Dheeraj Prakaash

excited to release a new benchmark for protein fitness prediction: FLIP2 FLIP2 has 7 new datasets spanning enzymes, PPIs, and light-sensitive proteins, + splits designed to test generalization in realistic protein engineering settings paper, data, code: flip.protein.properties

🚀 Introducing Protenix-v1, the first open-source model achieving AF3-level performance Highlights: 🔹 Verified inference-time scaling behavior 🔹 RNA MSA & protein template support 🔹 Additional release: Protenix-v1-20250630 trained on a larger dataset 🔹 PXMeter v1.0.0 for transparent evaluation (6k+ complexes, time-split & domain-specific subsets) 🔗 Code: github.com/bytedance/Prot… 🔗 Eval toolkit: github.com/bytedance/PXMe… 🔗 Online server: protenix-server.com

A new best-in-class structure predictor AND de novo design protocol Protenix-v2 claims to outperform AlphaFold3 in antibody-antigen structure prediction tasks, showing a 13% increase over its previous generation in DockQ scores. Available on @tamarindbio today. Protenix-v2 with only 5 seeds beats Protenix-v1 with 1000 seeds on antibody–antigen prediction. This implies a technical improvement, while not needing to massively scale inference of a given model like other providers previously showed. In addition, the authors use Protenix-v2 as a scoring and ranking mechanism for de novo antibody design. They report a 100% target-level success rate on the current soluble-target panel, meaning at least one confirmed binder for every tested target, with BLI-confirmed VHH-Fc hit rates from 2% to 48%. They also show that epitope choice matters a lot: on AMBP, one epitope gave 4% hit rate and another 48%. The GPCR result is probably the most impressive experimental result in the paper. With only 16–30 tested designs per target, the protocol shows VHH-Fc hit rates of 16%, 62%, 40%, and 88% across four GPCRs, and corresponding mAb hit rates of 0%, 17%, 50%, and 44%. They also report a best GPRC5D VHH-Fc binder of 112 pM under avidity conditions. Congratulations to the @ai4s_protenix team on the release!

Machine Learning-Assisted Local-to-Global Optimization Strategy for Accelerated Molecular Cluster Structure Prediction pubs.acs.org/doi/10.1021/ac… #JCIM Vol66 Issue5 #MachineLearning #DeepLearning

@tamarabidone 🥂👏🏻✨ inspiring!

Started my faculty role in 2019 with <300 citations and ~10 papers. Six years later: nearly 4× growth. If you feel disconnected from the broader scientific community, your path is valid. Your work matters. Your impact will find its way. Keep going.

Interested in #integrativemodeling of #biomembranes? Join us to hear Weria Pezeshkian talk about his recent work enabling the modelling of biological membranes across scales 🗓️ 14 April 2026, 15:00 CET ✍️ bioexcel.eu/y50u #ComputerSimulation #moleculardynamics

Holy smokes... Google Drive's doc scanner is wild. > multi-page real-time scanning > auto/continuous capture > duplicate page detection > redesigned beta UI Doc scanning will never be the same... 🤯

We developed a reproducible protocol to automate the equilibration of mesoscale all-atom lipid vesicles, with systems reaching 150M atoms. Grateful to @LCasalino88 and @RommieAmaro for their guidance and mentorship. Code: github.com/matteo-castell… Paper: pubs.acs.org/doi/10.1021/ac…

Today @GoogleMaps is getting its biggest upgrade in over a decade. By combining our Gemini models with a deep understanding of the world, Maps now unlocks entirely new possibilities for how you navigate and explore. Here’s what you need to know 🧵

LLMsFold: Integrating Large Language Models and Biophysical Simulations for De Novo Drug Design 1. A new computational framework called LLMsFold combines large language models with biophysical simulations to accelerate early-stage drug discovery, running on consumer hardware in minutes rather than requiring high-performance computing clusters. 2. The pipeline integrates three key components: geometry-based pocket detection using Convex Hull algorithms, molecular generation via Llama-3-70B with in-context learning, and structural validation through Boltz-2 co-folding for affinity prediction. 3. Unlike conventional approaches that require task-specific fine-tuning, LLMsFold leverages pre-trained LLM weights through carefully designed prompts with example molecules, enabling rapid target switching without retraining. 4. The system employs a reinforcement learning feedback loop where top-scoring molecules from Boltz-2 evaluation are reintroduced as prompt examples, iteratively refining candidates toward optimal binding affinity and synthetic accessibility. 5. Applied to two challenging targets—ACVR1 for fibrodysplasia ossificans progressiva and CD19 for B-cell malignancies—the method generated novel candidates with predicted nanomolar potencies that passed drug-likeness and novelty filters. 6. For the kinase target ACVR1, the top candidate showed predicted IC50 of 129 nM with high confidence metrics, while for the protein-protein interaction target CD19, the best molecule achieved predicted IC50 of 188 nM at a clinically validated epitope. 7. All final lead compounds were confirmed as novel chemical entities with no matches in PubChem, demonstrating the method's ability to explore new chemical space rather than retrieve known inhibitors. 8. The entire workflow completes in under 6 minutes on a standard MacBook Pro with M3 chip, making advanced de novo drug design accessible to academic groups and small biotech companies with limited computational resources. 💻Code: github.com/tacciolilab/LL… 📜Paper: biorxiv.org/content/10.648… #DeNovoDrugDesign #ComputationalChemistry #LargeLanguageModels #AlphaFold #Boltz2 #DrugDiscovery #Cheminformatics #MachineLearning #RareDisease #OpenScience

It's only been just over 67 hours since OpenAI dropped GPT-5.4. And people can't stop getting creative with it. 10 wild examples. Bookmark this 👇

Assessing Boltz-2 Performance for the Binding Classification of Docking Hits #Docking pubs.acs.org/doi/10.1021/ac… #JCIM Vol66 Issue3 #MachineLearning #DeepLearning

🚨 BREAKING: Someone just open-sourced software that sees you through walls using only WIFI signals. it’s called WiFi-DensePose. It maps your exact body pose in real-time. no cameras. no sensors. just your living room router. 100% Open Source.

I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then: - the human iterates on the prompt (.md) - the AI agent iterates on the training code (.py) The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. In the image, every dot is a complete LLM training run that lasts exactly 5 minutes. The agent works in an autonomous loop on a git feature branch and accumulates git commits to the training script as it finds better settings (of lower validation loss by the end) of the neural network architecture, the optimizer, all the hyperparameters, etc. You can imagine comparing the research progress of different prompts, different agents, etc. github.com/karpathy/autor… Part code, part sci-fi, and a pinch of psychosis :)

When I said I wanted to do science this is what I meant.

Hello everyone, I am looking for my next role in the computational chemistry/biophysics space 🛫 I have 12 years of experience in protein simulations/SBDD, looking for national lab/industry positions anywhere in the world. Any connections/leads welcomed - thanks in advance! 🙏🏻

Check out our review on Graph #NeuralNetworks #GNNs in #MolecularDynamics! We show how #AI is used for #ForceFields, free-energy and analysis - including our recent attention-based GNN workflow - and how you can integrate your favorite AI in #compchem! sciencedirect.com/science/articl…

If you are a #GROMACS user / interested in #moleculardynamics simulations join us next week for our "What's new in GROMACS 2026.0" #webinar 🗓️ 3 March 2026, 15:00 CET ✍️ bioexcel.eu/cpok @GMX_TWEET

Interested in molecular animation and in SF for #BPS2026? Stick around for the evening Structural Biology workshop (Rm 204/205/206) and my talk at 8:45pm (New Tools for Visualizing Dynamic Molecular Models) to learn about ProteinBlender! animation-lab.github.io/ProteinBlender/