Guardado en:
| Autores principales: | , , , |
|---|---|
| Formato: | Preprint |
| Publicado: |
2025
|
| Materias: | |
| Acceso en línea: | https://arxiv.org/abs/2511.06930 |
| Etiquetas: |
Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
|
| _version_ | 1866911257808863232 |
|---|---|
| author | Khan, ShahZeb Pallara, Chiara Monti, Barbara Molina, Alexis |
| author_facet | Khan, ShahZeb Pallara, Chiara Monti, Barbara Molina, Alexis |
| contents | Alzheimers disease (AD), a progressive neuro-degenerative disorder, currently lacks effective therapeutic strategies that can modify disease progression. Recent studies have highlighted the circadian rhythm critical role in AD pathophysiology, implicating circadian clock kinases, such as the Salt-Inducible Kinase 3 (SIK3), as promising therapeutic target. Generative AI models have surpassed traditional methods of drug discovery, untapping the vast unexplored chemical space of drug-like molecules. We present a sequence-to-sequence Variational Autoencoder (Seq2Seq-VAE) model guided by an Active Learning (AL) approach to optimize molecular generation. Our pipeline iteratively guided a pre-trained Seq2Seq-VAE model towards the pharmacological landscape relevant to SIK3 using a two-step framework, an inner loop that iteratively improves physiochemical properties profile, drug likeliness and synthesizability, followed by an outer loop that steer the latent space towards high-affinity ligands for SIK3. Our approach introduces feedback-driven optimization without requiring large labeled datasets, making it particularly suited for early-stage drug discovery in under-explored therapeutic targets. Our results demonstrate the models convergence toward SIK3-specific small molecules with desired properties and high binding affinity. This work highlights the use of generative AI combined with AL for rational drug discovery that can be extended to other protein targets with minimal modifications, offering a scalable solution to the molecular design bottleneck in drug design. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2511_06930 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | De Novo Design of SIK3 Inhibitors via Feedback-Driven Fine-Tuning of Seq2Seq-VAE Khan, ShahZeb Pallara, Chiara Monti, Barbara Molina, Alexis Biomolecules Alzheimers disease (AD), a progressive neuro-degenerative disorder, currently lacks effective therapeutic strategies that can modify disease progression. Recent studies have highlighted the circadian rhythm critical role in AD pathophysiology, implicating circadian clock kinases, such as the Salt-Inducible Kinase 3 (SIK3), as promising therapeutic target. Generative AI models have surpassed traditional methods of drug discovery, untapping the vast unexplored chemical space of drug-like molecules. We present a sequence-to-sequence Variational Autoencoder (Seq2Seq-VAE) model guided by an Active Learning (AL) approach to optimize molecular generation. Our pipeline iteratively guided a pre-trained Seq2Seq-VAE model towards the pharmacological landscape relevant to SIK3 using a two-step framework, an inner loop that iteratively improves physiochemical properties profile, drug likeliness and synthesizability, followed by an outer loop that steer the latent space towards high-affinity ligands for SIK3. Our approach introduces feedback-driven optimization without requiring large labeled datasets, making it particularly suited for early-stage drug discovery in under-explored therapeutic targets. Our results demonstrate the models convergence toward SIK3-specific small molecules with desired properties and high binding affinity. This work highlights the use of generative AI combined with AL for rational drug discovery that can be extended to other protein targets with minimal modifications, offering a scalable solution to the molecular design bottleneck in drug design. |
| title | De Novo Design of SIK3 Inhibitors via Feedback-Driven Fine-Tuning of Seq2Seq-VAE |
| topic | Biomolecules |
| url | https://arxiv.org/abs/2511.06930 |