Announcing our NeurIPS 2026 Oral on Emergent Biological Capabilities in BonbonBonbon · NeurIPS 2026 Oral

Read the blog
Back

Bonbon: A vertical foundation model for biology with emergent capabilities

Read the paper ◆ NeurIPS 2026 Oral

We’re excited to share Bonbon, a vertical foundation model for biology with emergent biological capabilities. AI models have learned to write, code, and reason, developing abilities beyond what they were explicitly trained to do. Emergent abilities have appeared in models trained on individual types of biological data, such as protein or DNA sequences. But to our knowledge, no previous multimodal model for biology has demonstrated these emergent capabilities in molecular interactions, learning biological properties beyond its explicit training objective without task-specific supervision.

But biology works through interactions. What a molecule does depends on what it interacts with, and those relationships determine biological outcomes. Yet AI for molecular interactions remains fragmented into narrow, task-specific models, often trained with explicit labels and focused on the geometry of how molecules fit together.

Biology is a complex system. How a cell behaves comes from many relationships operating together, most of them still unknown to us. Understanding that system requires understanding what interactions do and how they influence one another, beyond whether two molecules bind or geometrically fit. We built Bonbon to learn these hidden relationships, so we could expand what is possible to understand and design.

Models for molecular interactions have not yet demonstrated this kind of emergence beyond their training objective. Co-folding models learn to predict three-dimensional structures, but adding a capability such as affinity prediction requires additional supervised training on affinity measurements. The new capability is explicitly taught rather than emerging from the original training.

Our hypothesis was that the relationships governing molecular biology are encoded in the sequences of interacting molecules, and that with the right architecture, data, and training, they could be revealed. As we scaled Bonbon, we saw capabilities emerging across the hierarchy of biology. Bonbon independently learned to answer biological questions without structural coordinates, physics-based calculations, or task-specific supervision.

Training for the relationships in biology

Training a model for multimodal biology brings different challenges from those encountered in language and vision. An image contains dense information about objects, their positions, textures, and lighting. A biological interaction example carries a much sparser signal, and not every example contains information relevant to every biological property. The patterns that distinguish activation from inhibition, or covalent from non-covalent engagement, become visible only across many combinations.

Molecular relationships are also many-to-many. One protein can interact with many molecules, and one molecule can interact with many proteins. The model needs to learn how the relationship changes with each combination.

We designed Bonbon’s architecture and training objective for these relationships and trained it on approximately 3.2 trillion tokens of molecular sequences. The dataset spans proteins, small molecules, peptides, and protein–protein interactions. We did not provide structural coordinates, affinity measurements, or physics-based calculations during pretraining, or labels for the emergent capabilities described below.

To account for the sparsity of biological data, we used an unusually large amount of training data relative to model size: approximately 2,000 tokens per parameter. We then scaled Bonbon across six model sizes, from 42 million to 1.6 billion parameters, to examine what it learned.

  • Bonbon 2,000×
  • ESM-2 (650M) 100×
  • Evo 43×
  • Chinchilla 20×
  • GPT-2 6.7×
  • ESM-2 (15B) 4.3×
  • GPT-3 1.7×

Training tokens per parameter

Learning biological relationships requires extensive exposure to data.

Biological knowledge is spread across many examples of molecular interactions. We trained Bonbon with approximately 2,000 sequence tokens per model parameter, giving it extensive exposure to those examples relative to its size. The chart compares this training ratio with those of several language and biology models.

Emergent capabilities at scale

As we increased model scale and training compute, pretraining loss decreased smoothly. This measures how well the model performs its training objective, but it gave no indication that new biological capabilities were about to appear.

Pretraining loss

Training compute (FLOPs)

Smooth. Predictable. Uninformative.
Training improved smoothly as we increased scale.

Across six model sizes, Bonbon steadily improved at its training objective. This measure, called pretraining loss, gave no clear indication of the sharp changes in biological capabilities shown in the next figure. Measuring training progress alone would have missed that transition.

At the functional level, Bonbon learned to distinguish the biological outcome of an interaction. At the residue level, it learned to locate binding sites and distinguish main-site from alternative-site engagement. At the bond level, it learned to recognize covalent engagement and identify the reactive residue. And at the atomic level, it learned to identify the active part of a molecule.

Across the five smaller models, these capabilities stayed within a relatively narrow band. At 1.6 billion parameters, they improved sharply. We used continuous evaluation measures to address the concern that gradual improvement can look sudden when reported only as passing or failing a test. The sharp improvement remained.

Mechanism of Action

Macro precision

Training compute (FLOPs)

Six-class mechanism precision.

Orthosteric/Allosteric Discrimination

AUROC

Training compute (FLOPs)

The five smaller scales track the fingerprint baseline, within 0.06 of it.

Covalent Bond Recognition

AUROC

Training compute (FLOPs)

Flat through five scales, then 0.775.

Active Moiety Identification

Mean attention coverage

Training compute (FLOPs)

At chance for five scales, then 0.846.
Biological capabilities emerged beyond the training objective.

At the largest tested scale, Bonbon showed a sharp improvement in understanding biological effects, distinguishing binding modes, recognizing covalent engagement, and identifying active molecular regions. We had not trained it explicitly to perform these tasks. These capabilities followed a different pattern from the smooth improvement in pretraining loss.

Panel 1 of 4

This matters because training loss is commonly used to decide how to scale a model. Our results show that it can miss the transition when new capabilities emerge. We extend this finding, previously observed in language models, to a multimodal foundation model for biology.

Earlier biological models demonstrated emergent abilities from individual types of sequence data, such as proteins or DNA. Bonbon extends this to multimodal biology, learning properties that depend on how different molecular components interact. To our knowledge, this is the first demonstration of these emergent biological capabilities in a model for molecular interactions.

Accuracy and speed from shared interaction representations

Bonbon encodes what it learns in interaction representations that capture what it has learned about the relationship between molecular partners. We use the same interaction representations to examine biological effects, binding location, bond formation, and other properties.

We first tested these interaction representations on two established tasks: whether a molecule binds a protein and how tightly it binds. The same interaction representations that support Bonbon’s emergent abilities also achieved strong binding and affinity performance.

On CASP16, Bonbon’s affinity estimates reached a correlation of 0.74 with experimental measurements, compared with 0.52 for Boltz-2 and 0.54 for IntFold. On FEP+4, Bonbon reached 0.78, matching the physics-based method FEP+ and surpassing Boltz-2 at 0.66.

For affinity prediction, we trained a small probe on experimental affinity measurements while keeping Bonbon frozen. Training took approximately 12 minutes, showing that useful affinity information was already present in its interaction representations.

FEP+ 4 Targets

Mean Pearson correlation (r)

OpenFE

Mean Pearson correlation (r)

CASP16

Mean Pearson correlation (r)
  • Bonbon
  • Physics-based >1h
  • Physics-based <1h
  • Structure-based ML
  • Sequence-based ML
The same interaction representations capture binding strength.

Binding strength helps researchers decide which molecules to pursue and improve. Bonbon’s interaction representations supported strong agreement with experimental measurements across three benchmarks, matching FEP+ on the four-target comparison. We trained a small readout for approximately 12 minutes while keeping Bonbon itself unchanged.

Bonbon processes approximately 2,500 interactions per second on a single H100, about 50,000 times faster than Boltz-2 in our comparison. Physics-based methods such as FEP+ and ABFE can require hours per calculation, which limits their use to small sets of selected candidates. Bonbon brings comparable accuracy on FEP+4 to a throughput that supports much broader exploration. This matters because investigating biological systems requires examining many combinations and multiple properties of each interaction.

  • Bonbon (cached) 153M
  • Bonbon (on-the-fly) 9.0M
  • TerraBind 8K
  • AutoDock Vina 4K
  • Boltz-2 302
  • AlphaFold3 11

Pairs evaluated in one hour

Throughput: protein–ligand pairs evaluated in one hour on a single H100 GPU.

Bonbon evaluates approximately nine million new protein–ligand pairs per hour, about 50,000 times faster than co-folding models in our comparison. Reusing precomputed interaction representations increases throughput to approximately 153 million pairs per hour, enabling large-scale screening of molecular libraries and evaluation of newly generated molecules.

Learning the functional outcome of an interaction

For drug development, knowing that a molecule binds is not enough. We need to know what happens when it binds. Does it activate the target, inhibit its activity, or block a signal that would otherwise activate it? Two molecules can bind the same protein and produce different outcomes.

We first visualized Bonbon’s interaction representations across six biological categories: agonists, antagonists, inhibitors, activators, substrates, and products. These describe molecules that regulate activity or participate in biochemical reactions. The interaction representations were organized by those roles, even though Bonbon had never been trained on their labels.

Hover a class to isolate it.
Interaction representations organize around biological roles.

Each point represents a molecular interaction, colored by its experimentally established biological role. As we move between model sizes, we can see how Bonbon’s interaction representations organize around these roles, even though their labels were not supplied during pretraining. This gives us a view into biological knowledge that developed within the model.

We extracted a fixed signature of 160 features from the frozen model that distinguishes the six categories. Using a dataset of 107,027 experimentally measured interactions, we tested it on held-out groups of structurally dissimilar molecules and achieved 81.1% macro precision, giving each category equal weight.

Six mechanism classes

True class
Agonist 0.79 ± 0.02 0.07 ± 0.02 0.06 ± 0.02 0.00 ± 0.01 0.00 ± 0.00 0.00 ± 0.00
Antagonist 0.12 ± 0.01 0.83 ± 0.02 0.04 ± 0.01 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00
Inhibitor 0.08 ± 0.01 0.10 ± 0.01 0.88 ± 0.02 0.01 ± 0.00 0.01 ± 0.00 0.02 ± 0.01
Activator 0.00 ± 0.00 0.00 ± 0.00 0.01 ± 0.01 0.99 ± 0.01 0.00 ± 0.00 0.01 ± 0.01
Product 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 0.00 ± 0.00 0.69 ± 0.07 0.28 ± 0.09
Substrate 0.00 ± 0.00 0.00 ± 0.00 0.01 ± 0.01 0.00 ± 0.00 0.30 ± 0.06 0.69 ± 0.09
Agonist Antagonist Inhibitor Activator Product Substrate

Predicted class

Learning what an interaction does.

A molecule’s biological role determines what its interaction means for the system. Using a fixed signature from Bonbon’s interaction representations, we distinguished six roles with 81.1% average precision across categories. The diagonal shows correct assignments, while the other cells show where the roles were confused.

To make the test harder, we asked whether Bonbon could distinguish agonists, which activate a receptor, from antagonists, which block its activation, on the same receptor. We collected more than 28,000 measurements across 222 GPCRs, membrane receptors that allow cells to respond to signals. GPCRs are among medicine’s most important targets: 36% of approved drugs act on GPCRs. Every receptor had both agonists and antagonists.

The model cannot answer the question just by memorizing the protein, because every receptor appears with both outcomes. It has to distinguish what happens with the particular molecule interacting with that receptor. Bonbon achieved 85.5% balanced accuracy, giving equal weight to both classes, on held-out molecule groups.

Agonist vs antagonist

True class
Agonist 0.88 ± 0.01 0.12 ± 0.01
Antagonist 0.17 ± 0.04 0.83 ± 0.04
Agonist Antagonist

Predicted class

Agonist against antagonist, the binary call.
The same receptor can respond differently to different molecules.

Agonists activate a receptor, while antagonists block its activation. We tested Bonbon on 222 receptors that each had examples of both outcomes, removing receptor identity as a shortcut. Bonbon distinguished the outcomes with 85.5% balanced accuracy on held-out groups of molecules.

Establishing these effects ordinarily requires functional laboratory assays. To our knowledge, no prior computational method has demonstrated this kind of zero-shot mechanism-of-action prediction across protein families. Bonbon learned the relevant distinctions without being explicitly taught them.

Locating where an interaction happens

Knowing where a particular molecule engages a protein helps researchers understand its effect and decide how to improve it. A protein may have several interaction sites, so identifying a pocket in general does not establish which one a particular molecule uses.

We compared Bonbon’s attention, its internal weighting of sequence positions, with experimentally determined structures. Across 1,176 structures, attention was higher at protein positions closer to the bound molecule and decreased as distance increased. This held on both training and held-out proteins, across families including enzymes, receptors, ion channels, and transporters.

Training (ID)

Attention weight (fold over mean) x = 2.5 · mean 1.74 · n = 240 x = 7.5 · mean 1.60 · n = 240 x = 12.5 · mean 1.24 · n = 240 x = 17.5 · mean 0.95 · n = 240 x = 22.5 · mean 0.72 · n = 240 x = 27.5 · mean 0.58 · n = 240

Distance to ligand (Å)

Binned means measured. Individual points illustrative, drawn from each bin's mean and spread.

Training (ID)

Held-out (OOD)

Attention weight (fold over mean) x = 2.5 · mean 1.52 · n = 240 x = 7.5 · mean 1.38 · n = 240 x = 12.5 · mean 1.16 · n = 240 x = 17.5 · mean 0.94 · n = 240 x = 22.5 · mean 0.78 · n = 240 x = 27.5 · mean 0.66 · n = 240

Distance to ligand (Å)

Binned means measured. Individual points illustrative, drawn from each bin's mean and spread.

Held-out (OOD)
Attention concentrates near the interacting molecule.

Across 1,176 experimentally determined structures, Bonbon assigned more attention to protein positions closer to the bound molecule. Attention decreased with distance, and the pattern persisted on held-out proteins. The model had learned information about spatial relationships from sequences alone.

  • KCNQ2
  • Binding site
Normalized signal (0–1)

Amino acid position (UniProt)

KCNQ2
  • BCL-2
  • Binding site
Normalized signal (0–1)

Amino acid position (UniProt)

BCL-2
  • GPR84
  • Binding site
Normalized signal (0–1)

Amino acid position (UniProt)

GPR84
  • Tankyrase-1
  • Binding site
Normalized signal (0–1)

Amino acid position (UniProt)

Tankyrase-1
Bonbon’s attention follows the regions involved in binding.

We compared Bonbon’s attention along each protein sequence with the experimentally observed position of the bound molecule. Peaks in attention often align with nearby protein regions, including in these challenging examples. Bonbon learned this localization without structural coordinates or binding-site labels.

Panel 1 of 4

  • KCNQ2
  • GPR84
  • BCL-2
  • Tankyrase-1
Sequence learning reveals where molecular interactions happen.

Here we display Bonbon’s attention on known protein structures to make its localization visible. Blue indicates lower attention and red higher attention, with the bound molecule shown in green. The structures provide a reference for evaluation; Bonbon did not receive them as inputs.

We also examined membrane proteins and proteins with disordered regions, where binding sites can be harder to characterize. Bonbon localized interactions despite learning without structural coordinates. On the COACH420 binding-site benchmark, Bonbon achieved an F1 score of 0.497, compared with 0.45 for Fpocket, 0.44 for P2Rank, and 0.32 for ESMBind. It did so from sequences alone, without structural coordinates or binding-site supervision.

The capability here is molecule-specific localization from sequences alone, without binding-site supervision. The compared approaches require structural inputs or annotations that Bonbon does not, allowing researchers to investigate an interaction site directly from its learned signals.

  • ID
  • OOD
Kinase
RTK
Protease
GPCR
Oxidored.
Epigenetic
Ion Channel
Nuclear Rec.
Transporter

F1 score

Binding-site localization extends to unfamiliar proteins.

We tested whether Bonbon could identify binding regions across nine protein families. Performance was broadly comparable between proteins represented during training and held-out proteins, showing that localization extends beyond familiar examples. This matters when investigating targets for which detailed prior knowledge is limited.

Distinguishing main-site and alternative-site engagement

We next examined a distinction especially useful for drug design. The orthosteric site is where a protein’s natural binding partner acts. An allosteric site is a different location through which a molecule can influence the protein’s activity. Alternative sites can offer greater specificity, finer control, and ways to address proteins whose main sites are difficult to target.

We constructed a parameter-free signature of 145 features from Bonbon’s interaction representations to distinguish orthosteric from allosteric engagement. We then tested it on proteins excluded from training. Across 6,032 interactions involving unseen proteins, Bonbon achieved an AUROC of 0.834, compared with 0.589 for a molecular-fingerprint baseline.

AUROC = 0.834 Density

Signature score

Two populations, not one.
Learning two different ways to influence a protein.

A molecule can act at a protein’s main binding site or influence it through an alternative site. Across 6,032 interactions involving unseen proteins, Bonbon’s interaction representations distinguished these binding modes with an AUROC of 0.834. Recognizing the distinction opens additional ways to pursue selectivity and control.

Signed feature vector, 1 to 145: orthosteric-associated left of the rule, allosteric-associated right

Cohen's d (clipped at ±1)
-1 0 1
One ordering, five classes.
Shared interaction representations contain information about binding mode.

We identified a fixed pattern of 145 features associated with main-site or alternative-site engagement. The heatmap shows how that pattern appears across different protein families. It allows us to examine binding mode through the same interaction representations used for other biological questions.

We then selected only proteins with both orthosteric and allosteric interactions. Holding the protein constant removes its identity as a shortcut: the model must distinguish the site used by the particular molecule. Bonbon achieved an AUROC of 0.821 across 1,860 interactions. Performance varied across protein groups, with transcription factors remaining a difficult case.

To our knowledge, no prior computational method has demonstrated this per-molecule discrimination. Existing approaches struggle with the allosteric distinction: fine-tuned protein models have shown very low performance on allosteric-site detection, while co-folding models have repeatedly placed allosteric molecules incorrectly in the protein’s main binding site instead of their actual allosteric site.

Recognizing allosteric engagement opens additional ways to design medicines. Researchers can explore sites that offer finer control over protein activity, greater selectivity between related proteins, and access to targets whose main binding sites have been difficult to address.

Recognizing covalent bonds and their attachment points

Covalent drugs form a chemical bond with their target, which can help address proteins that are difficult to engage through other approaches. Designing such an interaction requires knowing whether a bond will form and where. Structure-based approaches show whether two atoms can sit close together, while covalent docking assumes in advance that a covalent bond forms. Proximity alone does not establish that the atoms will react: that also depends on their chemistry and the surrounding protein environment.

We found a fixed signature of 36 features in Bonbon’s interaction representations that distinguished covalent from non-covalent engagement. We then tested whether the model was relying on an obvious clue: an electrophilic warhead, a reactive chemical group that can participate in covalent bonding.

0.00 0.35 0.69
Frequency

Representation feature

Two populations, one feature ordering.
Covalent engagement leaves a pattern in interaction representations.

We found 36 features whose patterns distinguish covalent from non-covalent engagement across protein families. Together, they provide a fixed readout of bond-forming interactions from Bonbon’s interaction representations. The model was not explicitly trained to recognize this distinction.

First, we compared molecules that all carried a warhead, but only some of which formed covalent bonds. Bonbon distinguished them with an AUROC of 0.834. We then tested interactions in which neither class carried a detectable warhead. It still distinguished covalent engagement, achieving 0.775, including cases where bond formation depends on the protein’s catalytic machinery.

AUROC
Recognizing bond formation beyond the presence of a reactive group.

A reactive group, often called a warhead, does not guarantee that a molecule will form a covalent bond with a protein. We tested Bonbon in two settings: when all molecules carried detectable warheads, and when none did. In both, Bonbon distinguished molecules that formed covalent bonds from those that did not, outperforming methods based on molecular fingerprints. These controls show that detecting a warhead alone cannot explain what Bonbon learned about covalent engagement.

We next asked whether Bonbon could identify the attachment point. Some amino-acid side chains can react with a molecule to form a covalent bond. We examined three relevant residue types: cysteine and serine, established targets of covalent drugs, and lysine, another important opportunity for covalent design. The full pipeline first identifies the likely amino-acid type and then selects the specific residue on the protein where the bond forms. Without being told that position in advance, it selected the correct residue in 71.8% of cases.

Covalent vs non-covalent

Chance

AUROC
  • Which residue

Chance

AUROC
Recognizing bond formation beyond an obvious chemical clue.

A reactive chemical group does not by itself establish whether a molecule will form a covalent bond with a protein. Bonbon distinguished covalent engagement both when all molecules carried such a group and when none carried a detectable one. Its interaction representations also supported a localization pipeline that selected the correct attachment point in 71.8% of cases.

Bonbon recognizes covalent engagement zero-shot from sequences, without structural inputs, and its interaction representations also support localization of the reactive residue. To our knowledge, no prior method has demonstrated this zero-shot, structure-free recognition of covalent engagement.

Identifying the active part of a molecule

The final emergent ability concerns the active moiety, the part of a molecule involved in engaging its target. Identifying it matters when optimizing a drug: researchers need to know which atoms to preserve and which parts they can change to improve properties such as potency or selectivity.

Researchers typically investigate this through structure–activity relationship, or SAR, studies. They make and test dozens of related molecules, changing parts of the structure to learn which changes preserve activity and which destroy it. These repeated synthesis-and-testing cycles can take months. We wanted to know whether Bonbon had learned to recognize active regions without being taught where they were.

We used experimental structure–activity relationship (SAR) data to identify the active regions against which we compared Bonbon’s attention. This let us test whether the model highlights atoms whose modification changes biological activity. Across more than 241,000 interactions involving 145 proteins, approximately 85% of Bonbon’s molecular attention fell within the annotated active region, compared with a 36% baseline for uniformly distributed attention. Attention remained high inside the region and dropped sharply one bond outside it.

Mean attention per atom
Attention falls off a cliff exactly at the boundary.
Attention drops sharply at the active region’s boundary.

Knowing which part of a molecule engages its target helps researchers decide what to preserve during redesign. Bonbon’s attention remained high inside the experimentally annotated active region and dropped sharply just one chemical bond outside it. The pattern reveals a distinction between the active region and its surroundings without explicit training on those boundaries.

We saw the pattern across different protein classes and in out-of-distribution evaluations on proteins and molecules. Bonbon continued to identify active regions beyond the proteins and chemical groups represented in training. To test whether Bonbon had simply associated each protein with a familiar chemical pattern, we held the protein constant and changed the molecules.

  • sulfonamide O N S N O O O N N N N
    Enzyme coverage 0.82 sulfonamide
  • piperidine N N S O O
    GPCR coverage 0.84 piperidine
  • fluorobenzene O S O N F F N N N N S N
    Ion Channel coverage 0.85 fluorobenzene
  • sulfone S O O F F F F O
    Nuclear Receptor coverage 0.88 sulfone
  • pyrimidine N N N N N N N O N
    RTK coverage 0.89 pyrimidine
relative attention
Five target classes, five chemotypes, the same behaviour.
Identifying active molecular regions across different protein families.

These examples show Bonbon concentrating attention on the annotated active region of molecules acting on five protein classes. The boxed region was identified from experimental activity data, while the colored atoms show the model’s attention. This provides information that could guide which parts to preserve while exploring changes elsewhere.

We repeated this test on JAK2, BACE1, and the adenosine A2A receptor, spanning different protein families. For each protein, we examined four structurally distinct molecules with different active moieties. Bonbon’s attention followed the appropriate active region as the molecule changed.

  • amide / fused N-bicycle O N Cl O F F N N S N N N
    Amide bicycle coverage 0.83 amide / fused N-bicycle
  • aniline N N N N N S O O N N
    Aniline coverage 0.80 aniline
  • aminopyrazole N N N N S O O N N O N N
    Aminopyrazole coverage 0.87 aminopyrazole
  • phenol / aryloxy O N N N F O F F F N N F N S O O
    Aryloxy phenol coverage 0.84 phenol / aryloxy
JAK2 relative attention
Four chemotypes against one protein.
Different molecules require different active regions, even on the same target.

We held JAK2 constant and examined four structurally distinct molecules, each with a different active region. Bonbon’s attention followed the appropriate region as the molecule changed. This matters when pursuing several chemical approaches to one target, because each approach may have a different core that needs to be preserved.

  • amide / fused N-bicycle N N S N N O N N
    Amide bicycle coverage 0.85 amide / fused N-bicycle
  • pyrazole / imidazole N N N N N N O O N N O
    Pyrazole imidazole coverage 0.84 pyrazole / imidazole
  • fused N-heteroaromatic N N N N O O O O S S O N N N N
    Fused heteroaromatic coverage 0.87 fused N-heteroaromatic
  • pyrimidine N O N Cl N O O O N
    Pyrimidine coverage 0.89 pyrimidine
Adenosine A2A relative attention
Four chemotypes against one protein.
Different chemical ways to engage the same receptor.

We held the adenosine A2A receptor constant and examined four structurally distinct molecules with different active regions. Bonbon’s attention followed the appropriate region as the molecule changed, with 84–89% falling within the annotated active region. This shows why recognizing the particular interaction matters: different chemical approaches to the same target can require different molecular cores to preserve during redesign.

Docking and co-folding show which atoms are predicted to contact a protein, but contact does not establish which atoms are essential for activity. Researchers still use expensive, often months-long campaigns of making and testing related molecules to determine what must be preserved. Bonbon identifies active-region information zero-shot, without structural coordinates or explicit training on active-moiety labels.

This creates a useful connection to generative design. We could preserve the active region and explore changes elsewhere while evaluating properties such as binding strength. It gives the design process a more informed starting point for deciding what to keep and what to change.

Discovering new molecules in the lab

We chose nine diverse disease-related targets, grouped into three levels of difficulty. For each, we generated candidate molecules and used Bonbon to rank them. Up to 24 selected compounds were synthesized and tested by an external contract research organization, without iterative medicinal-chemistry optimization before that first experimental round. Across nine disease-related targets, Bonbon achieved a 61.5% dose-response hit rate in assays performed by an external contract research organization.

Well-characterized targets
  • TYK2 23 of 24 confirmed IC50 146 nM
  • PRMT5 16 of 24 confirmed IC50 5.9 µM
  • JAK1 21 of 24 confirmed IC50 3 nM
  • CDK1 22 of 22 confirmed IC50 114 nM
Beyond established pharmacophores
  • HDAC6 15 of 24 confirmed IC50 115 nM
  • MNK2 19 of 24 confirmed IC50 39 nM
  • MMP-9 3 of 23 confirmed IC50 480 nM
Out-of-distribution targets
  • ELANE 2 of 24 confirmed IC50 1.22 µM
  • VEGFR2 10 of 24 confirmed IC50 38 nM
The same interaction representations lead to novel inhibitors in the lab.

We used Bonbon to rank newly generated molecules for nine disease-related targets, then synthesized and tested up to 24 compounds per target. Without iterative medicinal-chemistry optimization, we achieved a 61.5% dose-response hit rate, including structurally novel inhibitors and active molecules for targets excluded from training. These experiments show that the same interaction representations behind Bonbon’s emergent abilities support the discovery of new, experimentally active chemistry.

Novel inhibitors on well-characterized targets

The first group included TYK2, PRMT5, JAK1, and CDK1. These targets have been extensively studied, with established inhibitor chemistry. We wanted to know whether scoring each candidate in a single forward pass through Bonbon could identify potent, novel molecules without iterative medicinal-chemistry optimization. We then synthesized and tested the selected candidates.

On TYK2, we confirmed 23 inhibitors, including 17 novel molecules. On PRMT5, we confirmed 16 inhibitors without reproducing individual known compounds. On JAK1, we confirmed 21 inhibitors, with the strongest at 3 nanomolar. On CDK1, all 22 compounds tested were active, including 10 novel molecules.

New chemistry beyond established approaches

The second group tested whether Bonbon could find alternatives to established chemical strategies. HDAC6, implicated in cancer and neurodegeneration, is an example: much of its inhibitor chemistry relies on hydroxamate groups, and moving beyond their associated liabilities is a longstanding research goal.

We confirmed 15 HDAC6 inhibitors. Twelve were novel molecules that lacked the hydroxamate group, providing new chemical approaches to a target where that chemistry has dominated.

We confirmed 19 MNK2 inhibitors, including 18 novel molecules outside the documented inhibitor chemistry. Each of the 18 novel molecules has a different chemical scaffold, and several showed activity at nanomolar concentrations.

For MMP-9, we confirmed three inhibitors, each without the strong zinc-binding warheads used by many established MMP inhibitors. This offers a different chemical direction to investigate for a target where selectivity across related proteins has been difficult. These results establish new chemical starting points worth investigating; their safety and therapeutic potential require further work.

Generalization to targets excluded from training

The final group posed the hardest test: VEGFR2 and ELANE were absent from training, along with proteins sharing more than 90% sequence identity with either target. We wanted to know whether Bonbon could identify unfamiliar, active molecules for proteins it had never learned from, including molecules novel against the known inhibitor literature.

Bonbon discovered confirmed inhibitors for both unseen targets, including a novel molecule for each and a VEGFR2 inhibitor active at 38 nanomolar. The same interaction representations behind its emergent abilities translated into new, experimentally active chemistry on proteins excluded from training.

Connected biological knowledge in shared interaction representations

These emergent abilities arise from a shared space of interaction representations. We did not fine-tune Bonbon separately or add a new architecture for each capability. Each capability addresses an important drug-discovery problem on its own. Together, they bring biological effect, binding location, bond formation, and active molecular regions into one shared space of interaction representations, making these questions accessible from the same model.

These results establish a starting point for what Bonbon’s interaction representations can support. Task-specific fine-tuning gives us a further route to develop and extend the capabilities already present in the model.

What makes these results interesting is that the new molecules work despite being different from known inhibitors. This suggests that Bonbon has learned something about what an interaction requires and can recognize different chemical ways to meet those requirements. That is what we mean by biological reasoning.

Our explanation is that Bonbon learns the components of an interaction and, at sufficient scale, becomes able to compose them. Recognizing covalent engagement, for example, requires combining information about a reactive group, a candidate protein residue, and its local environment. We propose that the sharp improvement reflects the model becoming able to bring these pieces together.

Researchers can therefore examine several questions before making a molecule: what will the interaction do, where will it happen, how will it engage, and which atoms should be preserved? Bringing these questions into one inference pass reduces the cost of exploring a drug design before committing to synthesis and experiments. This opens a path to more ambitious medicines and puts some of the hardest drug programs within reach.

From interactions to systems, and from understanding to design

Bonbon’s throughput creates a foundation for moving toward the scale at which biology operates. Functional information is especially important: knowing that one interaction activates a component while another blocks it helps us reason about how effects could propagate through a pathway. The next step is to connect these interactions into pathways and living cells. We are already extending Bonbon toward entire cells, building on its ability to reason about both molecular engagement and functional outcomes.

The same interaction representations open a path toward one-shot drug design. Today, improving one property of a molecule can undermine another, sending a program through another round of redesign and testing. Shared interaction representations create a path to considering those requirements together from the outset, making more ambitious combinations of properties practical to pursue.

We could begin with a specification: the mechanism we want, the site to engage, the targets to influence, the interactions to avoid, and more. Generative design could then compose candidates around those requirements, with experiments serving to confirm whether they achieve the intended biological behavior.

We are extending Bonbon toward entire biological systems and using its interaction representations to guide the design of new molecules. Together, these efforts connect understanding how biology works with composing new ways to influence it.