-pathway process, and the interest has shifted to multi-nutrient formulations. What is missing is a rigorous way to choose the combination. Most stacks are assembled from whatever has a mechanism paper attached.
The cellular senescence and chronic low-grade inflammation axis is the clearest opening. The candidate actives are permitted for use and widely available, the mechanistic literature is substantial, and human evidence for the combinations is almost absent, which means the composition space is largely unclaimed. Inflammatory readouts also sit inside our qualified endpoint set, so a study can be designed against endpoints we know the model predicts reliably.
Silk is a protein material with three distinct streams, and the bioactivity in each is better characterised in the literature than it is commercialised. One fraction already holds recognised safety status in the United States. The regulatory position of the others is mapped.
Our interest is in defined fractions with specified molecular-weight ranges and standardised activity, and in their use as carriers for other actives. Enhancers are the least worked axis in this field: the same active at the same dose behaves differently depending on delivery, and delivery is patentable.
Discovering an unknown natural compound creates a decade of work before anyone can sell it. Safety studies, preclinical, human tolerability, and a novel food or new dietary ingredient authorisation. Pharma absorbs that cost because pharma can optimise a hit into a lead through chemistry. In nutraceuticals the ingredient is fixed, so the molecule is the one variable that cannot be improved.
Everything else is open. Which actives are combined, at what ratios, in what delivery system, for which indication. That space is larger than any compound library, and every point in it is built from ingredients that are already safe and already legal to sell.
How the platform worksTwo of our programs start from streams that industry currently discards. Cereal bran is milled and sent to animal feed. In silk processing, the protein fraction removed before spinning is a waste stream.
This is not a sustainability position added to the science. It is where the material comes from. An abundant byproduct substrate also means supply is not the constraint on scale, which is the failure point for most novel natural ingredients.
Selected by Innosuisse in 2023 for resource efficiency.
A controlled fermentation generates vitamin B12 in the substrate itself while reducing phytate, the antinutrient that binds iron and zinc in cereal matrices. B12 is the hardest micronutrient gap in a plant-based diet and is normally closed with a synthetic additive. Here it is produced in the food, by the fermentation, from a byproduct.
The protocol was designed computationally, then run at the bench. Vitamin B12 synthesised de novo to 151.43 µg per 100 mL of ferment, roughly thirty-five times the adult daily reference intake per millilitre. Phytic acid reduced from 30.5 to 18.3 mg per gram of dry weight, measured across sixteen fermentation vessels.
Two applications, one composition. Plant-based diets need a B12 source that is not a synthetic additive. Separately, anyone eating substantially less needs more micronutrient density per gram, and reduced phytate improves absorption of the iron and zinc already in the matrix rather than adding more of either.
Open for licence.
Metabolic health
Patent filed
Healthy ageing
In silico
Sustainable ingredients
Lab validation
Our work focuses on building a trustworthy AI model that captures complex molecular interactions, clinical outcomes, consumer insights and business information in a single framework. We have seen that putting the layers together benefits the overall quality of model. This approach allows us to ask key questions related to bioactive innovation, that requires cross-layer information flow. Here we provide some examples.
Point the model at a trial that is recruiting, or one still on paper, and it predicts whether that trial will hit its primary endpoint. Screen for bioactives expected to hit clinical biomarkers and vice-versa.
Given one trial that succeeded and one that failed, it picks which was which 89 times out of 100. Chance is 50.
Success ranked above failure, per 100 pairs
Inductive heterogeneous graph neural network, positive-unlabelled training, temporal holdout. 10,573 labelled pairs, 7,210 bioactives, 17,080 indications, 356 biomarkers. ROC-AUC 0.893, 95% CI 0.838 to 0.939.
Millions of people describe in public what a supplement did for them. Nobody had established whether that tracks what trials find, so nobody could use it. We read 216,350 comments, worked out which described a real benefit, and matched them against the trial record.
A benefit reported by a hundred or more people has a 38% chance of randomised trial support, against 9% across the corpus. One person reporting it tells you nothing.
Randomised trial support, per 100 claims
Natural language inference directionality pipeline on a MeSH and MedDRA aligned vocabulary. 86 forums, 329 ingredients, 13,875 trials. Odds ratio 7.14 at 100 endorsements, 81% concordance across 711 claim pairs.
Fermentation recipes are normally settled by trial and error over months. We had the model design one instead, for two jobs at once: make vitamin B12 inside wheat bran, and break down the compound that blocks iron and zinc absorption. Then we ran it at the bench.
It did both. It also predicted which variable would control the yield and which would not, and both predictions held.
Two objectives, one designed protocol
Generative design over roughly 5,000 bacterial genomes and 128,000 publications. 16 fermentation vessels. Precursor supply p = 0.0004; final cell density p = 0.974.
The inference approach the platform is built on was entered into CAFA, an open assessment where independent organisers withhold the answers, collect every entrant's predictions, then wait for real experimental results and score everyone against them.
On human targets in the Biological Process ontology we ranked 1st, out of 54 methods from 23 research groups.
Human targets, Biological Process ontology
Ranked 1st
Probabilistic graph inference over a biological network. 866 targets across 11 species. Radivojac et al., Nature Methods 2013, Supplementary Fig. 7H.
01
A nutraceutical trial costs six figures and takes a year. The record that would tell you which claims to pursue is thin, fragmented across registries and literature, and biased: successful trials post detailed results while failed ones frequently post nothing. Standard machine learning treats an untested pair as a failed one, which poisons the signal.
We built an inductive heterogeneous graph neural network over a biomedical knowledge graph linking natural products, proteins, pathways, indications and clinical biomarkers, trained under a positive-unlabelled framework so that untested is not scored as failed. Being inductive, it scores a molecule never tested in any trial from its structure alone.
Scored on confirmed clinical successes against confirmed clinical failures, with bootstrap confidence intervals on every metric across 1,000 resamples.
ROC-AUC, axis from 0.5 to 1.0
95% CI 0.838 to 0.939
95% CI 0.725 to 0.744
no interval
Forecasting the readout of a registered trial, trained only on evidence published before 2020 and tested on trials that ran after it: ROC-AUC 0.893, 95% CI 0.838 to 0.939.
From molecular structure and the knowledge graph alone, with no trial information: 0.734, 95% CI 0.725 to 0.744. Published methods predicting drug trial outcomes report figures in the region of 0.6 to 0.7, on compounds with a far richer clinical record than anything in this category. Corpora and label definitions differ between that work and this one, so the comparison sets a scale, not a like-for-like ranking.
Given a clinical biomarker, the model ranks 7,210 candidate bioactives. The top 50 contains a genuine modulator 62% of the time. The top 10 contains one 38% of the time. Selecting 50 at random from that space would find one in under 1% of cases.
Run the other way, given a bioactive the model ranks 356 clinical biomarkers, and the top 50 contains a real one 58% of the time.
Panel A. Given a biomarker, ranking 7,210 bioactives
chance 0.07%
chance 0.14%
chance 0.28%
chance 0.69%
Panel B. Given a bioactive, ranking 356 biomarkers
chance 1.4%
chance 2.8%
chance 5.6%
chance 14%
Hit rate at k, the share of queries where the top k contains a genuine pair. The dark tick on each bar is the chance rate for that pool size.
This is the question that decides whether a trial can succeed. A composition with real activity still fails if the endpoint chosen cannot move far enough to reach significance in the population studied. Endpoint selection is where we are strongest and where this category most often loses.
Performance is not uniform across a space of 17,000 indications and should not be expected to be. Some endpoints have dense mechanistic coverage and a real clinical record behind them. Others have neither.
We qualify endpoints before a program starts rather than after. Every indication and biomarker in the graph carries a measured reliability score computed against ground truth, and the first thing we tell a partner is whether theirs sits inside the qualified set. 1,350 indications and biomarkers currently qualify as high quality.
Table B. Examples from the qualified set
Reliability is ROC-AUC against ground truth for that endpoint specifically. n is labelled pairs.
Top ranking indications
Top ranking biomarkers
02
Millions of people report what a supplement did for them, in public, continuously. Nobody had established whether that signal relates to clinical efficacy, which means nobody could use it. This category makes claims faster than trials can support them, and the real-world evidence infrastructure that exists for pharmaceuticals does not exist here.
We built the first systematic framework linking community-reported benefits to the clinical trial record on a shared, MeSH and MedDRA aligned vocabulary. 216,350 distinct comments from 86 forums over four years, 329 ingredients, 581 benefit terms, joined against 13,875 clinical trials and 12,859 publications. Directionality on every comment assigned by a natural language inference model.
Association between distinct endorsements and demonstrated efficacy, 711 claim pairs with sufficient directional volume. The hollow marker at one endorsement is not significant. Share with RCT support is plotted at the two thresholds measured, so the dashed segment joins two points rather than describing a fitted curve.
| Endorsements | Odds ratio | Share with RCT support |
|---|---|---|
| at least 1 | 1.02, not significant | approx. 9% |
| at least 20 | 3.06 | not reported here |
| at least 100 | 7.14 | 38% |
| at least 300 | 9.8 | not reported here |
Consensus, not attention, is the informative quantity. A single endorsement carries no clinical signal. Above twenty distinct endorsements the association with demonstrated efficacy is significant and rises monotonically: odds ratio 3.06 at twenty, 7.14 at a hundred, 9.8 at three hundred.
A benefit endorsed by a hundred or more users has a 38% prior probability of randomised trial support, against roughly 9% across the corpus. Across 711 claim pairs with enough directional volume to assess, community experience and clinical outcome agree in 81% of cases.
Two corrections were required to see any of this, and both generalise. Only about 15% of registered trials post structured results, and posting skews toward industry sponsors, so heavily studied ingredients were initially mis-scored as unsupported. Crediting the published literature corrected it. Separately, broad condition crosswalks generated spurious matches, so a study is now counted only for the indication its abstract shows it tested.
Published on medRxiv, June 2026, doi 10.64898/2026.06.26.2635669003
Fermentation protocols are developed empirically. Strain pairing, inoculation order, precursor supply and temperature staging are usually settled by iteration, one variable at a time, over months.
We generated one computationally instead. A generative model over approximately 5,000 bacterial genomes, their biosynthetic pathways and roughly 128,000 publications produced an end-to-end multi-strain fermentation protocol for wheat bran, designed for two objectives at once: synthesise vitamin B12 in the substrate from scratch, and hydrolyse the phytate that blocks iron and zinc absorption in cereal matrices.
Then it was run at the bench.
151.43 µg per 100 mL
Vitamin B12 synthesised de novo, to 151.43 µg per 100 mL of ferment. That is roughly thirty-five times the adult daily reference intake per millilitre, which leaves room for substantial dilution in a finished food while still delivering a meaningful dose at low inclusion. Approximately 17 µg per gram of dry weight.
60% phytic acid reduction
Phytic acid reduced from 30.5 ± 1.2 to 18.3 ± 0.9 mg per gram of dry weight, a reduction of 60 percent, measured by enzymatic assay in duplicate. Enough to bring the phytate to iron molar ratio below roughly 6 to 1, the threshold under which iron absorption in humans is no longer significantly inhibited.
40% phytic acid reduction
Phytic acid reduced from 30.5 ± 1.2 to 18.3 ± 0.9 mg per gram of dry weight, a reduction of 40 percent, measured by enzymatic assay in duplicate. Enough to bring the phytate to iron molar ratio below roughly 6 to 1, the threshold under which iron absorption in humans is no longer significantly inhibited.
30.5 → 18.3 mg per g dry weight
Phytic acid reduced from 30.5 ± 1.2 to 18.3 ± 0.9 mg per gram of dry weight, measured by enzymatic assay in duplicate. Enough to bring the phytate to iron molar ratio below roughly 6 to 1, the threshold under which iron absorption in humans is no longer significantly inhibited.
Panel 1. Vitamin B12 by inoculation scheme, µg per 100 mL, observed range
Lowest of the three schemes, consistently. This is the predicted direction confirmed.
151.43, highest single result
adult daily reference intake, 2.4 µg
Panel 2. Phytic acid by condition, mg per gram dry weight
Below the 6 to 1 phytate to iron molar ratio, the threshold under which iron absorption is no longer significantly inhibited.
Enzymatic phytate assay, duplicate determinations. B12 across 16 fermentation vessels. Panel 1 bars are observed ranges, not means. Panel 1 axis runs to 160 µg per 100 mL. Schemes are labelled by inoculation order only.
| Simultaneous co-inoculation, B12 | 86.44 to 93.18 µg/100 mL |
| Sequential, B12 | 76.90 to 95.80 µg/100 mL |
| Inverted sequential, B12 | 49.43 to 61.10 µg/100 mL |
| Highest single result, B12 | 151.43 µg/100 mL |
| Starting wheat bran, phytic acid | 30.5 ± 1.2 mg/g |
| Single-organism control, phytic acid | 28.8 ± 1.4 mg/g |
| Designed co-fermentation, phytic acid | 18.3 ± 0.9 mg/g |
Getting a fermentation to produce B12 is not itself a modelling result. Predicting which variable governs the yield is.
The design held that output would be driven by precursor supply and the staged temperature regime rather than by how much bacteria you finish with. Both halves were tested. Precursor supplementation produced a highly significant increase in yield (two-way ANOVA, p = 0.0004, n = 16). Final cell density showed no significant relationship to B12 content at all (one-way ANOVA, p = 0.974).
The design also held that inoculation order would matter, and in a specific direction, because pre-colonisation by one organism depletes what the other needs. Inverted order gave the lowest yields of the three schemes tested, consistently.
Two mechanistic predictions, both confirmed. That is the result worth reporting.
Composition, strains, precursor set and thermal regime are covered by a patent filing and available under NDA.
04
The class of probabilistic graph inference the platform runs on was entered into CAFA, the first Critical Assessment of Protein Function Annotation. Independent organisers, 54 methods from 23 research groups, 866 targets across 11 species, answers withheld from entrants and scored eleven months later against experimental results no entrant could see.
On human targets, in the Biological Process ontology, it ranked first.
Biological Process is the harder of the two ontologies and the one the assessment reports as furthest from solved, because abstract cellular function is not recoverable from sequence similarity, which is what most entrants relied on. It is the category where inference over a biological network has to do the work.
The method predicted only on eukaryotic targets. The assessment's headline metric computed recall across all 866 targets regardless of whether a method had made a prediction, so the prokaryotic targets it deliberately did not touch counted against it. The organisers note this directly. Restricted to the targets it addressed, it placed third in Molecular Function.
Coverage, share of targets predicted on
Method: Kourmpetis et al., PLoS ONE 5:e9293 (2010). Assessment: Radivojac et al., Nature Methods 10:221–227 (2013), co-authored. Supplementary Fig. 7H for the human ranking, Fig. 6B for coverage.
The problem was protein function rather than clinical translation. The principle is the same: inference over a biological network, with a calibrated confidence attached to every prediction.
If something in them does not hold up, we would rather hear it from you than not hear it.
Science to business in a single AI model.
In most organisations molecular science, clinical evidence and intellectual property sit in three systems owned by three teams, so no question can travel between them. We model all three as one graph. A single query crosses every layer and returns its own confidence.
One graph, five stages. Signal travels from molecules through the clinical record and the patent corpus to a product and the people it suits.
Efficacy in nutrition is a network property. Single-target, single-molecule thinking is a category error here: a pathway rarely moves far enough from one point of intervention to produce a readout in a mildly affected population. Compositions are designed for pathway coverage and complementary mechanism, not for the potency of one active.
The same actives at the same dose behave differently depending on delivery. We develop enhancer systems, including silk-based carriers, where the bioavailability of a known active is the variable rather than the active itself.
An ingredient with an established safety record and an existing supply chain, matched to an indication nobody has connected it to. The fastest route from discovery to market in this category, because everything except the claim already exists.
They do occasionally, and we study them properly. They are not the strategy.
An extended knowledge model covering natural compounds, protein targets, pathways, indications, clinical biomarkers and clinical outcomes, held as relationships rather than as separate tables. Because biomarkers sit alongside outcomes, a composition can be traced from the molecules acting on a target through to the readouts a trial would measure.
Probabilistic inference over a heterogeneous graph. Every prediction carries a calibrated uncertainty range and an audit trail down to the molecules, proteins, trials and filings that produced it. The same inference runs at the level of a consumer profile, so a composition can be matched to the people most likely to respond rather than to an average.
The global patent corpus modelled in the same graph as the science, so filings connect to the molecules and claims they cover rather than sitting in a separate search tool. Novelty, whitespace and freedom to operate are computed against a composition as it is designed, not assessed after the fact. Claim scope is shaped while the formulation is still open to change.
The platform was tested with its evidence layers separated, then integrated. Any single layer on its own performs near chance. Integrated, the same data reaches far above it.
That is the case for the architecture stated as a measurement. Molecular, clinical and patent evidence held in separate systems cannot answer the questions that cross them. Held as one graph, they can.
full protocolWe take an asset from computational discovery through composition, filing and study design. Getting it to market needs a partner with supply chain, clinical budget and channel. That is the division of labour and we are explicit about it before anyone signs anything.
A patent-filed Siftlink composition, exclusive in a defined field of use. You receive the composition, the filing position, the evidence dossier, the study design and the claim language.
Join a program at its current stage and carry it forward together. IP held jointly. You bring brand, channel and clinical budget. We bring the platform, the science and the patent function.
Best for a committed line, or for a supplier building a branded ingredient.
We take on a small number of innovation partners each year, in territories we are already building in. You set the target and fund the program. We run the discovery, design the composition and build the filing position, which we hold or hold jointly with you.
This is not contract research. We work in areas where we intend to build a portfolio regardless, which is why the fit has to be right in both directions.
Best where you have a category position to defend and no internal discovery pipeline.
Partner
Siftlink
Siftlink runs AI-discovery programs to identify novel, bioactive applications and compositions towards commercial development through partnerships.
Founder and CEO
Entrepreneur and AI x Bio x IP researcher. PhD in predictive modelling for protein and ingredient function. Two decades across R&D and patents in multinational food and nutrition.
Chairman
Former Global Head of Patents at multiple nutrition and biotech groups. Qualified European and Dutch patent attorney, PhD in Molecular Biology. Thirty years in biotech, pharma and nutrition.
Innovation Strategy
Leads partner engagement and commercial growth. PhD Plant Physiology. Twenty years across agtech, operations and supply chain at scale.
Head of Translational Research
Connects translational science with AI modelling. PhD in inflammation and immunity, trained in molecular pathology at Verona, published in liquid biopsy and biomarker validation.
Head of AI-Ops and Data Engineering
Computer Scientist. Architect of Sensi our agent-AI based discovery platform and data infrastructure the platform runs on.
Switzerland's life-science campus. Admission is reviewed.
Phenyl Building, Route de la Corniche 3A, 1066 Epalinges
We do not publish client work, in any form, including anonymised. Confidentiality is a condition of the work.
Publications and technical notes.
Concordance between clinically supported and Reddit-reported ingredient benefits.
medRxiv, June 2026 · doi 10.64898/2026.06.26.26356690
Read the preprintBenchmarking report: three tasks, bootstrap confidence intervals, temporal holdout, per-endpoint reliability.
Request the reportKourmpetis et al., PLoS ONE 5:e9293, 2010
Read the paperRadivojac et al., Nature Methods 10:221–227, 2013
The organisers' assessment of CAFA, co-authored. Supplementary Fig. 7H carries the human Biological Process ranking; Fig. 6B carries the coverage figures.
Read the paperContact
One call, thirty minutes, no deck. Describe the category, the claim or the gap. We tell you whether the platform has signal in it and what the next step involves.
Book a callOr write to info@siftlink.com
We will ask for name and company in the reply.
Siftlink SA
Biopôle, Phenyl Building
Route de la Corniche 3A
1066 Epalinges, Switzerland
