Alex Lubbock

Automation has been used in chemistry for decades. The problem for many though is applying these tools is not intuitive. Many papers describe these expensive or "home built" setups that also require the ability to code; it's not practical. Opentrons Labworks Inc. along with AI-assisted coding tools from Anthropic and OpenAI are changing that. It is now easy and affordable to start. At Cime, we are working to push past the simple one step transformations to explore new chemical space and sharing some of what we learn along the way. Patrick Doerner Barbour perfectly highlights how we do more together by open sourcing tools that are enabling us to automate synthesis (with more to come!).

18

Hey LinkedIn, 👋 I'm starting a new series of post which will give readers a bit of an insight into the life of a data scientist. I'll be talking about some of my favourite tools/programs/softwares that I use in my day-to-day work. 🛠️ Data Science Tools #1: Vim! Vim is a low-level, retro text editor that you open as a command in the terminal. It's automatically installed in most linux/unix operating systems. To try it out just type `vim` and it will pop up. 💻 If you've never opened it before you might be confused or frustrated. (i.e. "ugh, how do I close vim!? 🤬 ") Don't worry I was there too. The idea behind programming in Vim is that your fingers never have to leave your keyboard. You can navigate across the whole editor using keyboard shortcuts. If you want to actually type something you click `i` to enter insert mode and then escape to get back to command mode (another top tip: I remapped CAPS-LOCK to escape, because who really needs CAPS-LOCK). Some useful commands I use often to navigate the editor 🚢 - `gg` go to the top of the file - `G` go to the end of the file - `ctrl+u` page up - `ctrl+d` page down - `w` move forward one word - `b` move backword one word - `/` search for something The other great thing about Vim is that it is customizable! You can install so many plugins and colorschemes onto it, make your own commands, you can really make it your own. Opening up my Vim terminal is often where I feel most at home 🏡 . Here are some of my favorite plugins - NerdTREE: Gives you a small filebrowser in pane next to your editor - Copilot: LLM generated code autocompletions (subscription required) - Colorscheme: Catppuccin-macchiato/tokyo-night - vim-airline: Gives a nice looking bar on the bottom showing your filename and folder and environment you are in - Comment: enable block commenting with `,gc` There's loads more that you can do with it. I love Vim and often people ask me why not just use VSCode or Cursor. For me it boils down to a couple reasons: 1. Speed 🏎️ : The slowest unit operation at your desk is moving your hand from your keyboard to your mouse. When you eliminate that operation your work is much faster. Albeit this does come with a learning curve first of learning the commands but that is a small cost for many gains in the future. 2. Focus 😌 : Maybe the speed argument was more relevant 5 years ago... now I admit with AI coding agents it's hard to beat that. However, I find that I never focus better on my programming then when I have a just the terminal open in dark mode and nothing else. There is no distractions and it feels great. 3. Fun 🥳 : Once you learn the commands Vim is a lot of fun. It's like this subtle game I am playing in the background whilst working. I love it. It's hacker, its retro and it's one of the small things that I lowkey love about my job. If you want to try out vim (I actually use NeoVim) you can install my setup here: https://lnkd.in/evGQBiYz

51

3 Comments

We’re delighted to welcome Dr Shruti Mishra for this week’s DIS seminar: 📅 Thursday 5th March 2026 ⏰ 2–3pm 📍 Room A, West Hub Title: Reinforcement Learning in Locomotion and Navigation Environments Dr Mishra will introduce the reinforcement learning framework and Markov Decision Process formulation, before presenting applications to locomotion and navigation tasks across natural and synthetic environments. The talk will explore practical design choices, such as reward shaping, observation spaces, and algorithmic considerations, and conclude with an open-source environment for experimentation. About the speaker Dr Shruti Mishra is an Encode: AI for Science Fellow in the Department of Applied Mathematics and Theoretical Physics, where she develops reinforcement learning methods for continuous control problems such as fluid flows. Her research spans reinforcement learning, mathematical modelling, and continuum mechanics. Previously, she was a Research Scientist at Sony AI and DeepMind. She holds a PhD in Applied Mathematics from Harvard University and an MEng in Mechanical Engineering from Imperial College London. ☕ Coffee and cakes to follow — all welcome! #ReinforcementLearning #ArtificialIntelligence #MachineLearning #ContinuousControl #Locomotion #Navigation #AppliedMathematics #AIforScience #ResearchSeminar #AcademicEvents

1

"At the same time, compute is becoming essential to modern science from genomics, to climate modelling to materials discovery, and the science of AI itself." Last month, the UK government released its UK Compute Roadmap (link below). It's great to see that amongst all the hype around Gen AI, decision makers still have a broader vision and understanding of the importance of the development of computing technologies for the entire modern economy beyond just this, important but often overemphasized, use case. The UK has long been at the forefront of computing technologies but it will require a concrete vision and more, importantly, a decisive and patient implementation strategy to maintain, and grow, its leading position. Reading through the roadmap, the vision is clearly there but now comes the hard part. Excited to see what comes next! https://lnkd.in/dbdUW9Mb

27

1 Comment

I'm incredibly proud to announce a productive NeurIPS for our Post-training AI Research team. We had two papers on test-time scaling accepted into the Workshop on Efficient Reasoning, contributing to a fantastic ~6 acceptances for the wider Applied AI Research @ Cerebras. Both of our papers are now available on ArXiv:  - The Conductor and the Engine: A Path Towards Co-Designed Reasoning. Link - https://lnkd.in/g5Dnnz6C. - Calibrated Reasoning: An Explanatory Verifier for Dynamic and Efficient Problem-Solving. Link - https://lnkd.in/gH2jTGeP. In the spirit of collaboration, we are also open-sourcing our updated CePO flow as part of our OptiLLM library on GitHub. This is the same methodology that achieved top scores on the Artificial Analysis Leaderboard. Link - https://lnkd.in/gd6anwFN Congratulations to the team on this incredible milestone: Anisha Garg, David Bick, Engin Tekin, Michael Wang, Pawel Filipczuk, Amaan Dhada, Yash More, Nishit N. and rest of the Post-training Team! Looking Ahead A key enabler for these results and our upcoming work in Coding Agents has been using Reinforcement Learning (RL) to provide LLMs with new, specialized expertise, which makes them highly amenable to test-time compute scaling. This brings me to our next exciting step... 🚀 Announcing Limited Early Access to our RL Service! 🚀 Is your team excited by the potential of powerful LLM models — closed source like GPT-5/Claude/Gemini or Open-source like Qwen3, GPT-OSS — but frustrated when they fail that "last mile" on your specific, critical tasks? We are opening up a limited early access program to help you solve this. Our RL service is designed to transform general models into world-class experts for your unique domain. If you're interested in building an AI system that is an expert at solving your tasks, reach out to me to see if you're a fit for the program. 📧 Email: ganesh.venkatesh@cerebras.net When you reach out, please include: - Subject: "RL Service Early Access" - Body: A brief description of your application and the "last-mile" challenges you're facing. Looking forward to post-training your problems away!

122

1 Comment

New on the R Consortium blog: “Contributing to base R with Coding Equity and Joy — Inside the R Contributors Project.” Ella Kaye, Senior Research Software Engineer, University of Warwick, shares how the R Contributors project is making it easier—and more welcoming—to contribute to base R: R Developer Days, monthly contributor office hours, and a C Study Group for R contributors. She also explains why using GitHub (issues, discussions, labels) can lower barriers vs. Bugzilla. Bonus: a fun case study on learning-with-joy through the “aperol” R package—and how community feedback turned a silly idea into real learning. Bonus-bonus: Ella covers the history of rainbowR, a community that connects, supports and promotes LGBTQ+ folk who code in R, and spreads awareness of LGBTQ+ issues through data-driven activism. Read it all here: https://lnkd.in/ggnbiVjH #rstats #opensource #RProject #Community #ResearchSoftware

48

Winners of SoTA’s Opening the Black Box interpretability hackathon, Nitish Mital, (PhD) and Simon Malzard, PhD explore how synthetic data can be made smarter, not just bigger. In their letter to SoTA, they introduce a SHAP-guided synthetic data refinement approach - using explainable AI to identify and correct the features that mislead models. By closing the loop between explainability, simulation, and data design, their work points toward a future where synthetic data actively teaches models to see the world more accurately. Read their letter in full (link in the comments) to learn about their project, how it relates to their PoC at the SoTA Hack last January, and what their future plans are.

23

1 Comment

How is AI changing our lives? 🎧 The Turing's Adrian Weller joined Brian Cox, Neil Lawrence, Steph Wright and Jeanette Winterson to explore this fundamental question. Listen to the latest episode of The Francis Crick Institute’s A Question of Science podcast: https://lnkd.in/eR3yedD3 #AI #AIEthics #AIforGood #AIGovernance

163

1 Comment

AbBFN2 is out! We introduce a new generative AI model which has a flexible, multi-objective approach to antibody engineering. 🚀 Try AbBFN2: https://lnkd.in/dN43HPT8  📄 Preprint: https://lnkd.in/dFhehcSw  💻 Access the code: https://lnkd.in/dNdB6sC2

40

5 Comments

We are pleased to share that Dr Sergii Strelchuk, Tutorial Fellow in Computer Science at Jesus College and Associate Professor at Oxford, has been selected to lead the final phase of Quantum Pangenomics within the prestigious Wellcome Leap Q4Bio Challenge. This international project brings together experts from quantum computing and the life sciences to explore how quantum approaches might transform the analysis of large and complex genomic datasets, with potential benefits for personalised medicine and disease detection. Supported by Wellcome Leap and involving partners including the Wellcome Sanger Institute, the University of Cambridge and the University of Melbourne, the initiative places Oxford at the centre of a rapidly developing field. 🔎👉 Read more: https://lnkd.in/eKV7Ez2m

21

We're delighted to release another "Conversations with PhD Researchers" interview - this time with PhD Student Emilio McAllister Fognini. In this episode, Emilio introduces us to the world of Neural Operators - a powerful and abstract class of machine learning techniques used to model differential equations. At the CDT, we’re proud to support research pushing the boundaries of computational science and creating pathways toward real‑world impact. Emilio’s work is a fantastic example of the innovation emerging from our centre, and we would like to pass on our thanks to Emilio for the fantastic explanation and to Reuben Adams for producing another great interview! UCL Computer Science UCL Centre for AI https://lnkd.in/gEnVnVbY

9

1 Comment

Wrapping up London Calling 2025 - Oxford Nanopore Technologies’ annual meeting - and its satellite Informatics Day with a few personal highlights (in no particular order): 1. R10 chemistry is here to stay No updates are foreseen in the R10 chemistry, which is welcome news for those working toward clinical implementation of Oxford Nanopore - more on this below. 2. Dorado as a variant calling framework "Dorado is not just a basecaller, it's a framework. [...] It's time for ONT to take more ownership of variant calling.". ONT developing its own variant caller as part of the Dorado engine (it still will use Minimap2 for mapping). Preliminary results show promising performance beyond Clair3, ONT’s current recommendation. 3. Diverse and growing use in clinical care Biased as we might be as event attendees, the breadth of real-world clinical genomics applications is impressive. From early adopters like Danny Miller - who now predicts broader adoption within 3 years - to national programs and clinical labs transitioning from short-read to ONT, the shift seems both credible and imminent. 4. Adaptive sampling love A favorite among users for its flexibility: a cost-effective method to target specific regions while gaining bonus data from off-target regions (e.g. imputation, methylation). Hard to imagine its use declining unless WGS costs drop significantly. 5. Genome assemblies for humans Unexpectedly (to me), ONT is now highlighting its value for more than just microbes. In human genetics, ONT offers better resolution of complex regions (e.g. *SMN1/2*), accelerating pan-genome research and utility. 6. Proteomics: still early days While ONT’s expansion into protein profiling is often mentioned at London Calling, it still feels a few years out for routine users. That said, Jeff Nivala’s work in this space is highly promising and clearly setting the stage. 7. Data standards & sharing A maturing field. There was a strong multidisciplinary focus on how data is stored, represented, and shared among stakeholders - clinicians, patients, collaborators - which signals a maturing ecosystem striving for responsible, trusted data use. 8. Clinical reporting and literacy Important conversations around how to upskill clinicians in genomics and ensure they feel confident reporting findings to patients. Essential for real-world clinical integration. A few quotes that stayed with me: * Ewan Birney on data standards: “If we look back, we’ve come a long way - but we’re not at the top of the mountain.” * Rowan Gardner on adoption: “You’ll never change clinical practice if you try to hit on it.” * Georgina Dawson: “People can only understand what they can perceive.” #nanoporeConf

33

3 Comments

Quantum Motion brings together an amazing group of people to build amazing things. Listen to the team talk about delivering our first, full-stack quantum computer to the UK's National Quantum Computing Centre (NQCC). "This could only have been done here at Quantum Motion because we've gathered the expertise of all the necessary teams to deliver this full-stack system. We have quantum hardware engineers, intelligent automation professionals, we have theorists, we have experts in modelling, and without all these people, this endeavour wouldn't have been possible. It's not that we have all these teams, it's that we have the best people in each one of these teams." - M. Fernando González Zalba, Principal Quantum Engineer at Quantum Motion. Drive the future of quantum computing. Join the team: https://lnkd.in/eJWgAGj6 Team featured: April Carniol, Anna Stockklauser, David J. Ibberson, Katarina Brlec, Giovanni Oakes, David Wise, Tuula Ritakari

122

5 Comments

I'm looking forward to attending the CoSeC conference and sharing the metadata work we've been developing within the ARC team at the Rosalind Franklin Institute. FAIR pipelines aren't built overnight, but automating metadata capture is a big step towards more transparent, reproducible, and efficient data management across the research lifecycle. This talk is a snapshot of where we are now and where we're heading next

24

Is there a bitter lesson in antibody modelling? In 2019, Rich Sutton published a now famous blog post titled "The Bitter Lesson". The core argument is that generic models that scale with compute tend to overtake specialized, domain-specific approaches. For LLMs, much of their emergent behaviour can be attributed to scale. This convinced researchers to spend less time worrying about minor architectural tweaks to account for novel edge cases and keep scaling their compute, dataset size, and parameter count. For protein models, consider the results from the Ginkgo Datapoints Developability Competition (held by Ginkgo Bioworks). 113 teams tried to predict antibody properties from sequence alone. The best models showed moderate correlation for clearer properties like hydrophobicity, but for complex ones like self-association and expression titer, predictions were falling short. Models that looked strong in validation struggled on the held-out set. The authors name the bottleneck directly: available datasets are too small and too heterogeneous. If you follow the bitter lesson to its conclusion: as the protein language models tested in this competition scale their data, their ability to generate zero-shot developable antibodies will continue to improve. A recent finding is starting to demonstrate this! A team at JURA Bio, Inc. recently used a specialized method - variational synthesis - to design ~10^17 antibody sequences for roughly $1,000. The method folds the generative model into the DNA synthesis process itself, so you define a target distribution over antibody sequences, optimize the physical synthesis parameters so the library coming off the machine approximates that distribution, and let the chemistry do the sampling. In their blog post, the authors train a transformer with cross attention and CLIP-style contrastive learning on their dataset to predict scFv-pHLA interactions, which is simpler than other models leveraging MSAs/structural data. The model generalised to unseen antibodies and unseen targets, with some targets reaching over 100,000x enrichment for specific binders. This also reminded me of Leash Bio's work with Hermes, described well in a blog post by Abhishaike Mahajan from late last year. Hermes is a lightweight 50M-parameter transformer trained exclusively on Leash's proprietary protein-ligand binding. On held-out chemistry with zero overlap to training, it outperformed Boltz-2, a far more complex structural model. The takeaway was that dense, high-quality, internally consistent data can outweigh architectural sophistication. So, is there a bitter lesson in antibody modelling? These evidences suggest yes. When you get the data right, simpler models can keep up with or outperform more complex ones. We're thinking carefully about the data pipelines at Actimo Labs and how we can leverage simpler models for better, more developable antibodies. #antibodyDiscovery #drugDiscovery #machineLearning #computationalBiology

35

6 Comments

While there's been a lot of interest in using NNPs to study protein–ligand interactions with DFT-level accuracy, Ishaan Ganti found that NNPs struggle for these complexes and physics-based xTB methods do much better. "Quantity has a quality of its own," as the saying goes...

66

5 Comments