How it works

How a prediction
becomes protection.

What if we could change an insect’s decision to land? We predict which compounds could do it. Labs test them and share what happens, so each experiment can inform the next across the network.

Participating labs share methods and data within the network so each experiment can inform the next. Only compound identities and basic repellency data on those compounds are released publicly each quarter. Public release of anything else requires separate written permission from Monarch and the contributing institutions. We’re building toward a cycle that also learns from field trials, mechanism studies, and practical use.

Model

Predict from shared evidence

From millions of possibilities to a box of compounds.

Increasingly capable models still need evidence about the living animal. The challenge is to predict how an insect will respond to a compound we have not tested, then run the experiment and find out.

The question is specific: for this compound, in this insect, at this exposure, what response should we expect? Different model architectures learn from the shared experimental evidence to help us choose what to test next.

Every participating lab receives the full shared dataset and ranked candidates, including results from other labs. Researchers can use that evidence to ask their own questions, test another explanation, or build a different model. The work of one group becomes a starting point for others.

A prediction becomes something a scientist can hold: a shipment of compounds to test.

Why start with a prediction?

There are hundreds of millions of compounds to consider. At 100 tests a day, testing 100 million compounds just once would take about 2,740 years. That is a scale illustration, not our screening rate, and it leaves out controls, repeats, and different doses.

The models help us decide where to look. The experiments tell us what happened. After the first data has been recorded and checked, the mosquito protocol calls for 20 compounds each month. Some are sent to several labs so we can measure how much the results vary between them.

Start with a question the experiment can answer.

Before fitting a model, define the response it must predict. For a mosquito, that might be time spent on the landing surface relative to a matching control. For a fly, it might be preference for the treated end of a tube. The insect, exposure, attractant, and observation period are part of the question.

The model can predict a measured value, or the probability that it crosses a stated threshold. A compound classified as active under one set of conditions is not automatically active under another. We retain the underlying measurements and fit or evaluate each task separately before asking whether learning across tasks helps.

How does a molecule become something a model can learn from?

A molecular fingerprint records which chemical fragments a structure contains. Physical descriptors record properties such as molecular size, polarity, and solubility. Learned representations offer another route: a model develops its own numerical description from molecular structure. Compound identity, including stereochemistry where relevant, must remain traceable to the material actually tested.

These descriptions become inputs; the measured insect response is the target. We compare approaches such as random forests, gradient boosting, and neural models on the same prediction task. A more complex architecture needs to improve performance on compounds it has not learned from. The amount and diversity of the evidence matter more than the name of the model.

How do we know the model is useful?

With the same experimental budget, can the model help us find more compounds whose effects reproduce? We keep compounds out of training and compare predictions with their measured responses. Holding out whole chemical families tests a harder question: how well does the model handle different chemistry? We also check similarity across the split, because different families can still share important features.

For a test on unseen compounds, all recordings and repeat measurements of a compound stay together when dividing the data. Choices about features, thresholds, and model settings use the training and validation data. Results reserved for final evaluation remain untouched. Predictions recorded before the next experiments provide a prospective test; evaluation in another lab tests a change of setting.

Each round asks five practical questions:

Does it add to what simple properties tell us?
Compare it with a model using physical descriptors alone. Additional complexity needs to buy useful predictive information.
Is it relying on a shortcut?
Check whether a single property, such as molecular weight, explains the ranking. Investigate whether that pattern reflects biology or the way the dataset was assembled.
Has the shortlist narrowed to one chemical family?
Count the distinct structural families among the leading candidates and the share occupied by the largest. A high success rate within one family leaves much of the library unexplored.
Does it beat similarity search?
Compare it with simply choosing compounds that resemble a known active. The model should earn its place against this inexpensive alternative.
Would another modeling approach help?
Compare alternatives using the same compounds, targets, and evaluation rules, with model selection kept inside the training process.

Read the scores against the decision.

For a fixed testing budget, we ask how many useful candidates appear near the top. Enrichment compares the active fraction in that shortlist with the active fraction in the same evaluation pool. A value of one means no improvement over random selection. The threshold defining active and the size of the shortlist must be stated in advance.

Rank agreement asks whether predicted order follows measured response. It does not tell us whether the predicted magnitude is right. For probability predictions, the Brier score measures squared error against the observed outcome. We also check calibration: among predictions near 70 percent, does the defined outcome occur about 70 percent of the time? Those probabilities refer to the stated target, not automatically to the chance of an individual insect landing.

Scores need uncertainty estimates, appropriate baselines, and results broken out by the task being predicted. A good average can conceal failure on an unfamiliar family or a different insect. The strongest evidence is a prediction written down before an independent experiment.

Check the measurement as carefully as the model.

Measure the same compound in independent runs. How much does its response change? Do those repeats agree on whether it crosses the threshold? This tells us how reproducible the target is and where more evidence is needed.

Repeatability estimates depend on the assay, the compounds, and the number of repeats. They help us interpret prediction errors; they are not a universal hard ceiling on performance. If a result is unstable, another model may be less useful than a better experiment.

The standard we’re working toward.

Our laboratory challenge sets a demanding target: predict the probability that an insect lands at least once during a fixed observation period, to within 10 percentage points for every compound in a predefined, held-out test set. The conditions, sampling plan, and treatment of uncertainty must be fixed in advance. Predictions must precede the outcomes, and an independent lab must reproduce the result using the same locked model and protocol. This is a goal, not performance already achieved. It requires individual landing outcomes; landing-zone occupancy alone cannot establish it.

THE FIRST SHIPMENTMOSQUITO PROTOCOL

DEETA known repellent, used as a positive control.

10 candidates, plus a positive control. An illustration of what a participating lab receives.

Learn

Use what we learn to choose the next experiment.

We train the models on the growing body of shared behavioral data. A result can support a prediction, expose an error, or show that the measurement needs work. More data is useful when it helps us make a better next decision.

The models identify new candidates. Labs test them. Everyone shares the results. A result from one lab can change what another tests next.

What do we test next?

The next batch has four jobs. The balance depends on what the last round taught us and how many experiments we can run.

Find promising compounds.
Test candidates the model ranks highly and check whether their predicted effects reproduce.
Explore unfamiliar chemistry.
Sample structural families with little or no experimental evidence.
Test a close comparison.
Include compounds that resemble known actives but are predicted to respond differently. Matched experiments can help identify which structural differences matter.
Resolve useful uncertainty.
Choose cases where a measurement could change the next decision. A score near 0.5 is not enough on its own; uncertainty must be meaningful for that model and that chemistry.

A useful batch gives us a chance to find a better compound and a chance to improve the next prediction. Testing only the highest scores would leave us with little evidence about the rest of the library.

What happens when predictions are wrong?

We need to understand what limited the prediction. Does the effect repeat? Were the experiments comparable? Have we tested enough different chemistry? Did the model miss something that matters?

The next useful step might be a repeated assay, a different set of compounds, a better description of the experiment, or a different model. Sharing the recordings and conditions gives the network a way to investigate together.

Recorded conditions do not automatically become model inputs. Run order, apparatus position, or a decision to retest a promising compound at another dose can reveal how the experiment was organized. A model can learn those shortcuts instead of the relationship we want to understand.

Before changing the assay or model, state the suspected problem, what should improve, and what result would count against that explanation. If the shortlist is too narrow, test different chemistry. If repeated measurements disagree, investigate the experiment. These are hypotheses to test, not diagnoses established by one score.

Learn from work beyond our own experiments.

Published screens and larger chemistry datasets can help us describe molecules we have barely tested. One approach is transfer learning: take a molecular representation learned from a larger dataset, then fit and evaluate a model against our measured responses.

Older experiments may use different insects, doses, exposure routes, or definitions of success. Their labels cannot simply be treated as interchangeable with ours. We preserve the source and conditions, check for overlap with evaluation compounds, and test whether the added information improves held-out predictions. Shared evidence becomes more useful when its differences remain visible.

Walk through a 1,000-compound example

An illustrative example of the process, not a report of 1,000 compounds tested or results achieved.

Start without a model.
Imagine 1,000 candidate compounds and an initial random sample of 100 for each insect. Run the assays with controls, define the targets, and reserve evidence for evaluation.
Investigate the first predictions.
Suppose the model mostly ranks by molecular size. Check the source records. A label mixing knockdown with avoidance would teach the wrong task; correcting it requires preserving what was actually measured.
Broaden what the next round can teach us.
Suppose the leading candidates now come mainly from one family. Add unfamiliar structures and close comparisons. A lower immediate hit rate can be worthwhile if it gives the following round better evidence.
Test what will improve the next prediction.
If progress slows, examine repeatability, coverage, and the model comparison. Test whether additional chemical or receptor information helps, rather than assuming more columns will make the model better.

Each round leaves a shared record: what was tested, what was predicted, what happened, and what changed. The useful outcome is both a better next experiment and a dataset another scientist can question, reproduce, and build on.

From every participating labShared measurements

Learn from the shared evidenceUpdated predictions

Choose what to testThe next experiment

↶ Back to the next prediction

Lab data

Test

Watch what the animal does.

We have to run the experiments. Our lab in Emeryville and partner labs run shared assays. Cameras record how insects respond to a compound, alongside controls. Shared methods let another scientist repeat an experiment and check the result.

This connects researchers across North America and East Africa, with different insects, expertise, and field conditions, all working on the same question.

Participating groups bring questions their lab is well placed to answer: improving a measurement, testing whether a result reproduces, or investigating a response the model cannot explain. We work together to define the experiments and the support needed, with shared protocols, compounds, recordings, and results.

What happens inside the assay?

The mosquito protocol uses 20 adult females above a landing platform. Warmed blood beneath it provides heat and other cues while keeping the blood meal out of reach. A camera records five minutes of behavior. Solvent-only controls help us measure what the compound changed.

Cross-sections showing the compound above the cap in the contact assay and below the cap in the non-contact assay. — In the non-contact assay, mosquitoes can encounter the compound’s airborne components without touching it.

For spotted wing drosophila, a glass tube has raspberry puree at both ends. One end receives the test compound; the other receives control solvent. We record where the flies go over 30 minutes.

Separate the compound’s effect from the experiment around it.

A preference can follow the light, the position of a tube, the order of filming, or the condition of the insects. Repeating an experiment precisely can also repeat its bias. Solvent controls and independent repeats help us identify which differences need investigation.

For example, switching the treated end is not enough if one treatment side always occupies the same bench positions. The design needs to vary those factors independently to separate their effects. If it cannot, the record must preserve that limitation. A calculation cannot recover a comparison the experiment never made.

When a method changes, we retain its version and check comparability with earlier runs. Sharing the protocol, controls, and recordings lets another lab investigate the measurement itself, as well as the compound.

Explore the research network

World map highlighting the United States, Canada, Kenya, and Tanzania. Research partners and collaborators are listed below. — Research partners and collaborators across North America and East Africa.

United States: Monarch’s lab in Emeryville, California; Rutgers’ Marucci Center in New Jersey; USDA research, including Jana Lee’s field work in Oregon; Spatial Repellents at the University of Notre Dame in Indiana; and Dweck Lab in Connecticut.
Canada: The University of Toronto’s Department of Chemical Engineering & Applied Chemistry.
Kenya: The Kenya Medical Research Institute (KEMRI).
Tanzania: Ifakara Health Institute.

Real mosquito assay recordings

Control. The baseline configuration, without the test compound.

Non-contact. Airborne exposure, with the compound out of reach.

The red ring marks the landing surface. These clips show the assay configurations. A result requires repeated measurements and comparison with matching controls.

A recording becomes evidence everyone can use.

Computer vision measures what the insect does, frame by frame. Each result travels with the compound, dose, and experimental conditions. Participating labs can examine one another’s recordings, metadata, and results through the lab portal.

A precise estimate of little or no effect is useful evidence. A noisy or incomplete experiment may leave the effect unresolved. We share both, with their uncertainty and quality checks, so an inconclusive result does not quietly become a negative training label.

Record behaviorWhere the insects go, over time.
Measure the responseCounts, positions, and time in each zone.
Compare with controlsWhat changed under matching conditions?
Train and check the modelsThe measured effect, with its conditions and uncertainty.

From a frame to an experimental result

Six mosquitoes present for one second is six mosquito-seconds. Combining detections with time estimates landing-zone occupancy. Comparison with a matching control tells us whether it increased or decreased. This is not a count of distinct insects or separate landings.

For flies, we compare counts at opposite ends of the tube and ask how that preference changes against controls. The record identifies the conditions at each end, the treatment side, and the number of flies in each tube. Testing the compound at opposite ends across tubes helps us check whether a preference follows the treatment or the apparatus.

Egg-laying experiments ask where an insect chooses to lay its eggs. We retain the counts on each surface. If both counts are zero, there is no preference to estimate from that trial. Different behaviors require different measurements.

The response measure and time window should be defined before comparing models. Video measurements need checks against human-scored observations. A fall in occupancy must be examined alongside movement, time out of view, and survival before it is interpreted as avoidance. We retain the time course, including whether an effect appears early or fades during the run.

Treatments need matching solvent controls from the same experimental session. Comparisons across doses or solvents require experiments designed to connect those conditions.

Reproducing the experiment means knowing both the insect and the material: species, strain, age, sex, physiological state, compound identity, batch, purity, concentration, and solvent, alongside the apparatus and environment. Lab and User IDs preserve who did the work. Missing information must remain visible rather than be guessed.

Keep the different answers in the data.

For mosquitoes, time on the landing surface, time in the surrounding zones, and the number visible answer different questions. A lower count on the surface might reflect movement elsewhere in the arena or movement out of view. Keeping those measurements together helps us interpret the change.

For flies, an end preference compares the number at the treated and untreated ends relative to the total at those ends. The average position along the tube adds information about the whole group. The fraction reaching either end helps distinguish a preference from little participation. If no flies reach either end, an end preference is undefined; it is not evidence of no effect.

The full time course matters too. A response in the first few minutes may fade by the end. We preserve the observation windows and the original values, even when one defined endpoint is chosen for modeling. A threshold creates a category for a particular question; it does not replace the experiment.

What does the model learn from?

A training example connects a compound and an experiment to a measured response. Species, dose, solvent, attractant, contact or non-contact exposure, and the time window all help define what that response means.

We need to preserve those differences. A change in mosquito landing-zone occupancy and a fly’s preference for one end of a tube answer different questions. They should not become a single, unexplained percentage of insects “repelled.”

Repeated experiments show how consistent an effect is. Hundreds of frames from one cage or tube are repeated observations, not hundreds of independent experiments. Uncertainty must reflect how treatments were assigned and experiments repeated. The recording and control comparison remain traceable even when a model uses a summary.

A model can predict a measured response directly, or a category derived from a stated threshold. The original value stays in the record, so the threshold can change without repeating the experiment. Untested compounds remain unknown; their model scores are predictions.

The shared record brings four kinds of information together: compound identity and structure; experimental conditions; measured responses and their uncertainty; and any available chemical or receptor evidence. Computational estimates and experimental measurements remain distinguishable. Missing receptor evidence stays unknown.

The same compound can have several records because the insect, exposure, endpoint, or time window differs. Averaging those into one score would hide where it worked and where it did not. Another researcher should be able to trace a label back to the measurement, test a different interpretation, or build a different model.

What does sharing mean in practice?

Every participating lab can access the full shared dataset. In the portal, select “all labs” to see research uploads from across the network. Researchers can examine the evidence and use it in their own work. Public release beyond compound identities and basic repellency data requires separate written permission from Monarch and the contributing institutions.

The source worksheets and footage are shared within the network. Only compound identities and basic repellency data on those compounds are released publicly on a quarterly basis. We ask labs to share within 24 hours of completing a run or the minimal quality review it needs.

Our ambition is to build the largest shared dataset of insect responses to chemical compounds, with enough detail to understand what happened and under which conditions.

A real recording. Its actual measurements.

Control assay · September 2, 2026 · 918 sampled frames. Play the recording or choose a moment.

The green circle encloses the inner zone. The narrow band between the green and gold circles is the boundary zone.

Measured at 0.00 s

Inside the circle: 3
On the boundary: 0
Outside the circle: 2
Total detected: 5

The experimental log records 20 mosquitoes for this control. The counts shown here are computer-vision detections in one sampled frame, not unique mosquitoes or new landings. The same insect can appear in many frames. Detections can be missed or mistaken.

This is a control recording showing the measurement process, not a claim that a compound works.

Inspect the chart and source data

Detections inside the circle over the whole recording

Selected readings from the same recording. Times are rounded to two decimal places.
Time	Inside	Boundary	Outside	Total
0.00 s	3	0	2	5
100.01 s	10	1	9	20
200.02 s	11	0	4	15
300.03 s	10	3	2	15

Source: the processed frame measurements for control recording IMG_0952, linked to its experimental record.

Without the interactive view, the counts above show the first measured frame. The table remains available.

Field data

Shared field evidence

Does it work in real conditions?

A compound can change behavior in an assay and still fail to protect a crop or a person. Temperature, wind, humidity, surrounding odors, and how it is delivered can all change the result.

Which changes in behavior produce useful protection, for which insects, in which settings? A result in one species or location is a starting point. Field experiments need to establish where it holds, how long it lasts, and whether insects are displaced to other people or plants.

Promising candidates move to crop trials and mosquito testing sites. The next phase we’re building toward brings field results back into the same shared learning cycle, helping us choose better compounds and formulations.

The same open-science commitment applies in the field. Sharing the protocols, conditions, and results, including where a candidate falls short, gives other groups a way to examine the evidence and test whether a finding holds in a different setting.

How will field results improve the next prediction?

The field test must define success before it starts. For mosquitoes, landing, blood-feeding, and duration of protection are distinct outcomes. A claim of reduced infection or disease needs health-outcome evidence; an assay or hut result alone cannot establish it. For crops, the question reaches beyond insect position to infestation, damage, and harvest outcomes.

Trials need appropriate controls, randomized treatment allocation, and independent replication, with the formulation, dose, delivery method, and environment recorded. The concentration in a formulation does not by itself tell us the exposure reaching an insect. When a compound falls short, delivery and persistence need investigation alongside its biological effect.

We also need to measure effects on other insects, plant health, and human safety. Bringing field evidence back into model training is part of the next phase we’re building toward. The test is whether it improves predictions in subsequent trials, including settings that were not used to train the model.

Carry the same compound from the bench to the field.

The compound’s identity connects the records; the outcomes retain their own meaning. Reduced landing-zone occupancy in a cage, fewer blood meals in a hut, and less infestation in harvested fruit are different measurements. Field results add evidence about protection rather than replacing the laboratory record.

If a promising laboratory result fails in the field, the next experiment should distinguish possible reasons: insufficient exposure, a short-lived formulation, a different insect population, or a setting in which the behavioral effect does not provide protection. Sharing those conditions and results makes that failure useful to the next lab and the next model.

Rwambwa field station. Outdoor experimental huts in western Kenya let researchers evaluate protective tools against local mosquitoes under field conditions.

Photograph supplied by Eric Ochomo, KEMRI-CGHR. This shows a research setting, not results from a Monarch compound trial.

Explore the field sites and crop trials

Two small experimental huts stand inside separate compartments of a large screened enclosure at KEMRI in Kenya. — **Semi-field testing.** Huts inside a screened enclosure reproduce a household setting while keeping the experiment contained. KEMRI uses these structures to study mosquito behavior and evaluate repellents and other protective tools.

A researcher in protective equipment uses a backpack sprayer to treat blueberry bushes in a field plot. — **2025 · Blueberry plots.** A spray-on repellent is applied to the bushes. The trial monitors fruit infestation, effects on other insects, and possible damage to the plants.

Ripening blueberries on a living plant sit inside loosely hung, large-hole mesh that leaves the fruit accessible to flies. — **2026 · Open fruit clusters.** The mesh catches falling berries for later evaluation. Its large holes leave flies free to reach the fruit.

Crop photographs: Jana Lee / USDA Agricultural Research Service, from the 2025–2026 Monarch trial protocols. Semi-field photograph: Eric Ochomo, KEMRI-CGHR, showing facilities operated with Research World.

In crop trials, harvested fruit is held so researchers can observe flies emerging from it. This measures infestation, a step toward the farmer’s question: how much usable fruit is protected?

Chemistry & mechanism

From the molecule to the animal

Understand why a compound changes behavior.

How does an insect turn a smell into a decision to approach, land, feed, or lay eggs? We want to connect the chemistry of a compound to what happens in the living animal. That means bringing molecular properties, receptor biology, and behavioral evidence together.

Open science means sharing enough of that work for another scientist to question it: the methods, the measurements, and what remains uncertain. A group studying receptors can build on the behavioral evidence; a group studying behavior can test a proposed mechanism.

The chemistry starts before a molecule reaches a receptor. Volatility, solubility, and stability help determine how a candidate can be delivered and how long it may remain available. A promising biological interaction still needs a formulation that produces useful exposure, with plant and human safety evaluated separately.

Our structural approach uses experimental receptor structures and AlphaFold predictions to examine possible interactions with insect odorant receptor complexes, including the co-receptor Orco. Docking proposes how a molecule might sit within a pocket or another site. We ask whether the fit is plausible and how it compares with reference compounds.

An interaction fingerprint records which parts of the receptor a predicted pose contacts and the types of interaction it could make. Comparing these patterns across compounds and receptors may help us find different chemistry worth testing. The comparison depends on reliable structures, correct receptor identities, and a common frame of reference.

A plausible pose does not establish binding, activation, inhibition, or repellency. Those are separate questions. Structural and mechanism studies can proceed alongside behavioral and field experiments, with each informing the next.

Chemical and structural evidenceWhat could the molecule do?
Examine its properties and predicted interactions. Keep measured properties distinct from computational estimates.
Functional evidenceDoes the receptor respond?
Measure the response across concentrations, including whether the compound activates, inhibits, or changes a response to another cue.
Behavioral evidenceDoes the insect behave differently?
Test the whole animal under defined conditions, with controls. A receptor response alone cannot establish protection.

Complementary kinds of evidence, not a report of a completed mechanism study. No single score answers all three questions.

How does this feed back into the model?

The aim is to connect a compound’s identity and properties, its predicted interaction pattern, measured receptor responses, and observed behavior. A missing measurement stays unknown. Evidence about plant damage, human safety, or practical delivery retains its own meaning rather than disappearing into a single “repellency” score.

The receptor work documented so far emphasizes mosquitoes. A pattern found in those receptors does not establish the same effect in a crop pest. Both the species represented and the reliability of each structure matter when deciding how far a prediction can travel.

A stronger explanation makes a testable prediction. We can alter a proposed sensory pathway and ask whether the predicted behavioral change follows, while checking general activity and responses to other cues. Contact and airborne effects may involve different pathways. Disagreement between structural, functional, and behavioral evidence helps define the next experiment.

Receptor experiments can add a concentration–response curve: how the measured response changes as the amount of compound increases. Any summary of potency belongs with the receptor identity, species, assay conditions, and type of response. It is not a context-free property of the molecule or a measure of field protection.

Bringing these evidence streams into a combined assessment is a development goal. We can compare a model using chemical structure alone with one also using receptor information, holding the compounds, behavioral targets, and evaluation splits constant. The comparison must use inputs that would actually be available when making the next prediction. An improvement supports the usefulness of the added evidence; it does not establish a causal mechanism by itself.

A computational pose, a functional response, and an insect’s behavior may disagree. No response in one receptor assay does not exclude another receptor or another pathway. These disagreements help us choose what to test and give collaborating groups a specific question to investigate.

Commercialization

Open discoveries, practical use

Bring it to the people who need it.

Turn promising compounds into formulations that work in the field and on human skin. Test safety and performance, obtain the required approvals, and manufacture them at a cost people can afford.

Compound identities and basic repellency data on those compounds are released publicly each quarter, so researchers, companies, and other organizations can build on those discoveries. Public release of anything else requires separate written permission from Monarch and the contributing institutions.

The work continues beyond identifying a compound. Can a formulation maintain protection for long enough, be manufactured reliably, and fit how a farmer or a family will actually use it? For public health, that includes affordability, consistent use, and added protection alongside existing prevention tools. These practical constraints should shape the next formulation and experiment.

Use the evidence to choose what to develop.

A compound that changes behavior on contact may need a different delivery system from one that works through airborne exposure. The experimental record preserves that distinction. Concentration–response measurements, volatility, solubility, and stability help define the formulations worth testing; no single model or docking score establishes that a product will work.

The comparison moves from a promising molecule to a practical formulation: protection at an achievable exposure, for a useful duration, with acceptable safety and cost. Sharing discoveries and supporting evidence lets others pursue applications the original experiment did not address. What they learn can guide the next question.

Who can build on the discoveries?

Our public-release commitment covers only compound identities and basic repellency data on those compounds, published each quarter. Public release of anything else requires separate written permission from Monarch and the contributing institutions. Researchers, companies, and other organizations can build on the findings. Turning them into useful products still requires formulation, safety testing, manufacturing, and the approvals appropriate to each use.

01
Effective formulations
Make the compound work where it will be used.
02
Affordable production
Manufacture it at a cost people can afford.
03
Farmers and communities
Build the partnerships that bring it to them.

Open discoveries give useful solutions more ways to reach people.

Everyone contributes. Everyone can build on it.

Join the research network

The Monarch flywheel

Everyone contributes. Everyone can build on it.