C2Bio

  FlashABB is a fast antibody structure prediction library from the Oxford Protein Informatics Group. (OXPIG) with a preprint here . It predicts paired VH/VL structures, scores, developability (FlashTAP), and produces structure-aware sequence embeddings (FlashABB-SSS) — all from a single "HEAVY|LIGHT" sequence string, with weights bundled in the repo. The library was written for CUDA. Getting it running on Apple Silicon required fixing five separate issues. I worked with Claude Code and got the following fixes:   --- Fix 1 — pip ≥ 26 editable install pip 26 dropped the legacy setup.py-only editable install fallback. A minimal pyproject.toml was all that was needed:   [build-system]   requires = ["setuptools>=64", "wheel"]   build-backend = "setuptools.build_meta"   ---   Fix 2 — Device auto-detection   Three places hardcoded device='cuda'. Replaced with a helper:   def _default_device():       if torch.cuda.is_available(): ...

Great to see that DeepSeek R1 is published as a research paper, which allows people to understand more about the engineering and design decisions made by their incredible team.  According to a comment by the editor of Nature, this paper was reviewed by 8 experts. (Most scientific journals require only 3 independent reviewers.) This reminds me of AlphaFold, a revolutionary protein model that can fold most proteins sequence into 3D structures with accuracy comparable to crystallography.  AF was published as a scientific paper followed by a Nobel prize. However, a lot of details are missing in the paper, especially training methods, which is crucial for fine tuning the original model (if it is open sourced). Open file mimics AF with full open source and well documented training methods, which allows scientists to understand the model in detail and appreciate the engineering marble in addition to being able to fine tune the model. Love to see more publications of mode details ...

A fundamental challenge in generative artificial intelligence involves sampling from carefully constructed high-dimensional latent spaces and utilizing these samples as inputs to decoder networks for generating novel entities. In the context of computational protein design, this process typically involves sampling regions within protein embedding spaces where specific biochemical properties are anticipated, such as enhanced binding affinity or improved developability characteristics in therapeutic antibodies. The sample-decode paradigm presents several significant technical challenges that must be addressed for effective protein generation. First, determining the optimal sampling distance from training datasets remains a critical consideration—sampling too close may limit diversity, while sampling too far may compromise biological relevance. Second, identifying which directions in the latent space merit more extensive exploration requires careful consideration of the underlying protein...

After tracking Nvidia’s RTX 50-series cards for a while, I still feel they are too expensive and likely not powerful enough to fully support my AI enthusiasm. As a result, I’ve decided to explore alternatives, particularly cloud solutions. Google Colab has turned out to be a very appealing option for now. It provides a Jupyter Notebook interface backed by CPUs, GPUs, and TPUs managed by Google. There is a free tier that offers limited GPU access (availability-based). During my test run this Sunday morning, I was able to access a GPU with 16 GB of memory. Paid users receive access to more powerful GPU and CPU resources. This makes Colab a viable alternative to running an on-premise AI lab—provided you’re comfortable with Google hosting your data and models. I haven’t checked AWS’s GPU offerings yet, but based on my prior EC2 experience, I feel Colab is much more convenient. Colab simplifies many aspects (from billing to system management), allowing you, as an AI scientist, to focus on m...

I was really impressed by Claude Code. For just $5 and under an hour, Claude scanned my project repository and generated the skeleton of a new functionality based on my prompt. While the code wasn’t error-free, the overall logic was sound. Reaching that point on my own would have taken me hours, if not days. About half of the errors could be resolved with additional prompting, and I spent roughly another hour fixing the remaining bugs and testing. In total, I had working, productive code in about four hours—work that would normally take me closer to twenty hours. In other words, that $5 effectively saved me 15 hours of effort!