Postdoc @RIKEN, PhD from @ColegroupNCL Newcastle University. Computational Chemist/Drug Design & Open Source Dev. Fan of SC2. An English flavoured scientist.
Joined February 2020
- Tweets601
- Following1.1K
- Followers240
- Likes430
BenCree retweeted
Introducing Langevin Flow Maps
TLDR: Neural Network Potentials that learn their own stochastic pathwise solution for faster simulation.
- Transferable across molecular systems.
- Transition dynamics are preserved.
- No trajectory data are required.
Work together with @sgmccallum, supervised by @AlexanderTong7, and @JamesFosterBath.
Arxiv: arxiv.org/abs/2610.05998
Data: huggingface.co/datasets/nikl…
Code: github.com/sammccallum/lfm
BenCree retweeted
As agents perform AutoResearch they get fixated on their old ideas and do only incremental innovations. We introduce a simple context management trick that forces the agents to forget everything they have done except the best solution they have found so far.
This simple modification in context management allows the agents to get out of their idea basins and come up with more novel solutions.
Closely related to the work 'Fork and Flush' and 'Idea-basins' that @DimitrisPapail posted a few days ago.
BenCree retweeted
reading feynman lecturess as feynman intended: with accompanying ai-made demo sections for each chapter
rapid acquisition of knowledge is possible thanks to superintelligence
BenCree retweeted
The Kalman filter community uses the term “smugness” to describe a model that ignores new data. That’s a good name for it. Should be used more widely.
BenCree retweeted
i managed to photograph my camera's sensor!
the architecture shows across multiple frames that received a very particularly angled direct sunbeam, which was collimated by the lens, bounced off the chip surface into the rear lens element, and back to the sensor legibly magnified
BenCree retweeted
Can’t make this up…the MacBook is using its webcam to look at its screen in a mirror to improve AMD Radeon chip support in Omarchy.
BenCree retweeted
📢 Call for proposals: Boltz small-molecule design collaboration! 🧬
Can we help design your ideal molecule? Can you help us improve our open-source models? Please reach out or share with scientists you know! More details below!
It has been great to see the level of excitement around our Boltz-2 release. Along with the structure and affinity model, we are working on new workflows for hit-discovery and hit-to-lead of small molecules. We have seen very strong computational results and we are now interested in validating them across a range of different targets and real-world settings.
Do you have a target on which you would like to design or optimize small-molecule binders? Do you have an associated chemical or functional assay that you can run to validate the designs? If the answer is yes to both questions, we would love to hear from you and collaborate.
As part of these collaborations, we would work alongside you to tailor our workflows for your task, run the large scale design on our cluster, and return to you the designed molecules alongside their expected scores. We ask you to commit to testing the top (according to your judgement) designed compounds and send us back the experimental results within 2 months of the selection.
You’ll have full IP of the compounds we send you and have the opportunity (not necessity) to be included alongside with the experiments you run in a joint journal publication. If this is something that might interest you, please reach out to me at gcorso at csail.mit.edu or DM me on the boltz-community Slack!
BenCree retweeted
A surprising & little-known results in classical statistics:
Mean (μ) and median (m) are within one std deviation:
|μ−m| ≤ σ
For unimodal densities, bound is even tighter
|μ−m| ≤ 0.7746 σ
This beautiful results first appeared in a 1932 paper by Hotelling & Solomons
1/3
BenCree retweeted
You can estimate logarithms base 10 without a calculator with the approximation (x - 1)/(x + 1).
BenCree retweeted
Transforming molecular machine learning with stereoelectronics-infused molecular graphs (SIMGs)
Traditional molecular representations in machine learning often overlook essential quantum-mechanical effects, limiting predictive power in complex chemical systems. In my latest Substack post, I discuss recent work by Boiko, Reschützegger, Sánchez-Lengeling, Blau, and Gomes, who propose SIMGs—a novel molecular graph approach explicitly embedding quantum-chemical stereoelectronic details.
Their method leverages Natural Bond Orbital (NBO) analysis to significantly boost predictive accuracy, outperforming classical models on diverse molecular properties and even large biological systems previously inaccessible to quantum-level precision. This innovation not only advances molecular predictions but enhances interpretability and insight into subtle quantum interactions.
Link to your Substack post: open.substack.com/pub/bravoa…
BenCree retweeted
A new flagship quantum chemistry dataset is out! Omol25 contains more than 100 million DFT calculations across an expanse of chemical space.
Collaborating with the FAIR team on this has been super fun!
arxiv.org/abs/2505.08762
BenCree retweeted
Today in @Nature, we’re presenting GenCast: our new AI weather model which gives us the probabilities of different weather conditions up to 15 days ahead with state-of-the-art accuracy. ☁️⚡
Here’s how the technology works. 🧵goo.gle/49trAOv
BenCree retweeted
Here's the last paper from my time in @FrankNoeBerlin's group.
We use normalizing flows (NF) to build a whole phase diagram from just two MD simulations. Especially useful for ML potentials, since it requires much fewer energy evaluations than traditional methods!
Efficiently #MachineLearning phase diagrams with conditional Boltzmann Generators for Materials. By Max Schebeck, @inve_michele and Jutta Rogal!
arxiv.org/abs/2406.12378
BenCree retweeted
Boltz-1: Democratizing Biomolecular Interaction Modeling
1. Boltz-1 introduces the first fully open-source model achieving AlphaFold3-level accuracy in predicting 3D structures of biomolecular complexes, democratizing access to cutting-edge structural biology tools.
2. The model incorporates several innovations, including improved MSA pairing, unified cropping strategies, and robust pocket-conditioning algorithms, enabling precise modeling of protein, nucleic acid, and ligand interactions.
3. Benchmarked against Chai-1 and CASP15 datasets, Boltz-1 demonstrated comparable or superior performance across multiple metrics, including local structure accuracy (LDDT) and global quality (TM score).
4. By leveraging a diffusion-based generative model, Boltz-1 improves training efficiency and inference robustness while significantly reducing computational costs compared to AlphaFold3.
5. Boltz-1 supports user-defined binding pocket specifications, allowing for enhanced flexibility in real-world applications such as drug design and functional annotation.
6. Released under the MIT license, Boltz-1 includes training code, model weights, datasets, and evaluation benchmarks, fostering community-driven innovation and collaboration.
7. While exhibiting strong performance, Boltz-1 highlights challenges in resolving overlapping chains and ligands, pointing to future opportunities for model refinement and expanded datasets.
@BarzilayRegina @Itamarchinn @GabriCorso @jeremyWohlwend @pas_saro @WojTechnology @KenLeidal @meteos_97
💻Code: github.com/jwohlwend/boltz
📜Paper: biorxiv.org/content/10.1101/…
#BiomolecularModeling #ProteinStructure #OpenSource #AlphaFold #DrugDiscovery
BenCree retweeted
Hmm. We can put boundaries around it. NICE use ~£25k as value of extra quality year for humans. £100m is value of 4000 extra years. Bats live ~15 years, so that 266 bats if valued same as humans….. How big is the colony and how do you value bats vs humans?
BenCree retweeted
Introducing SRME: a new metric for further assessment of foundational machine learning interatomic potentials!
👉A smaller SRME value indicates better performance for calculating thermal conductivity
🖥️matbench-discovery.materials…
Huge thanks to Janosh @jrib_ & Yuan @cyrusyc_tw
🚀Introducing autoWTE: using foundational machine learning Interatomic Potentials, phono3py & fine-tuning to accurately predict Wigner Thermal conductivity for crystals of any structure or composition🧊🔍
🌡️: github.com/MPA2suite/autoWTE
#ThermalConductivity #MatbenchDiscovery
BenCree retweeted
We’ve just finished writing the missing 15,616 Wikipedia articles to get complete coverage of all 19,255 human genes. We used PaperQA2, which has higher accuracy than existing human-written Wikipedia articles, as judged by blinded biology PhD students and postdocs. 1/5
BenCree retweeted
OpenDock: A PyTorch-Based Open-Source Framework for Protein-Ligand Docking and Modeling
• A paradigm shift in molecular docking: OpenDock, built on PyTorch, offers an open-source framework with unparalleled flexibility for protein-ligand docking, integrating both classical and deep learning-based scoring functions for enhanced docking accuracy.
• Comprehensive sampling strategies: OpenDock supports multiple sampling methods, including Monte Carlo (MC), genetic algorithms (GA), and particle swarm optimization (PSO), enabling more refined docking poses and side-chain orientations.
• Custom scoring and constraint functionality: Users can define custom scoring functions and constraints, such as covalent docking or enzyme-substrate models, to guide the optimization of docking poses beyond standard methods like AutoDock Vina.
• Seamless integration of deep learning models: The framework incorporates DeepRMSD and other DL-based scoring functions, allowing for the generation and refinement of docking poses with state-of-the-art performance in benchmark tasks.
• Validated against competitive tools: In redocking and cross-docking tasks, OpenDock demonstrates improved docking success rates compared to popular alternatives, including AutoDock Vina, Gnina, and DiffDock.
• Built for flexibility and future innovation: Its modular design allows users to experiment with scoring strategies and introduce new constraints, fostering innovations in drug discovery and enzyme engineering.
💻Code: github.com/guyuehuo/opendock
📜Paper: doi.org/10.1093/bioinformati…
