The PharosBio blog
Plain-English guides to R&D data infrastructure: lab informatics, the systems that run modern laboratories, and the autonomous analysis layer emerging on top of them.
- AI for research
Best AI tools for scientists
Chatbots, literature AI, coding agents, research agents: what each tier of the 2026 AI toolkit actually does for a lab, and where validated data analysis comes from.
Read article - R&D data infrastructure
LIMS vs ELN vs SDMS vs LES
What each lab informatics system actually does, how the categories blur in practice, and where the software landscape still fails on data analysis.
Read article - Lab automation
Lab automation meets computational biology
AI reasoning agents can design experiments, but only the wet lab makes results true: the 8 automation levels, four ways to buy automation, ten companies to watch, and the closed loop.
Read article - AI in biotech
AI-native biotech in 2026
The three-layer biology-native data stack from Bessemer's market map, AstraZeneca's three tiers of AI adoption, the companies building each layer, and where the two views meet.
Read article - AI in biotech
From AI pilots to real ROI in pharma
Most pharma AI runs as free-floating tokens, not structured projects. The data from Benchling, MIT, and Bain, and the radical-collaboration principles that turn token burn into ROI.
Read article - AI for research
10 ways to use Claude for scientific work
Scheduled pipeline scans, life-science connectors, skills behind slash commands, projects, browser control, Cowork, and Claude Code: the ten workflows that take Claude beyond the chat box.
Read article - AI in biotech
How to validate agentic AI
Berkeley's failure taxonomy for multi-agent systems, translated into the six phases of a bioinformatics analysis: where agents break, what catches it, and what stays human.
Read article - AI in biotech
From autocomplete to autonomous colleague
Eight years, four capability eras, and the infrastructure decisions that mattered more than any single model: a practitioner's history of language models in the life sciences.
Read article - AI for research
Seven underused ways to use LLMs
Scored judgments, prioritisation, adversarial review, scheduled monitoring, literature triage, run-to-run variance, and agent panels: seven judgment tasks where models earn their keep, with the prompts and the published evidence.
Read article - AI for research
Identifier hell
Reasoning is no longer the scarce resource in agentic biology. Resolution is: the four layers of identifier failure, what the gget virus benchmark proved, and where a model belongs in the pipeline.
Read article - AI for research
Will AI replace bioinformaticians?
AI absorbs the pipeline work and leaves the judgement. Three transitions in the bioinformatics role, why multi-modal integration and explainability need experts most, and what that means for the tools you buy.
Read article - R&D data infrastructure
MCP for scientific data
Eight design rules for an MCP server over biological data, a bad and a good tool definition side by side, the failure modes that hide inside tool output, and the bio servers worth connecting today.
Read article - AI for research
Where the human checkpoint goes
A meta-analysis of 106 studies, a controlled experiment on research ideas, 70 teams analysing one dataset, and a 90% alert override rate. What the evidence says about where a human checkpoint belongs, and where it becomes theatre.
Read article - AI for research
When agents disagree
Three reasons the same question returns different answers, why majority voting fixes one of them and quietly ruins another, and how consensus, judge and validator agents actually earn their cost.
Read article - Target discovery
Inside-out proteins
Two 2026 papers independently put intracellular proteins on the outside of tumour cells, by two different mechanisms. The defining feature is an absence, which is exactly why surfaceome predictors miss them, and exactly what a classifier would have to encode.
Read article - Target discovery
When the cell lines are not the patients
DepMap has essentially one NPM1-mutant AML line for a genotype in roughly 30 percent of patients, and the lines labelled MDS were established after transformation. How to find synthetic lethal targets without them.
Read article - AI in biotech
Validating AI agents under GxP
FDA cited the quality unit clause, not a software rule. Europe proposed a ban and is walking it back. US banking superseded SR 11-7 and carved out agentic AI. Five controls for the gap they all left.
Read article - AI strategy
Where to dial AI governance up
Governance-everywhere is a tax paid on every use case to insure against the few that carry risk. How to set the tier by the consequence and reach of a wrong output, with the cases that priced each rule.
Read article - AI governance
EU AI Act delay: what still applies
The Digital Omnibus deferred the AI Act's high-risk chapter by sixteen months and softened Article 4. Neither change touches the obligations regulators are actually fining people for.
Read article - AI in biotech
AI-native pharmacovigilance
France codes patient reports at AUC 0.97, but median under-reporting is 94%. Seven rules from the evidence, including the two that the arrival of reasoning models has just reopened.
Read article - AI strategy
Can an agent take over CDMO paperwork?
Certificate of analysis review is the most automatable work in pharma supply chain and the least measured. Why the first agent should watch the coordinator rather than replace them, and what the regulation actually asks for.
Read article - AI strategy
Vendor-agnostic AI, seven layers
Most vendor-agnostic programmes are a router and nothing else, which is the one layer where switching was already cheap. Where the lock-in actually sits, what each layer costs to change, and how a French regulator let its data pick the model.
Read article