AI

  • DeepSpot

    Byscientist

    Kalin Klonchev – the winner of a competition for AI based data analysis from Broad in 2024 had also created a tool called DeepSpot. Worth looking at for spot analysis of H&E sections by converting a full H&E slide pictures to “spots” which are analyzed. Some good links: DeepSpot paper: https://www.medrxiv.org/content/10.1101/2025.02.09.25321567v1 DeepSpot GitHub repository: https://github.com/ratschlab/DeepSpot…

  • AI Automations

    Byscientist

    The AI automations have only increased. There is one interesting one that has been receiving publicity. Check it out: https://knowledgework.ai It takes notes while the person is working and becomes the second brain. Privacy and access may be of concern but capability is available with AI tools.

  • Observability

    Byscientist

    Observability is important for AI and AI tools. It is the ability to monitor them for token usage, response quality and model drift. Typically, an AI system is monitored through logs, traces and metrics but an AI system on AI agent may need other metrics. Troubleshooting a complex AI system that produces its output probabilistically…

  • Error codes

    Byscientist

    There is no error code in the answers that are provided by AI prompts. It will return an answer that is the best fit to the prompt or the question, but it does not tell you the probability that it is not correct or that it is low probability of answer. The conversational AI will…

  • Judging art

    Byscientist

    Art especially with pencil and paper is the culmination of human esthetic and often times is dependent on the eye of the beholder. Some artwork that is meaningful to one person is not so understandable or sometimes even ugly to the other person. So how do you judge art. There are several ways to do…

  • |

    AI for drug discovery

    Byscientist

    There are many companies that promise AI for drug discovery such as ChatGPT from OpenAI, Anthropic’s AI suite and Perplexity and there are more companies coming up too. One specific company that is unusual is that it seems to be promising much more deeper understanding of the drug discovery pathways then the others. It can…

  • | |

    Functional prediction of microbial sequences

    Byscientist

    Even with E.coli and M.tuberculosis we only know ½ of them. Can you use ML model to define function: as natural language, or molecular interaction or chemical reaction. Function as molecular interaction: protein – protein interaction. Genomics: Learn association between genes (just like words). It is called gLM2. A multi-modal single residue resolution gLM. GLM2…

  • |

    Virtual cell

    Byscientist

    Hani Goodrazi – Arc institute has been working on virtual cell. Drugs fail due to overfit experimental models, You need screen drugs with better models of human biology. Geneva is platfrom that brings tumor models into perturbation model – that is a transcriptomics assay that deconvolves into effect. Take multiple cell line and then treat…

  • |

    Open source protein models

    Byscientist

    A company called Profluent (profluent.bio) has been developing protein models that can be used for designing new proteins (https://www.nature.com/articles/s41587-022-01618-2), modeling of new CRISPR-Cas sequences (https://www.nature.com/articles/s41586-025-09298-z) and developing LLM for protein generation (https://www.biorxiv.org/content/10.1101/2025.11.12.688125v1.article-info). What is amazing is that they have open sourced all their models and Profluent-E1 is available in GitHub to download and use. (https://github.com/Profluent-AI/E1)…