Topos Bio ships Topos-2 and Topos-Bind for “undruggable” proteins

San Francisco AI biotech Topos Bio said Wednesday it is releasing Topos-2, a next-generation model that generates protein shape ensembles across the order–disorder spectrum, and previewing Topos-Bind, which it calls the first generative model for disordered proteins bound to small molecules.

Most human proteins have floppy, disordered stretches that refuse a single crystal structure, which is why drug designers long treated them as undruggable. Topos-2 and Topos-Bind bet that generating the full ensemble, not one best fold, can put those moving targets on the map for small-molecule design.

On Wednesday, 30 September 2026, Topos Bio announced Topos-2 and previewed Topos-Bind. The company blog, dated that day, is bylined Tomas Salgado, Andre Graubner, and Malhar Kute. It says Topos is releasing Topos-2, a next-generation model for the full range of shapes a protein can take, across what it calls the order–disorder spectrum, and previewing Topos-Bind. Business Wire, as Financial Content carries it, stamps the item September 30, 2026, at 9:00 a.m. Eastern. The wire calls Topos Bio an AI-powered biotechnology company and says it introduced two AI foundation models. A foundation model, here, is a large model trained so it can be reused on many related problems, rather than a model built for one protein. Those lines are Topos Bio’s.

What the company means by disordered. Many proteins do not sit still in one shape. The floppy stretches are called intrinsically disordered regions. A whole protein that stays floppy is an intrinsically disordered protein, which the wire shortens to IDP. The blog says three out of five human proteins contain disordered regions longer than 30 residues. A residue is one amino acid, one link in the chain. Thirty links is a short stretch, not the whole protein. The blog cites a 2026 paper in Proceedings of the National Academy of Sciences for that three-out-of-five line. The wire’s phrase is close to two-thirds of all proteins in humans. Both counts are Topos Bio’s. The blog says these regions matter in signaling, in the control of how genes are read, and in the droplets cells use to gather molecules, and that they show up in cancer, neurodegenerative disease, and metabolic disorders. Because they do not hold one stable shape, the kind a crystal structure would freeze, structure-based drug design has had little to grip. The company says that is why they have largely been called undruggable. The wire says the same class drives aggressive cancers and neurodegenerative diseases including Alzheimer’s, Parkinson’s, ALS, and Huntington’s, and that Topos puts the commercial opportunity at more than $200 billion a year. That dollar figure is Topos Bio’s. The posts do not name a market-research firm behind it.

Topos-2, as the blog describes it, is an all-atom generative model. All-atom means it places the atoms, rather than a simplified stand-in for each amino acid. Generative means it produces possible shapes, rather than only scoring a shape someone else drew. It reads an amino acid sequence, the order of building blocks, and generates a conformational ensemble: the set of shapes the chain can take, and how common each one is. Topos-1, the company says, modeled disordered proteins and disordered regions. Topos-2 expands that design across the full order–disorder spectrum, from folded proteins to floppy ones and the proteins that are both. The blog says the largest gains over existing models are on partially and fully disordered proteins. The wire says Topos-2 extends Topos-1 to proteins that have both disordered and ordered regions, closer to how these proteins work in the body. Those lines are Topos Bio’s.

The score Topos prints is on PeptoneBench. The blog calls it an independent benchmark of NMR chemical-shift measurements on 659 proteins and SAXS data on 439 proteins, across that same spectrum. NMR, nuclear magnetic resonance, reads how atoms sit in a magnetic field. The reading shifts when the local shape changes. SAXS, small-angle X-ray scattering, reads how compact or spread out a protein is in solution. The blog says a score near 1.0 means the ensemble agrees with the experiment within the benchmark’s estimated uncertainty, which mixes the error in the measurement and the error in the math that turns a shape into a predicted measurement. Lower is better. Topos says Topos-2 scored 1.11 on the reweighted composite score, and that this is more than four times closer to 1.0 than the next-best model, BioEmu, at 1.49. The gap from 1.0 is 0.11 for Topos-2 and 0.49 for BioEmu. The larger gap is a bit more than four times the smaller one. The wire says Topos-2 ranked first among over 10 evaluated models on agreement with NMR chemical shifts and SAXS, and that early testing put it nearly four times closer to experimental ground truth than the next-best model. The blog’s wording is more than four times. The wire’s wording is nearly four times. Both are Topos Bio’s. The wire names Microsoft Research’s BioEmu, Google DeepMind’s AlphaFold-2, and Boltz’s Boltz-2 among the models it says Topos-2 beat. The blog says the baseline scores on the chart are as reported by Invernizzi and colleagues in a 2025 preprint. A preprint is a paper posted before a journal’s outside review. Topos calls the benchmark independent. It does not say a journal reviewed Wednesday’s ranking. It also says the benchmark sequences were held out of training, including any sequence that lined up with 80 percent or more of a benchmark sequence at 30 percent identity or higher, so the test proteins were not memorized copies.

A supplemental note on the same post says Topos-2 scores best both before and after reweighting, and that it improves more than the other models when the weights change, from 2.05 down to 1.11. Reweighting, Topos explains, cannot add a shape the model never produced. It can only change how often each shape is counted. A large drop, the company says, suggests Topos-2 was already generating shapes consistent with the experiment, and that much of the remaining error was in how often each shape appeared. That 2.05 to 1.11 line is on the blog. The wire does not print it.

Topos-Bind, on the blog, is a preview. The blog calls it an all-atom generative model that outputs the ensemble of a protein bound to a small molecule, starting from a protein sequence and a SMILES string. SMILES is a line of text that spells a small molecule the way a sequence spells a protein. Topos says, to its knowledge, this is the first generative model built to capture the range of shapes for disordered proteins bound to small molecules. The wire calls Topos-Bind first of its kind for how these dynamic proteins interact with potential drug molecules, and says no other existing model does this today. Both phrasings are Topos Bio’s, including the limit “to our knowledge” on the blog.

The case study is amyloid-beta, a peptide tied to Alzheimer’s disease. A peptide is a short protein. The blog says Topos tested Topos-Bind on amyloid-beta bound to 22 different small molecules. It says the peptide is highly flexible, moving among many shapes, and that some of those shapes are more likely to stick to other copies and seed aggregates. The therapeutic goal, as the blog states it, is to shift the ensemble so those aggregation-prone shapes are less common, rather than to hit one frozen structure. The reference Topos uses is all-atom molecular dynamics from its own simulation engine. Molecular dynamics, shortened to MD, is a physics simulation of how the atoms move. Topos says physical experiments do not directly resolve the ensemble, so the simulation is the stand-in. Across the 22 ligands, a ligand being the small molecule, Topos says Topos-Bind reproduces the broad spread of shapes in that simulation, while the other evaluated cofolding models collapse the peptide into a much narrower range. A cofolding model proposes a shape for a protein and a molecule together, usually one confident shape. The blog names Boltz-2, Chai-1, and OpenFold3, and says they are built to output one best-guess shape, which leaves them poorly suited to a chain that will not sit still. The wire says Topos-Bind was the only evaluated model able to predict how amyloid-beta’s shape changes in response to drugs. Those lines are Topos Bio’s. The posts do not say the model treats Alzheimer’s, and they do not say a drug from this test has entered a trial.

The blog’s footnote puts numbers on that spread, for amyloid-beta 1-42, the 42-amino-acid form. Radius of gyration is how spread out the chain is. An angstrom is a tenth of a nanometer, about the length of a chemical bond. Topos says the simulation’s mean radius is 15.9 angstroms, with a standard deviation of 3.2. Standard deviation, here, is how wide the spread is. Topos-Bind’s mean is 16.1 angstroms, with a standard deviation of 4.0. The cofolding models, Topos says, sit at a mean of 10 to 12 angstroms, with a standard deviation of 0.5 to 1.0, a tighter and more compact cluster. Topos says it knows of no experimental radius measurement for ligand-bound amyloid-beta 1-42, which is why the simulation is the reference. Measurements of the peptide with no ligand, it says, are limited and vary widely, though Topos-Bind’s ensembles cover the reported values. Those figures are Topos Bio’s comparison with its own simulations. They are not a hospital result.

The training claim is two datasets, and the posts do not treat them as one number. The wire says Topos assembled the largest experimental datasets of small-molecule binding measurements against intrinsically disordered proteins, at least two orders of magnitude larger than previously reported datasets, along with the largest physics-based corpus of IDR-ligand systems. An IDR is an intrinsically disordered region. Two orders of magnitude is a factor of about a hundred, as a floor of “at least.” The blog describes Topos-DB, more than 100,000 disordered protein–ligand systems simulated with all-atom molecular dynamics on Topos’s physics engine. Topos says, to its knowledge, that is the largest all-atom protein molecular-dynamics dataset by the count of distinct systems, at roughly three times the next largest, and the first such set to simulate disordered regions with small molecules at that scale. The experimental binding sets are the wire’s claim. The 100,000 simulated systems are the blog’s. The posts do not print the earlier dataset’s name beside the “two orders” line.

Ryan Zarcone, chief executive and co-founder, said the hardest problems in biology are not hidden for lack of computing power. They are hidden, he said, for lack of the right data and models. He said Topos is built on a premise: if the company can understand how these proteins move and interact, it can design medicines for diseases that have stayed beyond modern drug discovery, and change the trajectory of patients’ lives. That quotation is his, in the wire. Amir Khosrowshahi, chief technology officer and co-founder, said most AI protein models predict one best structure per sequence, while a disordered protein has no single shape. A single-structure model, he said, is answering the wrong question. He said Topos’s models generate the full ensemble a protein can adopt, and then learn how a drug acts across that shifting landscape, not only where it binds in one snapshot, but how it changes the protein’s behavior. That quotation is his, in the wire.

Steven Finkbeiner, M.D., Ph.D., director of the Center for Systems and Therapeutics at Gladstone Institutes, is quoted as a scientific collaborator with Topos Bio. He said many proteins most relevant to neurodegenerative disease are among the hardest to study because they do not stay in one shape. He said the field has lacked enough experimental data to connect that motion to molecular interactions. He said combining large-scale data generation with AI models designed for dynamic proteins could help close that gap, and advance the understanding of disease biology and the search for new therapeutics. That quotation is his, in the wire. The release calls him a collaborator. It does not print a contract, a dollar figure, or a trial.

Where the models are going, as the posts state it. The wire says Topos-2 and Topos-Bind are being applied across Topos Bio’s internal discovery programs in neurodegeneration and oncology. Oncology is cancer. The blog says the platform is meant to generate ensembles, model them with small molecules, and design compounds that shift proteins away from disease-driving states, and that Topos-2 and Topos-Bind are steps toward doing that at scale. It says the company is applying the platform, across collaborations and internal programs, to targets long called undruggable. The about box says Topos is headquartered in San Francisco, that the platform integrates physics with generative AI, and that the company is backed by Boldstart, Threshold, Neo, and notable angel investors. The posts do not name those angels, and they do not print a valuation.

The picture is Topos Bio’s Fig. 1 from the Wednesday post, the PeptoneBench bar chart. Topos-2 is the short highlighted bar, closest to the line at 1.0. The other bars, in order, are BioEmu, PepTron, Boltz-1x, AlphaFold2, Boltz-2, and ESMFold. The axis is the reweighted composite score, lower is better, and it starts at 1.0. It is the company’s chart. It does not print a date.

In plain terms, Topos Bio said on Wednesday that Topos-2 generates the range of shapes a protein can take, including the floppy ones drug designers have skipped, and that on the company’s reading of PeptoneBench it sits closer to NMR and SAXS experiment than BioEmu, AlphaFold-2, and Boltz-2. Topos-Bind is a preview for the same kind of ensemble when a small molecule is bound. The 1.11 and 1.49 scores, the amyloid-beta comparison with 22 molecules, and the training-set claims are Topos Bio’s. The company does not say a drug from this work is approved, or that the ranking was reviewed by a journal.