Topos-Chem and Topos-AP: designing past the edge of known chemistry

By Connor Bybee, Research Scientist

Chemical exploration with generative modeling

Movie 1. Topos-Chem generating new analogs of six literature kinase inhibitors (osimertinib, afatinib, and lapatinib with EGFR; ibrutinib with BTK; and imatinib and dasatinib with ABL1). For each drug, one group (orange) is removed and regenerated by Topos-Chem while some parts of the molecule (aqua) are fixed. Each segment first replays one generation, then transitions to two separate analogs generated with a different number of atoms.

In a single iteration, a medicinal chemist working with Topos Bio’s generative chemistry platform produced the two most potent new compounds in a live kinase program, both far from published chemistry. Topos-AP, the affinity model that scored them, had the highest observed binder AUROC among five models on that program, with its clearest lead on the most novel and covalent compounds.

Key results

Novel and potent designs. The two most potent new compounds in our program were designed by a medicinal chemist using Topos-Chem and Topos-AP. With binding affinities of 1.3 nM and 2.5 nM (pKi 8.9 and 8.6), neither compound has a Tanimoto similarity above 0.38 to any published compound with Ki below 1 µM (Fig. 2).
Highest observed accuracy on the most novel chemistry. For candidates least similar to reported structures, Topos-AP scores an area under the receiver operating characteristic curve (AUROC) of 0.82, against 0.73 for the top baseline, OpenFold3 with AQAffinity.
Highest observed accuracy on covalent compounds. Among compounds with a reactive warhead, Topos-AP is the best predictor of the five methods evaluated (0.73, against 0.41 for OpenFold3 with AQAffinity; Fig. 1).

Topos-AP is the most accurate predictor on the program’s compounds

Topos-AP
Baselines
Chance (0.5)

Fig. 1. Binder AUROC for Topos-AP and four established models (OpenFold3 with AQAffinity, Boltz-2, AutoDock Vina, and FLOWR.root) on reversible compounds, covalent compounds, and all measurements.

Generative chemistry needs affinity prediction that generalizes

A generative chemistry model will propose more new molecules than can be economically synthesized and tested. Choosing which to make therefore relies on an in silico method. This choice is hard in the areas where generative chemistry is most valuable, which includes understudied targets implicated in disease such as disordered proteins, diverse chemical modalities such as covalent modifiers, and unique, patentable molecules. Our targets, including protein kinases, are dynamic and contain disordered regions, which affect ligand binding and may call for different drug modalities. See our companion post for how kinase dynamics affect binding affinity. Our generative chemistry platform has two parts that work together.

Topos-Chem is the generative part. It explores chemical space beyond known chemistry and optimizes molecules for binding affinity to the therapeutic target, selectivity against off-targets, drug-likeness, and ADMET (absorption, distribution, metabolism, excretion, and toxicity). Movie 1 shows Topos-Chem regenerating one group of each of six literature kinase inhibitors.
Topos-AP is our affinity-prediction platform. It scores how strongly each molecule binds the therapeutic target, and these scores are used to rank which molecules to make and test.

Novel and potent compounds in a single iteration

The design series is potent and novel

Potency rises with similarity to the nearest published binder: Spearman ρ = 0.56 over 41 unpublished compounds (p = 0.00014). Published compounds are shown in their own column.

Unpublished compound
Published binder
Parent
Product
Parents combined into products

Fig. 2. Each point is one reversible binding compound with a measured pKi for the target kinase. The x axis is the Tanimoto similarity to the nearest published compound measured with a pKi > 6 (Ki < 1 µM); the published binders are shown in their own column. A medicinal chemist combined features of the parents into two novel products, Product 1 and Product 2.

Fig. 2 shows binding data derived from designing reversible inhibitors for an internal kinase program. In the first part of the discovery campaign, Topos-Chem and Topos-AP were used to generate a large and diverse set of designs for the target. We selected and tested 41 compounds.

Assays identified 7 active binders, defined as molecules with verified Ki values of 20 µM or lower (pKi 4.7 or higher). Two of them are Parent 1 (Ki 5 µM, pKi 5.3) and Parent 2 (Ki 16 nM, pKi 7.8). Then, with these results, a medicinal chemist designed two novel and potent chemical derivatives, Product 1 and Product 2, which have Ki values of 1.3 nM and 2.5 nM (pKi 8.9 and 8.6), respectively. They are the two most potent compounds in the program. Product 1 is more potent than 11 of the 13 published compounds measured in the program’s assay.

Topos-AP ranks binders best on the program’s compounds

Designing high-affinity and novel binders is a challenging problem when guided by machine learning affinity prediction models trained on binding data. Data leakage can inflate benchmark accuracy¹. Models that rely on chemical similarity may not learn key features of binding interactions and therefore fail to generalize. We therefore tested every model on the program’s compounds, most of which are unpublished.

We scored Topos-AP and four established models, OpenFold3 with AQAffinity²^,³, Boltz-2⁴, AutoDock Vina⁵, and FLOWR.root⁶, on every compound the program has measured. This covers 235 measurements of 105 reversible compounds and 51 measurements of 28 covalent compounds. Topos-AP was trained only on public data, after controlling for data leakage by removing every record that matched a test compound, the target protein, or a close analog. None of the internal program data was used in model training or validation.

Topos-AP has the highest AUROC on reversible compounds (0.77), on covalent compounds (0.73), and on all measurements together (0.75; Fig. 1). OpenFold3 with AQAffinity is the closest baseline, at 0.74, 0.41, and 0.72, respectively. The largest advantage is on covalent compounds.

To test generalization, we split the reversible compounds into four bins by their maximum Tanimoto similarity to public compounds measured on the target. The bins are below 0.30, 0.30–0.40, above 0.40, and identical to a published compound. In the least similar bin (21 compounds, 37 measurements), Topos-AP’s AUROC is 0.82, against 0.73 for OpenFold3 with AQAffinity, 0.63 for AutoDock Vina, 0.61 for Boltz-2, and 0.57 for FLOWR.root.

Designing covalent compounds

At Topos Bio, we are interested in different modalities for drugging disordered proteins. Designing covalent binders has several therapeutic advantages⁷. Topos-Chem supports generation of covalent candidates by positioning a reactive warhead near the specified target residue. Movie 2 shows an example of generating covalent compounds with the covalent warhead held fixed close to the target residue. Covalent drug design depends on both reversible affinity and the covalent reaction. Across the 28 covalent compounds (51 measurements, 12 binders), Topos-AP’s AUROC is 0.73 against 0.41 for OpenFold3 with AQAffinity (Fig. 1).

Growing a molecule from a warhead

Movie 2. Completion from a warhead. Only osimertinib’s acrylamide (aqua) is kept in its crystal position next to Cys797 of EGFR (PDB 4ZAU), and a generative model builds the other 32 heavy atoms of a new molecule (orange). As the molecule takes shape, a dotted line marks its hydrogen bond to the hinge residue Met793; pocket regions and the drawn side chains are labeled.

Conclusion

These results demonstrate the performance of the Topos Bio drug discovery platform in a real-world drug discovery campaign. In one kinase program, a medicinal chemist working from Topos-Chem’s designs, with Topos-AP among the ranking inputs, made the two most potent new compounds, both far from published chemistry. Across every compound the program measured, Topos-AP, trained only on public data and molecular-dynamics ensembles, had the highest observed binder AUROC of five models, with its largest observed lead on the least familiar chemistry and covalent compounds.

Supporting information

A measurement counts as a binder when its pKi is 4.7 or higher (Ki of 20 µM or lower). We score each model by binder AUROC⁸.

Models

Model	Description	Score used for ranking
Topos-AP	Our affinity-prediction platform, combining structure-based and structure-free components, trained on molecular-dynamics ensembles of a leakage-guarded kinase panel	Binder probability
OpenFold3 with AQAffinity²^,³	OpenFold3 co-folding with SandboxAQ’s AQAffinity affinity head, used as released	Predicted pIC₅₀, mean of 5 seeds
Boltz-2⁴	Co-folding model with an affinity head, used as released	Binder probability
AutoDock Vina⁵	Physics-based docking into each compound’s own co-folded complex after 1 ns of molecular dynamics and minimization	Docking score, negated so higher is better
FLOWR.root⁶	Flow-matching model for structure-aware ligand generation and affinity prediction	Predicted pKd (it has no probability head)

No model used a molecular-weight correction. Boltz-2’s correction does not change the binder probability ranking, only its predicted affinity value; FLOWR.root ran with its correction switched off; and our OpenFold3 with AQAffinity runs had none. Vina is ranked by its raw docking score. Fig. S1 shows the ROC curves behind the AUROC values in Fig. 1.

ROC curves, reversible compounds

Topos-AP · AUROC 0.768
OpenFold3 + AQAffinity · AUROC 0.744
FLOWR.root · AUROC 0.665
AutoDock Vina · AUROC 0.671
Boltz-2 · AUROC 0.694
Chance

Fig. S1. ROC curves pooled over all 235 reversible measurements. Each curve plots the fraction of binders ranked above a threshold (true-positive rate) against the fraction of non-binders ranked above it (false-positive rate) as the threshold moves down that model’s ranking. The area under a curve is that model’s AUROC.

Leakage audit

Fig. S2 shows how far the test compounds are from each training set. No reversible or covalent test compound is closer than 0.48 to any Topos-AP training ligand. AQAffinity, OpenFold3’s affinity head, has seen more of this chemistry, with 14% of the reversible compounds in its released training data, all measured on other proteins. Separately, those data include records for the target itself.

Maximum similarity to the training set

Training ligands (Topos-AP)
Public chemistry on the target
OpenFold3 affinity training set (AQAffinity), any protein

Fig. S2. For each of the 105 reversible test compounds, the maximum Tanimoto similarity to each training set, plotted as a cumulative distribution in the style of Runs N’ Poses⁹. The further left and steeper a curve rises, the more novel the test set is against that training set. For Topos-AP, no test compound is closer than 0.48 to any training ligand, and the same holds for the covalent compounds. Against public chemistry for the target, 18% of compounds are themselves published, and that step is drawn at 1.0. The third curve is AQAffinity’s released training and validation data¹⁰ on any protein. It contains 14% of the compounds, all measured on other proteins, and the median is 0.54.

Contributions

Conceptualization. Connor Bybee, Robert A. Galemmo, and Fernando Martinez.
Data curation. Connor Bybee, Jiaqi Zhu, and Anthony M. Giannetti.
Formal analysis. Connor Bybee and Anthony M. Giannetti.
Investigation. Connor Bybee, Jiaqi Zhu, Xiaotong Lu, Patrick Visperas, Benjamin Rousseau, Malhar Kute, and Anthony M. Giannetti.
Methodology. Connor Bybee, Xiaotong Lu, Patrick Visperas, and Anthony M. Giannetti.
Project administration. Connor Bybee.
Resources. Connor Bybee and Jiaqi Zhu.
Software. Connor Bybee, Andre Graubner, Tomas Salgado, Scott Leishman, and Anthony M. Giannetti.
Supervision. Anthony M. Giannetti.
Validation. Anthony M. Giannetti.
Visualization. Connor Bybee and Jiaqi Zhu.
Writing – original draft. Connor Bybee.
Writing – review & editing. Jiaqi Zhu, Robert A. Galemmo, Anne Pipathsouk, and Anthony M. Giannetti.

References

Mattsson, B. & Walters, W. P. Identifying and addressing systematic data leakage in protein–ligand affinity benchmarks. bioRxiv (2026). ↩
The OpenFold3 Team. OpenFold3-preview, v0.3.1. Zenodo (2025). ↩ ↩²
SandboxAQ. AQAffinity: open-source structure-free affinity on OpenFold3. SandboxAQ Blog (28 July 2026). ↩ ↩²
Passaro, S. et al. Boltz-2: Towards accurate and efficient binding affinity prediction. bioRxiv (2025). ↩ ↩²
Eberhardt, J., Santos-Martins, D., Tillack, A. F. & Forli, S. AutoDock Vina 1.2.0: new docking methods, expanded force field, and Python bindings. J. Chem. Inf. Model. 61, 3891–3898 (2021). ↩ ↩²
Cremer, J., Le, T., Ghahremanpour, M. M., Sługocka, E., Menezes, F. & Clevert, D.-A. FLOWR.root – a flow matching-based foundation model for joint multi-purpose structure-aware 3D ligand generation and affinity prediction. Nat. Commun. 17, 5883 (2026). ↩ ↩²
Singh, J., Petter, R. C., Baillie, T. A. & Whitty, A. The resurgence of covalent drugs. Nat. Rev. Drug Discov. 10, 307–317 (2011). ↩
Hanley, J. A. & McNeil, B. J. The meaning and use of the area under a receiver operating characteristic (ROC) curve. Radiology 143, 29–36 (1982). ↩
Škrinjar, P., Eberhardt, J., Studer, G., Tauriello, G., Schwede, T. & Durairaj, J. Evaluating generalization in protein–ligand cofolding methods. Nat. Struct. Mol. Biol. 33, 782–794 (2026). ↩
SandboxAQ. AQAffinity model card, with its training and validation data. Hugging Face (2026). ↩