Highlights
- Platform: BIOPTIC B1, a SMILES-based transformer pre-trained on 160M+ molecules from PubChem and Enamine REAL Space, fine-tuned on experimental protein-ligand binding affinities from BindingDB
- Library searched: Enamine REAL Space s_ subspace, 3.2 billion make-on-demand small molecules
- Input required: SMILES of a single known Miro1 ligand (MR3); no 3D protein structure data
- Candidates selected: 62 chemically novel compounds (Tanimoto similarity < 0.4 vs. MR3); 53 synthesized at >90% purity by Enamine Ltd.
- Neuroprotection hit rate: 14/53 (26.4%) enhanced neuronal viability under oxidative stress in FA patient-derived sensory neurons
- Top hits: WO24 and WO17 matched or exceeded MR3 neuroprotection at 1–5 μM
- Thermal shift confirmation: 13/53 compounds produced a ≥1°C Tm increase; WO24 showed dose-dependent binding to the Miro1 C-terminal GTPase domain
- Patient line validation: WO17 and WO24 confirmed across three independent Friedreich's ataxia patient lines from distinct donors
- IP: Patent filed jointly by Stanford University and Optic Inc.
Methods
BIOPTIC B1 is a SMILES-based transformer pre-trained on more than 160 million molecules from PubChem and the Enamine REAL Space, then trained on all experimental protein-ligand binding affinities from BindingDB. The system requires only the SMILES of a single known ligand as input, bypassing the 3D structural data that conventional structure-based virtual screening depends on.
For this study, the SMILES of MR3, the sole published Miro1 ligand at the time of screening, served as the query input. BIOPTIC B1 searched the s_ subspace of Enamine REAL Space (3.2 billion compounds, selected for >80% reported synthesis success rate). The top 30,000 ranked candidates were filtered for chemical liabilities via REOS, novelty (ECFP4 Tanimoto similarity < 0.4 vs. MR3), and CNS drug-likeness (MW ≤ 400, LogP ≤ 5, HBD ≤ 3, HBA ≤ 7). The resulting set was clustered into 150 groups via k-means, and 62 exemplars were assessed for synthesis feasibility; 53 were produced at >90% purity.
Experimental evaluation proceeded through two orthogonal pipelines: an MTT-based high-throughput cell viability assay in Antimycin A-stressed FA patient-derived sensory neurons, and a thermal shift assay (TSA) using purified truncated human Miro1 C-terminus (residues 408–592). In benchmarking across seven diverse protein targets, BIOPTIC B1 showed predictive performance comparable to the state-of-the-art ligand-based method ChemProp, while outperforming ECFP4 fingerprint similarity and Gradient Boosting.
Key Results
Fourteen of 53 compounds (26.4%) enhanced neuronal viability under oxidative stress in the MTT assay. Thirteen (24.5%) produced a ≥1°C Tm increase in the TSA. Two compounds, WO24 and WO17, ranked as top hits across both pipelines. Re-tested at lower concentrations (1 and 5 μM, below the initial 10 μM screen), both matched or exceeded MR3 neuroprotection, confirmed by TUNEL staining and replicated across three independent FA patient lines from distinct donors.
WO24 showed dose-dependent Tm increases in the TSA, confirming direct and specific binding to the Miro1 C-terminal GTPase domain. WO17 increased Miro1 iron binding in ICP-MS assays, suggesting potential iron-chelating activity that warrants further investigation. Neither compound affected Miro1's intrinsic GTPase activity in fibroblasts, supporting the druggability of this domain without disrupting mitochondrial transport.
Scientific Rigor
- Hits evaluated through two independent assays (MTT cell viability and thermal shift)
- WO24 binding confirmed by dose-response TSA
- Neuroprotection independently confirmed by TUNEL staining
- Efficacy replicated across three FA patient lines from distinct donors
- Compounds confirmed not to inhibit Miro1 GTPase activity or mitochondrial transport in healthy neurons
- BIOPTIC B1 benchmarked against established ligand-based methods (ECFP4 fingerprint, Gradient Boosting, ChemProp)
- This is the second peer-reviewed publication demonstrating BIOPTIC B1 in a drug discovery application, following earlier published work on the LRRK2 Parkinson's disease target
Links & Availability
Full paper (Cell Chemical Biology): https://www.cell.com/cell-chemical-biology/fulltext/S2451-9456(26)00156-X
Proteomic data browser (MiroScape): https://miroscape.github.io/MiroScape/#/miroProteome/FA
Raw proteomic data (ProteomeXchange via PRIDE): http://doi.org/10.6019/PXD070711
Citation
Chandra S, Kwak CS, Du Z, Barisano G, Nguyen KT, Vinogradov V, Wang X, et al. Chemical modulation of Miro1 alleviates cell-type-specific vulnerabilities in Friedreich's ataxia. Cell Chemical Biology, 2026. https://www.cell.com/cell-chemical-biology/fulltext/S2451-9456(26)00156-X
Authors & Acknowledgements
Kong T. Nguyen and Vlad Vinogradov (Optic Inc.) performed the AI-guided virtual screen and managed chemical synthesis. Andrey Doronichev (Optic Inc.) provided technical support. The study was led by Sujyoti Chandra and supervised by Xinnan Wang at Stanford University's Department of Neurosurgery. A patent on the identified compounds has been filed jointly by Stanford University and Optic Inc.
