Topos Bio AI Models Target Proteins That Refuse to Hold Still

Topos Bio launched two AI models on September 30, targeting proteins that shift among many shapes instead of holding one stable structure. The Topos Bio AI models challenge a dominant assumption in structure-based drug discovery: that researchers can design around one reliable molecular snapshot.

Topos-2 generates collections of protein conformations from an amino acid sequence. Topos-Bind adds a potential drug molecule and predicts how the protein-ligand system changes across those conformations. Topos calls these collections conformational ensembles, meaning the range and relative distribution of shapes a protein can adopt.

That distinction matters for intrinsically disordered proteins, or IDPs, which lack a single stable three-dimensional structure. Such proteins are linked to neurodegeneration, cancer, and metabolic disorders. Yet the models enter a competitive field that already includes BioEmu, AlphaFold-derived methods, Boltz-2, Chai-1, and OpenFold3.

Topos reported strong benchmark results and an amyloid-beta case study. However, its most consequential claims still depend on company-run comparisons, simulated reference data, and models that have not yet produced a clinically validated medicine.

The Topos Bio AI Models Address Two Different Gaps

Topos-2 models how proteins move, while Topos-Bind asks how a small molecule changes that moving system.

Topos-2 is an all-atom generative model. It accepts an amino acid sequence and produces multiple structures representing possible states of the protein. “All-atom” means the output includes individual atoms rather than a simplified representation of larger molecular units.

The model expands the scope of Topos-1. That earlier system focused on disordered proteins and disordered regions within larger proteins. Topos-2 covers the broader spectrum from ordered proteins to partially and fully disordered systems.

This design makes the model different from systems optimized to return one likely structure. A single structure can be useful when a protein folds into a relatively stable shape. It becomes less informative when motion and structural variability influence the protein’s biological role.

Topos says Topos-2 performed best on the independent PeptoneBench framework. That benchmark compares computational ensembles with measurements from nuclear magnetic resonance and small-angle X-ray scattering experiments.

The company reported a reweighted composite score of 1.11 for Topos-2. BioEmu, the next-best system in its comparison, scored 1.49. Within this benchmark, 1.0 represents agreement inside the estimated uncertainty of the experimental and computational measurements.

The phrase “four times closer” needs context. It compares each score’s distance from 1.0, rather than saying Topos-2 is four times more accurate across all uses. The reported distances are 0.11 and 0.49, respectively.

PeptoneBench itself is separate from Topos. Its public benchmark repository provides code for evaluating predicted ensembles against experimental observables. It also instructs participants to generate at least 100 configurations for each sequence.

Topos-Bind tackles a second problem. It accepts a protein sequence and a small molecule represented as a SMILES string, which encodes molecular structure as text. The model then generates an ensemble for the combined protein-ligand system.

According to the company’s model release, Topos-Bind is the first generative model designed specifically for ensembles of disordered proteins bound to small molecules. That priority claim has not received broad independent validation.

Topos has previewed Topos-Bind rather than presenting it as a completed, generally available product. The difference is important. Topos-2 has an external benchmark result, while Topos-Bind currently rests on a narrower company case study.

That case study uses amyloid-beta, a flexible peptide associated with Alzheimer’s disease. Instead of asking for one preferred complex, Topos-Bind estimates how each compound changes the distribution of amyloid-beta shapes.

The launch therefore joins two stages of a proposed discovery workflow. Topos-2 maps the moving target. Topos-Bind attempts to map how a candidate molecule changes that motion.

Dynamic Proteins Put Single-Structure Workflows Under Pressure

The launch pressures drug-discovery workflows built around a best-guess structure, not every protein prediction system equally.

Structure-based drug design often begins with a three-dimensional target. Researchers identify pockets, simulate interactions, and optimize molecules that fit favorable arrangements of atoms. This approach works best when the selected structure represents a biologically meaningful state.

IDPs complicate that process because their flexibility is not merely experimental noise. Their shifting structures can support signaling, transcription, molecular recognition, and the formation of biomolecular condensates inside cells.

Topos cites research indicating that three out of five human proteins contain disordered regions longer than 30 amino acids. Other estimates place intrinsically disordered content at roughly 30% to 40% of the human proteome.

The exact percentage changes with definitions and prediction methods. The underlying problem does not. A sizable share of disease-relevant protein biology cannot be represented adequately by one fixed structure.

AlphaFold showed how accurately AI can predict stable protein folds from sequence. That achievement created an enormous resource for biology, but it did not make protein motion irrelevant.

A 2025 Nature Communications study described disordered proteins as systems that should be represented by structural ensembles. The researchers combined AlphaFold-derived distances with molecular dynamics to generate such ensembles.

That work illustrates the real competition. The field is not divided neatly between static AlphaFold predictions and one Topos alternative. Multiple teams are extending structure prediction with simulations, experimental restraints, and generative sampling.

BioEmu also generates protein conformational ensembles. Other approaches refine AlphaFold outputs or use them as priors for molecular dynamics. These methods differ in training data, computational cost, physical assumptions, and dependence on experimental measurements.

Topos is making a narrower strategic bet. It focuses on the order-disorder spectrum and aims to connect ensemble generation directly with small-molecule design.

This focus places pressure on teams using docking or cofolding systems that return a compact set of confident structures. Such tools can still be valuable for stable complexes. They become harder to interpret when a target remains flexible before and after binding.

Topos compared Topos-Bind with Boltz-2, Chai-1, and OpenFold3 in its amyloid-beta study. It said their outputs clustered into relatively narrow groups of compact shapes.

The company attributes that result to the competing models’ objectives. Those systems generally seek confident structural predictions, while Topos-Bind explicitly seeks a distribution of states.

That comparison does not show that Topos-Bind is universally better. It shows that a model optimized for ensembles answers a different question from one optimized for a likely complex.

For drug developers, the forced response is methodological. Teams pursuing dynamic targets must decide whether a single predicted complex is enough, or whether ensemble behavior must guide compound selection.

The response will unfold over years rather than months. Discovery teams need to connect predicted conformations with target engagement, cellular effects, toxicity, and eventual clinical outcomes.

Why Protein Ensembles Change the Drug-Design Question

For a dynamic target, the useful question is not only where a molecule binds, but how binding redistributes the protein’s possible states.

Consider amyloid-beta in its monomeric form. The peptide moves through many conformations, and some states are more likely to associate with other molecules. Those associations can contribute to aggregation processes linked with Alzheimer’s pathology.

A traditional workflow might search for one binding pose against one selected amyloid-beta structure. An ensemble workflow asks whether a compound changes the probability of aggregation-prone states across a broader distribution.

Topos tested Topos-Bind using amyloid-beta 1-42 with 22 small molecules. The model’s predicted ensembles reproduced the broad spread of molecular sizes found in Topos-generated molecular dynamics simulations.

The comparison used radius of gyration, a measurement of how compact or extended a molecular structure is. A broader distribution indicates that the protein explores structures with meaningfully different dimensions.

Topos reported that the other evaluated cofolding models collapsed their predictions into narrower distributions. If that pattern holds experimentally, a single-state workflow could overlook compounds that alter the ensemble without creating one dominant structure.

However, the reference was molecular dynamics produced by Topos itself. Molecular dynamics uses physics-based calculations to simulate atomic motion over time, but its results depend on force fields, starting conditions, and sampling choices.

Physical experiments do not directly reveal a complete atomic ensemble. Researchers infer ensembles by combining measurements with computational models. That makes evaluation more complicated than comparing a predicted crystal structure with an experimental structure.

Topos trained its approach with Topos-DB, a proprietary collection containing more than 100,000 simulated disordered protein-ligand systems. The company says this is roughly three times larger than the next-largest comparable corpus by distinct system count.

The dataset is central to the model’s reported advantage. Public structural databases contain many stable protein structures, but far fewer atomic ensembles for disordered proteins interacting with small molecules.

Generating simulation data offers scale where experimental data remain scarce. It also introduces a dependency: the model can learn patterns and biases embedded in the simulation engine that created its training examples.

Topos attempts to counter that problem by testing Topos-2 against experimental NMR and scattering measurements. Its PeptoneBench result therefore carries more weight than a comparison based only on simulated trajectories.

Topos-Bind has not yet reached the same evidentiary level. Its amyloid-beta work asks whether the model resembles molecular dynamics, not whether its predicted ligand effects match experimental observations across many targets.

This distinction defines the central tension behind the launch. Topos has built a mechanism suited to dynamic biology, but the data needed to validate that mechanism are unusually difficult to obtain.

The company is developing experimental feedback loops alongside its computational work. In March, Topos announced a collaboration with Dalriada Drug Discovery covering mass spectrometry, chemoproteomics, and live-cell measurements.

Dalriada’s experimental program includes target-engagement profiling, peptide mapping, covalent-binding kinetics, and mechanism-of-action proteomics. These methods can test whether predicted interactions occur in biological systems.

That partnership reflects what the models cannot settle alone. An attractive ensemble remains a hypothesis until experiments show binding, selectivity, functional effects, and acceptable behavior in cells.

Topos-2 Enters a Crowded Protein Modeling Race

Topos is competing on specialization, while larger and better-funded platforms are expanding the range of protein states they can represent.

BioEmu is the most direct reference point for Topos-2 because it also generates structural ensembles. Topos used BioEmu as the next-best comparator in its PeptoneBench result.

The reported 1.11 score gives Topos-2 a clear advantage on the selected composite measure. Yet one benchmark cannot resolve every question relevant to drug discovery.

Performance can change across protein lengths, sequence families, environmental conditions, and different forms of disorder. Benchmark datasets can also overlap conceptually with the assumptions used during model development, even without direct training-set leakage.

AlphaFold-derived methods present another route. Researchers can use AlphaFold’s predicted distance information as restraints, then run molecular dynamics to build an ensemble consistent with those distances.

This hybrid strategy combines a widely tested structure predictor with established physical simulations. It can also incorporate experimental measurements when those data are available.

Its disadvantages include computational expense and additional modeling choices. Topos is betting that a model trained to generate ensembles directly will make the workflow faster and more scalable.

Boltz-2, Chai-1, and OpenFold3 occupy a neighboring category. They model biomolecular structures and interactions, including protein-ligand complexes, but are not primarily designed to reproduce full equilibrium distributions for highly disordered targets.

Topos-Bind’s amyloid-beta comparison therefore tests a deliberate mismatch in objectives. That is informative, but it is not a complete contest among equally specialized ensemble models.

The commercial race extends beyond model accuracy. A useful drug-discovery platform must identify molecules that can be synthesized, reach their targets, avoid harmful interactions, and produce therapeutic effects.

Topos joined that race with limited capital compared with major pharmaceutical companies and the largest AI laboratories. The San Francisco company raised a $10.5 million seed round in January 2026.

Its backers include Boldstart, Threshold, Neo, and individual investors. The financing supports an ambitious platform, but clinical drug development can consume far more capital and time than model training alone.

Topos is also participating in TuneLab, an initiative providing federated access to models trained on Eli Lilly’s proprietary drug-development data. Federated learning allows participants to benefit from shared models without directly pooling all underlying private records.

That relationship can strengthen predictions involving absorption, distribution, metabolism, excretion, and toxicity. Those properties often determine whether a promising molecule can become a viable medicine.

Still, neither funding nor data access resolves the central validation problem. Topos must show that better ensemble predictions improve decisions at later stages of discovery.

A modest improvement in computational ranking can be valuable if it prevents expensive experiments on weak candidates. A dramatic benchmark improvement can be less valuable if it does not change which compounds succeed.

The competitive question is therefore not whether Topos can generate visually convincing molecular motion. It is whether its specialization creates measurable advantages in hit discovery, lead optimization, and therapeutic development.

The Evidence Is Stronger for Topos-2 Than Topos-Bind

Topos-2 has an independent experimental benchmark, while Topos-Bind remains a promising preview supported mainly by simulation-based evidence.

PeptoneBench is a meaningful test because it spans ordered and disordered proteins. Its datasets include nuclear magnetic resonance chemical shifts and small-angle X-ray scattering measurements.

Topos says the benchmark includes NMR data for 659 proteins and SAXS data for 439 proteins. The evaluation compares properties calculated from generated ensembles with experimental observations rather than selecting one visually plausible structure.

That design aligns with the task Topos-2 claims to solve. It rewards distributions that reproduce measurements averaged across many changing molecular states.

However, Topos disclosed a reweighted composite result. Reweighting adjusts the relative contribution of generated structures to improve agreement with experimental data. This is a recognized ensemble-refinement technique, but it is not identical to fully prospective prediction.

Readers should therefore distinguish three levels of evidence. The first is generation from sequence alone. The second is agreement after refinement against experimental information. The third is prospective success on unseen biological systems.

The public announcement emphasizes the composite result but does not provide a peer-reviewed Topos-2 paper establishing performance across every relevant evaluation condition. Independent reproduction would strengthen the claim substantially.

Topos-Bind faces a larger evidence gap. Its reference ensembles come from the company’s molecular dynamics engine, and its training database also contains company-generated simulation data.

This arrangement does not invalidate the result. Simulations are indispensable for studying systems that evade direct structural measurement. It does mean the evaluation may favor assumptions shared by the model and reference process.

The amyloid-beta experiment also tests a limited set of 22 ligands. That is enough to illustrate model behavior, but not enough to establish general performance across disordered targets and chemical classes.

Disease biology adds another layer of uncertainty. Accurately predicting that a molecule shifts an amyloid-beta ensemble does not establish that the shift reduces harmful aggregation in a living system.

The molecule must reach the relevant tissue, engage the target, produce the intended functional change, and avoid unacceptable toxicity. Each step can break the connection between a computational prediction and a medicine.

Topos acknowledges the importance of experimental work through its internal programs and collaborations. Its Dalriada partnership specifically targets engagement, selectivity, and pathway-level effects.

The company says it is applying the models to internal programs in neurodegeneration and oncology. It has not disclosed a clinical candidate produced by Topos-2 or Topos-Bind.

That absence is normal for an early-stage biotechnology company. It also defines what cannot yet be claimed. The models have not proved that previously unreachable targets are now druggable.

“Undruggable” is also a moving label. It often describes targets that lack conventional binding pockets or resist established development methods. It does not mean that no biological or therapeutic strategy can ever affect them.

Topos has presented a plausible route into that territory. The route now needs prospective experiments that were not used to build, tune, or select the models.

What to Watch After the Topos Bio AI Models Launch

Three signals will show whether ensemble modeling becomes a practical discovery advantage rather than another strong computational demonstration.

The first signal is independent technical validation. Researchers should look for a detailed Topos-2 or Topos-Bind paper, disclosed evaluation settings, accessible outputs, and reproduction by groups outside the company.

Independent results near the reported PeptoneBench performance would strengthen Topos-2’s position. Large performance changes under alternative settings would weaken the claim that the advantage generalizes.

For Topos-Bind, the most useful validation would compare predicted ligand-induced ensembles with experimental measurements across several unrelated disordered proteins. One amyloid-beta case cannot establish broad reliability.

The second signal is prospective experimental success. Topos and Dalriada can test compounds selected with the models for target engagement, selectivity, and changes in relevant cellular pathways.

A blind test would be especially informative. Topos could rank compounds before experiments, then disclose whether the highest-ranked molecules outperform conventional selections.

Repeated enrichment of active compounds would strengthen the argument for ensemble-based discovery. Failure to improve experimental hit rates would suggest that accurate-looking motion does not yet guide better chemical decisions.

The third signal is pipeline progression. Topos should eventually identify specific programs, development milestones, or candidate molecules influenced by Topos-2 and Topos-Bind.

A declared development candidate would not prove clinical value, but it would show that the platform supports decisions beyond retrospective benchmarks. Later toxicology and human data would provide progressively stronger tests.

Competitor responses also matter within these signals. BioEmu successors, AlphaFold-based ensemble methods, and specialized simulation systems will continue improving.

Topos may hold an early advantage on disorder-focused benchmarks without retaining a durable technical lead. Its defensibility could instead come from proprietary simulation data, experimental feedback, and experience selecting compounds.

For AI product teams, this launch offers a broader lesson about evaluation. A model’s objective, reference data, and real-world validation must match the decision it is supposed to improve.

Teams tracking evidence across model releases, experiments, and partnerships need more than a collection of press releases. A searchable AI knowledge base can help connect benchmark claims with later validation.

Topos has identified a genuine weakness in single-structure thinking. Dynamic proteins need representations that preserve motion instead of compressing it into one confident answer.

The open question is no longer whether protein ensembles matter. It is whether the Topos Bio AI models predict them accurately enough to change which compounds researchers build, test, and advance. Watch for independent benchmarks, blind experimental results, and a disclosed candidate before treating that question as settled.