From the 11th until the 16th of May, we (Gemma and Henriette) attended PEGS Boston 2025. First of all, we will share some tips for preparing to attend a conference. We will give some general feedback on the conference and share some general highlights. Lastly, we will mention some of the talks we found interesting.
Continue reading
Pose Prediction: Does Your Model Generalize? The Role of Data Similarity
In our recent work with the PoseBusters benchmark, we made a deliberate choice: to include both receptors seen during training and completely novel ones. Why? To explore an often-overlooked question: how much does receptor similarity to training data influence model performance?
Continue reading
On the Joys of vim-like Browsing
Reflections on Pointlessness
One of the great delights in this life is pointless optimisation. Point-ful optimisation has its place of course; it is right and proper and sensible, and, well, useful, and it also does, when first achieved, yield considerable satisfaction. But I have found I soon adjust to the newly more efficient (and equally drab) normality, and so the spell fades quickly.
Not so with pointless optimisation. Pointless optimisation, once attained, is a preternaturally persistent source of joy that keeps on giving indefinitely. Particularly if it involves acquiring a skill of some description; if the task optimised is frequent; and if the time so saved could not possibly compensate for the time and effort sunk into the optimisation process. Words cannot convey the triumph of completing a common task with hard-earned skill and effortless efficiency, knowing full-well it makes no difference whatsoever in the grand scheme of things.
Continue readingAttention Is All You Need – A Moral Case
It turns out that giving neural networks attention gives you some pretty amazing results. The attention mechanism allowed neural language models to ingest vast amounts of data in a highly parallelised manner, efficiently learning what to pay the most attention to in a contextually aware manner. This computational breakthrough launched the LLM-powered AI revolution we’re living through. But what if attention isn’t just a computational trick? What if the same principle that allows transformers to focus on what matters from a sea of information also lies at the heart of consciousness, perception, and even morality itself? (Ok, maybe this is a bit of a stretch, but hear me out.)
To understand the connection, we need to look at how perception really works. Modern neuroscience reveals that experience is fundamentally subjective and generative. We’re not passive receivers of objective reality through our senses, we’re active constructors of our own experience. According to predictive processing theory, our minds constantly generate models of reality, and our sensory input is then used to provide an ‘error’ of these predictions. But the extraordinary point here is that we never ‘see’ these sensory inputs, only our mind’s best guess of how the world should be, updated by sensory feedback. As consciousness researcher Anil Seth puts it “Reality is a controlled hallucination… an action-oriented construction, rather than passive registration of an objective external reality”, or in the words of Anaïs Nin, half a century earlier, “We do not see things as they are, we see things as we are.”
Continue readingEye of the World by Robert Jordan: A Concise Review
I was recently devastated to hear that Amazon Prime has cancelled the Wheel of Time TV Show, a fantasy epic based on the novels of Robert Jordan. I recently binge-watched the entire show and found it to improve throughout, with the third and most recent season being the best.
In my grief, I turned to something dark – reading the books instead.
I have recently finished the first book (of 12) and thought I would give my thoughts on the story and the storytelling of Jordan as a concise book review so I can get my final Blopig out of the way.
Continue readingChatGPT can now use RDKit!
All chemistry LLM enthusiasts were treated to a pleasant surprise on Friday when Greg Brockman tweeted that ChatGPT now has access to RDKit. I’ve spent a few hours playing with the updated models and I have summarized some of my findings in this blog.
Estimating uncertainty in MD observables using block averaging
When running molecular dynamics (MD) simulations, we are usually interested in measuring an ensemble average of some metric (e.g., RMSD, RMSF, radius of gyration, …) and use this to draw conclusions about the investigated system. While calculating the average value of a metric is straightforward (we can simply measure the metric in each frame and average it) calculating a statistical uncertainty is a little more tricky and often forgotten. The main challange when trying to calculate an uncertainty of MD oveservables is that individual frames of the simulation are not samped independently but they are time correlated (i.e., frame N depends on frame N-1). In this blog post, I will breifly introduce block averaging, a statistical technique to estimate uncertainty in correlated data.
Continue readingMemory Efficient Clustering of Large Protein Trajectory Ensembles
Molecular dynamics simulations have grown increasingly ambitious, with researchers routinely generating trajectories containing hundreds of thousands or even millions of frames. While this wealth of data offers unprecedented insights into protein dynamics, it also presents a formidable computational challenge: how do you extract meaningful conformational clusters from datasets that can easily exceed available system memory?
Traditional approaches to trajectory clustering often stumble when faced with large ensembles. Loading all pairwise distances into memory simultaneously can quickly consume tens or hundreds of gigabytes of RAM, while conventional PCA implementations require the entire dataset to fit in memory before decomposition can begin. For many researchers, this means either downsampling their precious simulation data or investing in expensive high-memory computing resources.
The solution lies in recognizing that we don’t actually need to hold all our data in memory simultaneously. By leveraging incremental algorithms and smart memory management, we can perform sophisticated dimensionality reduction and clustering on arbitrarily large trajectory datasets using modest computational resources. Let’s explore how three key strategies—incremental PCA, mini-batch clustering, and intelligent memory management—can transform your approach to analyzing large protein ensembles.
Continue readingOpen Source Pharma: From Idealism to Pragmatic Solutions
In an industry dominated by patents, proprietary data, and the race to get a first-in-class drug, the concept of open source drug development once seemed like an impossible dream. Yet as traditional pharma continues to leave many global health needs unaddressed—particularly for diseases affecting low and middle income countries1,2—the open source model has evolved from idealistic theory to pragmatic reality. In this post, I’ll lead us through how open source drug development has overcome key obstacles of funding and intellectual property (IP) management to deliver real-world solutions.
Continue readingAn insight into mega-conferences – attending ESCMID Global 2025
I suppose it really hit me when the Viennese border control officer asked, “Ah, you must be here for the conference?” That’s when I realised: this wasn’t just any event. ESCMID Global isn’t your average gathering of lab coat enthusiasts, but rather one of the largest clinical infectious disease conferences in the world. Over 16,000 attendees packed into Vienna for their 35th annual congress.
So, was flying across Europe to attend the Glastonbury of conferences, minus the mud, plus the microbes, worth it?
Well… it depends on what you’re hoping to get out of it. If you’re an academic, you might find that a lot of the sessions lean heavily towards the clinical side of things. On the plus side, it made it easier to narrow down my schedule – with over a dozen sessions happening at any one time, a bit of decisiveness goes a long way. Personally, I found the big-name, high-level keynotes and annual updates from organisations like EUCAST the most accessible and informative.