Integration of AI in Experimentation

Exciting to see AI agents highlighted as a “Method to Watch” in Nature Methods. https://lnkd.in/dipBKDPn Lin Tang's piece points to a future where AI agents do far more than automate lab tasks; they actively help generate hypotheses, design experiments, and accelerate the search for new scientific theories in biology. This resonates strongly with our own work. In the last years, we’ve begun developing active learning loops powered by perturbation models, where AI proposes experiments, wet labs execute them, and models improve iteratively. This hands-on experience makes me particularly excited about a future where robust multimodal foundation models can be supplied directly to AI agents, giving them a rich prior about biological systems. What inspires me most is reframing experimental design as a search through hypothesis space. AI doesn’t replace scientific intuition — it amplifies it. It helps us navigate complexity, connect molecular and cellular states to patient-level phenomena, and ultimately push biological understanding toward deeper causal theories. We’re only at the beginning, but I believe that providing reliable perturbation and system models to AI agents will meaningfully accelerate discovery across biomedicine and beyond.

Probably one of the most important articles in biotech AI this year...🧬👇 A new Nature Biomedical Engineering paper describes CRISPR-GPT, a large language model–driven, multi-agent system for automating CRISPR gene-editing workflows (link in the comments). ⚙️ The system coordinates multiple specialized AI agents to do these things: - Select appropriate CRISPR systems, design guide RNAs, and choose delivery methods. - Generate experimental protocols and assay plans. - Analyze results from wet-lab experiments and adapt subsequent steps. I can't stress enough the potential significance of this work, which IMO lies in the integration of computational reasoning with experimental execution, enabling “closed-loop” cycles where experiment design, execution, and analysis are connected and automated. The authors tested their agentic CRISPR-GPT system in real experiments, e.g., these: 🧬 Knockout of four genes in a human lung adenocarcinoma cell line. In the multigene knockout experiment targeting four genes (TGFβR1, SNAI1, BAX, BCL2L1) in A549 lung adenocarcinoma cells, the AI-generated protocol achieved consistently ~80% editing efficiency across all targets, as measured by NGS analysis 🧬 Activation of two genes in a human melanoma cell line. In the epigenetic activation experiments in a human melanoma cell line, the reported efficiencies were approximately 56.5% for NCR3LG1 and 90.2% for CEACAM1, based on flow cytometry comparing gRNA-edited groups versus negative controls. With this kind of agentic AI tools, it could become possible to explore larger experimental spaces more systematically and at greater speed. This really clicks with my vision of what the key role of AI systems is in drug discovery and biotech research: tying pieces together and allowing for holistic discovery vs classical reductionism (such as "target-ligand"-centric DD workflows). We outlined our ideas about this with Oleg Kucheriavyi earlier this year, in our industry report "Beyond Legacy Tools: Defining Modern AI Drug Discovery for 2025 and Beyond." The report is published with BioPharmaTrend.com, check it out via the link in the comments 📑 👇. Looking forward, systems like CRISPR-GPT could evolve into general-purpose lab intelligence, able to handle multi-modal data, integrate with robotics, and support continuous, iterative discovery. The space is worth watching. If you are following AI drug discovery and other deep tech trends, subscribe to Where Tech Meets Bio, a leading Substack newsletter in this niche (link in comments). Image from the article (citation in comments)

Self-driving labs can’t drive without data. Materials data are heterogeneous, sparse, and multi-scale. While compute power and model architecture are important, the underappreciated bottleneck in materials AI is data fragmentation. The field has datasets, yes ... but they live in silos, use inconsistent formats/metadata, span different scales, and often cannot be integrated easily. We’re typically working with datasets from tens to a few thousand samples, not the millions seen in image or language domains. Until we learn how to effectively integrate and fuse across modalities (experiment, simulation, processing), the promise of AI in materials will remain limited. Our new perspective, “Data integration and data fusion approaches in self-driving labs” (APL Machine Learning, Oct 2025 by Alexey Gulyuk, Nahed abu Zaid, Rada Chirkova, Yaroslava Yingling), reviews possible solutions: 🔹 Data integration brings order by harmonizing metadata and formats via ontologies, knowledge graphs, and FAIR principles. Emerging methods like federated learning and blockchain provenance extend this across labs while preserving data privacy and traceability. 🔹 Data fusion brings insight by combining spectroscopy, microscopy, and simulation outputs using Bayesian inference, graph neural networks, and physics-informed machine learning to uncover new patterns. Next-generation fusion frameworks now explore causal inference, reinforcement learning, and explainable AI for adaptive experimentation. 🔹 Together, they enable real-time reasoning and autonomous decision-making in #SelfDrivingLabs. Example: At the National Science Foundation (NSF) STEPS (Science and Technologies for Phosphorus Sustainability) Center, we’re building a phosphorus knowledge graph and fusion workflow that unifies adsorption kinetics, XPS/FTIR spectra, DFT-calculated binding energies, etc. Imagine a lab that detects when a phosphate-capturing material begins to degrade and autonomously adjusts regeneration protocols mid-experiment. That’s not science fiction; it’s data fusion in action. 📰 Featured in Scilight: Accelerating self-driving labs into the future by Ben Ikenson #AI #DataFusion #Materials #KnowledgeGraphs #Sustainability #Phosphorus #MachineLearning #SelfDrivingLabs #Scilight

AI is starting to challenge how we approach Design of Experiments (DoE) in industrial R&D. Traditional DoE requires you to define everything up front… factors, ranges, interactions. ML-guided optimization learns as it goes. Instead of fixing an experimental plan, the model watches results and decides where to test next, focusing on regions where behaviour is changing or still uncertain. Take Monolith AI's "Next Test Recommender" tool. After each batch, they fit a surrogate model, estimate uncertainty, and select the next points to maximize information gain. Classical DoE is planned exploration. ML-guided optimization is adaptive. But this only works if simulations are cheap. If runs are slow, licence-constrained, or organisationally painful, you don’t get enough data to train a useful surrogate. You fall back to intuition. That’s where infrastructure matters. Containerised FMUs (by the Modelica Association) remove licence bottlenecks. API-driven simulations let you parallelise, store, and replay every run. Tools like the Quix FMU runner are built for exactly this: turning simulations into something you can scale and automate. Now you have enough data for the model to actually learn. At that point, experimentation stops being manual. It becomes software that improves with every run. Less tweaking. More exploration. And once exploration is cheap, engineers stop asking “what should I tweak next?” and start asking “what parts of the design space haven’t we explored yet?”

I’m pleased to share a new publication on the “Current Opportunities for the Integration and Use of Artificial Intelligence and Machine Learning in Clinical Trials: Good Clinical Practice Perspectives.” This paper is the result of a cross-disciplinary working group of AI and clinical research experts convened by FDA’s Office of Scientific Investigations (OSI). The initiative reflects our attempt to assess the integration of AI/ML in clinical trials not just through the lens of technical performance but through Good Clinical Practice (GCP), inspectional oversight, and operational implementation at sites. While enthusiasm for AI continues to grow, its deployment in regulated clinical environments raises unique challenges related to data integrity, patient safety, and auditability. This paper offers a structured framework for addressing those concerns. Our key findings include: - AI/ML is already influencing trial design, monitoring, recruitment, and data capture; but formal governance and oversight remain inconsistent. - The current discourse often overlooks how AI affects real-world trial execution, particularly protocol adherence and inspection readiness. - The use of large language models (LLMs) in documentation and decision support is expanding rapidly, with limited guardrails. - Federated learning and privacy-preserving architectures offer promising alternatives to centralized data sharing. - Context-specific validation, not just general accuracy, is essential for safe, effective use in regulated settings. Based on these findings, we developed the following recommendations: - Align all AI/ML use in trials with GCP principles, ensuring traceability, transparency, and risk management. - Separate generative or adaptive systems from trial-critical decision pathways unless robust oversight is in place. - Establish clear SOPs, governance structures, and version control protocols for AI systems used by sponsors or sites. - Prioritize validation strategies tailored to the AI tool’s intended use, potential impact, and operational context. - Foster collaboration across stakeholders to build shared expectations for inspection readiness and responsible AI conduct. As AI becomes more deeply embedded in clinical research, structured, context-aware implementation will be critical. Our paper provides a foundation for moving forward responsibly as FDA continues to augment both its internal AI capabilities and its oversight mechanisms to advance national public health priorities. https://lnkd.in/dpbizggB

One of the most interesting shifts in pharma R&D right now is the emergence of continuous learning loops between computational models and lab experiments. Strategic investments are reinforcing this direction. A recent example is the $1B collaboration between NVIDIA and Eli Lilly and Company, aimed at building an AI factory for drug discovery, leveraging large-scale models trained on the language of biology and chemistry. At the core of this approach is a tight feedback loop between the wet lab and the dry lab, where experimental results continuously update computational models. Instead of the traditional discovery cycle: Hypothesis → experiment → analysis → new hypothesis AI enables something closer to: Model prediction → experiment → real-time data → updated model → next experiment This continuous loop allows research teams to iterate far more quickly. Industry analyses suggest that embedding AI directly into experimental workflows could reduce discovery timelines by as much as 40% in some cases. For pharma organizations, the implications are significant: • accelerating target validation • prioritizing experiments more effectively • reducing failed experimental cycles The companies that succeed may not simply use AI tools. They will build AI-native discovery systems in which computation and experimentation continuously inform one another. Article: https://lnkd.in/gzXqtj2Y #AI #DrugDiscovery #Pharma #Biotech #PrecisionMedicine#AI #DrugDiscovery #Biotech #PharmaR&D

When it comes to using AI in experimentation, here's the hill I'm willing to die on: 🔽 AI is genuinely useful in the creative, unbounded parts. When you're staring at a messy multi-metric readout, it can synthesise patterns, surface anomalies, and suggest plausible next iterations faster than you can alone. But the moment you let AI tell you what to *decide*? You've broken the discipline that makes experimentation worth doing. 🔽 I've seen teams (and platforms) use AI to "analyse results" and recommend a call. That's a failure mode. Here's why. Good experimentation means pre-registering decision criteria upfront. You run the stats. You check: did we meet the criteria or not? No AI needed. It's decision science and statistics, not storytelling. 🔽 If you're leaning on AI here, one of two things is true: → You never set clear criteria (you don’t know your goals and AI won't save you) → You're willing to override them (AI will happily spin stories and help you cherry-pick) Either way, it's not a technology problem. It's a discipline problem. 🔽 Use AI to explore faster. Never to decide. The hard work of experimentation, knowing what matters and holding yourself to it, is the one thing you can't automate.

Andrej Karpathy just dropped an open source repo called autoresearch that when you combine it with something like Claude Code becomes self-improving AI. Karpathy is widely regarded as one of the foremost voices in machine learning. He was a founding member at OpenAI and later head of AI at Tesla. So when he releases something, I pay attention. The idea is pretty elegant: You give an AI agent a small but real setup and let it experiment autonomously overnight. It modifies something, runs a test, checks if the results improved, keeps or discards the change, and repeats. You wake up in the morning to a log of experiments and hopefully a better outcome than what you started with. In his case, he was training a machine learning model. The agent would adjust hyperparameters, train for 5 minutes, measure validation loss, and loop. After just a few runs his model got significantly better without him doing anything. Now I'm not in machine learning. I take models other people have made and use them to make money. But the moment this dropped, I started thinking about ways to apply the same principle to my own business. To make a long story short, the requirement is simple. You need an objective metric you can track, an API you can call to measure it, and something the agent can modify between runs. If you have those three things, you can build an autonomous experimentation pipeline for almost anything. I've already set this up for my cold email campaigns. Reply rate is the metric, the Instantly API is the measurement tool, and the agent iterates on the email copy. It runs every hour via GitHub Actions. I'm not even in the loop anymore. I'm also running it on my Claude Code skills and I used it to take an old web app from 1,100ms load time down to 67ms across 67 autonomous runs. Landing pages, ad creatives, product descriptions, YouTube titles, pricing pages, chatbot scripts. If you can measure it and an agent can change it, autoresearch applies. I dove into more detail on YouTube: https://lnkd.in/gzWACx_i