Python is a shockingly slow language. A test on a raspberry pi of simply “turn this pin on and off as fast as you can” gave the results below.
| System | Library | Speed |
|---|---|---|
| Shell | /proc/mem access | 2.8 kHz |
| Shell / gpio utility | WiringPi gpio utility | 40 Hz |
| Python | RPI.GPIO | 70 kHz |
| Python | wiringPi2 bindings | 28 kHz |
| Ruby | wiringPi bindings | 21 kHz |
| C | Native library | 22 MHz |
| C | BCM 2835 | 5.4 MHz |
| C | wiringPi | 4.1 – 4.6 MHz |
| Perl | BCM 2835 | 48 kHz |
In the last five or so years, large language models (LLMs) have transformed from a novel regurgitator of haphazardly stitched together sentences to an almost ‘human’ personality standing by our side as we tackle life. Whilst the perceived humanity of these models is the topic for perhaps a future blogpost, it is almost undeniable to understate the impact of LLMs in our daily lives. Do you need someone to proofread your essay you’ve spent hours drafting? GPT (or one of its many counterparts) has you covered. Need help drafting an email from scratch? No problem. Want to write and/or heavily edit an entire academic article which would typically require days, if not weeks, of research? Surely just needs a push of a button… right?
Despite tremendous advances in LLMs, key issues mean they are not yet a fully dependable addition to our writing endeavours. They are known to fail when asked to generate new content with only a basic prompt. Some of these failures have made headlines 1. Some of the scariest instances are those of hallucinated information 2–4 . This refers to the phenomenon where AI tools generate convincing information which is factually inaccurate or simply fabricated 2 . In Belgium, the Ghent university rector came under fire for citing quotes, supposedly from influential thinkers, which were later found to be AI-hallucinations 1.
Whilst there are numerous examples of the poorly cited and often AI-hallucinated papers falling through the cracks of the peer-review process, today we focus on a Frontiers in Microbiology review titled ‘Broadly neutralizing monoclonal antibodies against influenza A viruses: current insights and future directions’ 5. This paper attempts to provide an overview of the current landscape of monoclonal antibodies (mAbs) which are being developed to confer protection against influenza A, highlighting ‘technological advances, clinical performance, and scalability’. This paper contains many of the hallmarks of text that has been created or edited with generative AI, despite the generative AI statement stating ‘The author(s) declared that Generative AI was not used in the creation of this manuscript.’
A useful way to understand AlphaFold 3’s sampling behavior is to look not only at the final predicted structure, but at what happens along the reverse diffusion trajectory itself. If we track quantities such as the physical energy of samples, noise scale, and update magnitude over time, a very clear pattern emerges: structures remain physically imperfect for most of sampling, and only take proper global shape in the final low-noise steps.
This behavior is a result of the diffusion procedure implemented in Algorithm 18, Sample Diffusion, which follows an EDM-style sampler with churn. Rather than simply marching monotonically from noise to structure, the sampler repeatedly perturbs the current coordinates, denoises them, and then takes a Euler-like update step. Because of the churn mechanism, AlphaFold 3 deliberately injects additional noise during part of the trajectory, which encourages exploration but also delays local geometric convergence. This mechanism is shown in step 4 -7 of the Sample Diffusion Algorithm from Alphafold3 Supplementary Information.
Machine-learned interatomic potentials (MLIPs) have become a cornerstone of modern computational chemistry, enabling simulations that approach quantum accuracy at a fraction of the cost of traditional methods such as density functional theory (DFT). However, a central challenge in designing MLIPs lies in respecting the fundamental symmetries of molecular systems, especially rotational and translational invariance, while maintaining scalability and flexibility.
In our recent work, we introduced TransIP, a novel framework that formulates how symmetry is incorporated into molecular models by learning symmetry directly in the latent space of an atomic transformer model, in which we treat atoms as tokens, instead of hard-coding equivariance into the neural network architecture.
At the core of TransIP is a simple yet powerful idea: instead of enforcing SO(3) equivariance through specialized layers, the model is trained with a contrastive objective that aligns representations of rotated molecular configurations. A learned transformation network maps latent embeddings under rotations, encouraging the model to discover symmetry-consistent representations implicitly. This design preserves the flexibility and scalability of standard Transformers while still capturing the geometric structure of molecular systems.
Continue readingThis year we (Fabian and Henriette) were invited to speak at the Biologic Summit. The conference took place January 20-22 in San Diego. Fabian presented his work on conformational ensembles of antibodies [1] in the “Data Strategies and the Future of AI Models” track. Henriette presented her work on LICHEN [2], a tool to generate an antibody light sequence for a specific heavy sequence, in the “ML/AI for Biologics Developability, Optimization and de novo Design” track. Below we give some general highlights of the conference, and some talks we enjoyed. We would like to thank the organisers for the opportunity to discuss our research and hear about the latest developments in harnessing ML for design and optimisation of antibodies.