Shengyao Liang @StellarCodeBio

(Thread 2/2) Our first #StellarCodeBio open-source project: A JAK2 pIC50 predictor (XGBoost, RDKit, Test R² ~0.68). Aiming to aid #DrugDiscovery. ChemRxiv: doi.org/10.26434/chemr… Feedback welcome! #OpenSourceAI #Cheminformatics

🚀 Launching #StellarCodeBio! My vision: next-gen AI learning "essence" – data-efficient, causal, neuro-symbolic & SNNs. Building "Research Models" for science. Hi @sama , thoughts on AI beyond current paradigms? GitHub: github.com/ShengyaoLiang/… #AIforScience #FutureofAI

手头有4个manus的邀请码，回复“求邀请”并转发，随机抽取4位朋友私信赠送，快来参与吧～ #manus #邀请码

Aspiring Innovator & Future Founder exploring the intersection of AI, Biotechnology, and complex systems. Driven high school student with a passion for deep learning and building impactful technology. Currently honing foundational knowledge and technical skills.

🔗 Learn more: For detailed insights, visit my LinkedIn post: 👉 linkedin.com/posts/shengyao… #AI #DrugDiscovery #Biotech #Collaboration (9/9)

Let’s collaborate! 🤝 Our AI-driven model can significantly enhance your drug discovery efficiency and accuracy. 📩 Contact us at: [email protected] (8/9)

6. High Efficiency & Seamless Integration Optimized with GPU acceleration and batch processing, our model screens millions of compounds in hours. ⚙️ It integrates with virtual screening, molecular generation, and ADMET prediction pipelines. (7/9)

5. Robust Generalization Capability Trained on a large-scale ChEMBL dataset, our model generalizes well across diverse targets. 🔹 It maintains accuracy even with novel scaffolds or low-data scenarios, making it highly adaptable. (6/9)

4. Multi-Scale Feature Fusion We integrate atomic, local, and global molecular features through a multi-scale fusion mechanism. 🧬 This enables the model to capture activity-determining factors from fragments to full conformations. (5/9)

3. Reliable Predictions with Quantified Uncertainty Our model uses Bayesian estimation to quantify prediction uncertainty. ✅ This provides confidence intervals, helping prioritize high-confidence hits and reducing costly false positives. (4/9)

2. Spatial-Aware Neural Network Architecture We’ve developed SpatialAwareGNN, combining E(3)-equivariant convolution with spatial attention. 🔹 It ensures rotational & translational invariance, accurately modeling molecular binding interactions. (3/9)

1. Breakthrough Molecular Representation Our model integrates 3D conformational info with RBF encoding, capturing spatial relationships. 📊 This enables it to identify how conformational changes impact biological activity—something 2D methods can’t do. (2/9)

Driving Drug Discovery with Advanced Activity Prediction We’re excited to introduce our next-gen drug activity prediction model, designed to enhance accuracy, reduce costs, and accelerate drug discovery. 🧬 Here’s what makes it unique: 👇 (1/9)

🔗 Learn more: For detailed insights, visit my LinkedIn post: 👉 linkedin.com/posts/shengyao… #AI #DrugDiscovery #Biotech #Collaboration (9/9)

Let’s collaborate! 🤝 Our AI-driven model can significantly enhance your drug discovery efficiency and accuracy. 📩 Contact us at: [email protected] (8/9)

6. High Efficiency & Seamless Integration Optimized with GPU acceleration and batch processing, our model screens millions of compounds in hours. ⚙️ It integrates with virtual screening, molecular generation, and ADMET prediction pipelines. (7/9)

5. Robust Generalization Capability Trained on a large-scale ChEMBL dataset, our model generalizes well across diverse targets. 🔹 It maintains accuracy even with novel scaffolds or low-data scenarios, making it highly adaptable. (6/9)

4. Multi-Scale Feature Fusion We integrate atomic, local, and global molecular features through a multi-scale fusion mechanism. 🧬 This enables the model to capture activity-determining factors from fragments to full conformations. (5/9)

3. Reliable Predictions with Quantified Uncertainty Our model uses Bayesian estimation to quantify prediction uncertainty. ✅ This provides confidence intervals, helping prioritize high-confidence hits and reducing costly false positives. (4/9)

2. Spatial-Aware Neural Network Architecture We’ve developed SpatialAwareGNN, combining E(3)-equivariant convolution with spatial attention. 🔹 It ensures rotational & translational invariance, accurately modeling molecular binding interactions. (3/9)