Efficient and Programmable Exploration of Synthesizable Chemical Space 1. A groundbreaking study introduces PrexSyn, a novel model for molecular discovery that ensures synthesizability while efficiently exploring chemical space. This model achieves near-perfect coverage of synthesizable molecules, setting a new benchmark in molecular design. 2. PrexSyn leverages a decoder-only transformer architecture trained on a billion-scale dataset of synthesizable pathways paired with molecular properties. This unique approach allows the model to generate molecules based on property prompts, enabling programmable objectives through logical queries. 3. The model demonstrates state-of-the-art performance in chemical space projection, achieving a 94% reconstruction rate on the Enamine REAL space and a significant improvement in similarity scores compared to previous methods. 4. PrexSyn excels in sampling efficiency for black-box oracle functions, outperforming both synthesis-agnostic and synthesis-based methods. This efficiency is attributed to its well-structured query space, making optimization more tractable than in discrete molecular graph spaces. 5. The study also highlights PrexSyn's ability to handle composite property queries, allowing users to specify complex molecular objectives using logical operators. This feature enhances the model's flexibility and applicability in real-world drug discovery scenarios. 6. PrexSyn's high-throughput data generation engine enables billion-scale training with modest computational resources. This innovation significantly reduces training time and cost while improving model performance. 7. The model's effectiveness is demonstrated in tasks such as scaffold hopping and docking-based molecular optimization, where PrexSyn generates high-scoring molecules with improved properties compared to existing baselines. 📜Paper: arxiv.org/abs/2512.00384v1 💻Code: github.com/luost26/PrexSyn #PrexSyn #MolecularDesign #ChemicalSpace #Synthesizability #AIinChemistry #DrugDiscovery

Many users of my `academic-homepage` template have customized it to showcase their pets, travels, art, and more. So, I made this very charming page that lets you display anything beyond publications! Try it! Repo: github.com/luost26/academic-… Demo: luost.me/academic-homepage/s…

🔥Spatial intelligence needs fast, *interactive* 3D world generation 🎮 — introducing WonderWorld: generating 3D scenes interactively following your movement and content requests, and see them in <10 seconds! 🧵1/6 Web: kovenyu.com/WonderWorld/ arXiv: arxiv.org/pdf/2406.09394

Just arrived in Vienna! I will be at #ICML2024 on Tuesday and Wednesday, and will be presenting our ChemProjector in Poster Session 3 on Wednesday! Drop by and have a chat!🤗

Acting on a piece of folded RNA (@PDBeurope 6D3P) with my hands in #AugmentedReality using a @MetaQuestVR 3, backed up by @openmm_toolkit running an @ambermdprog forcefield. All web and accessible to you in some months hopefully! By @fabiojcortes and @lucien_krapp :-)

Wow... Just tested the new keyframe update from Luma. 🤯

🧵1/n LLMs significantly improve Evolutionary Algorithms for molecular discovery! For 18 different molecular optimization tasks, we demonstrate how to achieve SOTA performance by incorporating different LLMs! Learn more in our new paper! Website: molleo.github.io/(w/ Code)

🧵1/7 Introducing an “Encyclopedia” of Molecular Design with Machine Learning @NatMachIntell! nature.com/articles/s42256-0… Collaboration with @arian_jamasb*, @JeffGuo__*, @TianfanFu, @charlieharris01, @yingheng_wang, @chenru_duan, @pl219_Cambridge, @pschwller, and Tom Blundell.

UPDATE: The footer text has been made configurable! and the cat image is hidden by default. Check out my GitHub Pages template for academics here! Feedbacks and contributions are welcome!! github.com/luost26/academic-…

I've created a Github Pages template based on my homepage. Feel free to use if you love it! github.com/luost26/academic-…

Tom was the first to lose his career to Al

I'm excited to share SyntheMol, a generative AI model for drug design optimized for creating easily synthesizable molecules. We applied our model to generate, synthesize, and validate 58 novel antibiotic candidates and found 6 potent hits. @NatMachIntell nature.com/articles/s42256-0…

impressive😎

Finally… after a decade

Introducing the new JupyterLab Desktop! blog.jupyter.org/introducing…