iiinova — where your digital self keeps evolving

Testing AI-designed proteins isn’t proof of a learning loop. Our LinkedIn post examines Claude’s protein-binder campaign and the missing evidence: did experimental feedback improve the next round against a baseline? A connected workflow ≠ demonstrated learning. www.linkedin.com/feed/update/...

linkedin.com

#proteindesign #aidrugdiscovery #biotech #lifesciences | HealthyData.Science

Claude protein binder design: testing the designs is not the same as closing the loop. That distinction is important in the discussion Freda Duan has prompted about experimental feedback in AI drug di...

000

Stephen @healthydata.bsky.social · 05/10/2026

AI didn’t “find a cure” for iMCD. It ranked adalimumab as a candidate, and the result was worth testing because proteomics, single-cell RNA, and lab evidence also pointed to TNF signalling. One patient entered remission, but this is a hypothesis, not proof. www.linkedin.com/feed/update/...

linkedin.com

#drugrepurposing #raredisease #aiinhealthcare #drugdiscovery #translationalresearch | HealthyData.Science

adalimumab idiopathic multicentric Castleman disease machine learning: the interesting part is not that AI “found a cure.” It is that a model-generated drug ranking became worth taking seriously only...

010

Stephen @healthydata.bsky.social · 01/10/2026

From healthcare & life sciences: independent oversight is standard when safety is at stake. Voluntary self-policing + almost no independent scientific voices at the table feels insufficient. AI is already in drug discovery and clinical decisions. Real independent guardrails still matter.

000

Stephen @healthydata.bsky.social · 28/09/2026

OpenAI’s review of “misaligned” agent activity is a reminder: safety is not just about the model. The buyer question is: what can an agent access, do, send or change across connected systems? No confirmed breach of sensitive U.S. systems. #AgenticAI #AIGovernance

Blue and cyan infographic titled “OpenAI Agent Misalignment: Third-Party Impact.” It asks: “What can an agent access, do, send or change?” An important notice states that there is no confirmed breach of sensitive U.S. government systems or access to non-public data. Two panels summarise what is known and what it means: OpenAI is reviewing agent activity beyond intended scope, has notified dozens of organisations, and says it strengthened isolation and monitoring. The graphic explains that agents can use tools, credentials and multi-step actions to reach browsers, APIs and connected systems; independent evidence on frequency and safeguard effectiveness remains limited. It asks what minimum controls—such as role-based access, data boundaries, tool permissions, audit logs, human approval and a stop mechanism—are needed before an AI agent acts across a regulated workflow.

011

Stephen @healthydata.bsky.social · 27/09/2026

AlphaGenome Atlas may be the clearest AlphaFold comparison in genomics. Its value is not just scale: it could help researchers prioritise which DNA variants deserve follow-up. But a prediction score is a starting point, not proof of causation or clinical meaning. #Genomics #AIinHealthcare

Graphic about Google DeepMind’s AlphaGenome Atlas, a searchable genomics resource that predicts molecular effects for 9 billion possible single-letter DNA variants. It explains that the tool may help researchers prioritise variants for study, but prediction scores are not proof of causation, clinical significance or treatment decisions.

000

Stephen @healthydata.bsky.social · 26/08/2026

EU AI Omnibus delays AI Act high-risk rules for relevant medical devices and IVDs until 2 August 2028. That is more preparation time—not a pause on MDR, IVDR, validation, quality systems or AI governance. What pharma, medtech and health teams should assess now: healthydata.science/ai-regulatio...

EU AI Omnibus timeline showing high-risk AI Act requirements for relevant AI-enabled medical devices and IVDs deferred to 2 August 2028.

000

Stephen @healthydata.bsky.social · 17/08/2026

Can Palantir replace your healthcare or life-sciences AI vendor? It may unify data and enable enterprise-scale AI. But specialised workflows, regulated validation, auditability and domain expertise remain separate buying questions. A buyer-focused assessment: healthydata.science/can-palantir...

000

Stephen @healthydata.bsky.social · 03/08/2026

Counterfeit medicines don’t always look fake. I mapped out 5 AI tools that detect super‑fakes with hidden branding, mobile vision, and chemical/API verification. And built a decision matrix to help pharma security teams choose the right one. Read it here: healthydata.science/top-5-ai-too...

healthydata.science

Top 5 AI Tools for Identifying Fake Pharmaceuticals with Hidden Branding

AI-powered tools to detect fake pharmaceuticals with hidden branding. Compare Digimarc, VRAIAI, TrueMed, Cypheme and RxScanner for safer, smarter drug security.

000

Stephen @healthydata.bsky.social · 16/07/2026

Only ~5% of healthcare LLM studies used real patient data. Most “evidence” is clean exams and synthetic cases. If a vendor can’t show real‑world performance, treat their benchmark scores as marketing, not assurance. healthydata.science/from-pilots-...

Health AI in the Real World infographic showing key findings on adoption, safety, clinician impact, autonomy gaps, and trust from 2026 healthcare AI evidence.

021

Stephen @healthydata.bsky.social · 23/06/2026

We’ve overhyped generic LLMs in pharma. They’re great at text, terrible at traceability. The $100B unlock comes from healthcare‑native systems with built‑in regulatory controls and clear lineage. This piece explains the Pharma ‘X‑Factor’ for evaluating vendors: healthydata.science/the-100b-val...

Pharma executive in a glass-walled office pointing at a transparent screen that shows a chart titled ‘The $100B Value Unlock: Why Generic LLMs Can’t Pass the Pharma “X‑Factor” Test,’ comparing rising pharma‑specific LLM value with a flat generic LLM line and highlighting domain knowledge, regulatory alignment, auditability, and X‑Factor requirements.

040