Episode — Agents in Research

La Agente Óptima for Self-Driving Laboratories: Researchers developed La Agente Óptima, an agentic framework that oversees Bayesian optimization campaigns in automated experimental facilities. The system decouples model reasoning from regular campaign execution, querying the language model only when anomalies emerge or parameters require modification. In physical flow-chemistry and contact-angle trials, the platform flagged measurement errors, altered reagent compositions, and raised reaction yields from 30% to 59% with lower material consumption than human baselines.
ABLE Benchmark for Protein Design Agents: The Agentic BAIM-LLM Evaluation (ABLE) benchmark tests the capability of language model agents to operate biological AI tools, including ProteinMPNN and AlphaFold3. Across 15 tested frontier systems, seven models refused all tasks. Among models that completed the evaluations, Claude Sonnet 4 and Gemini 3 Pro obtained the highest scores in tool selection and retrieval, though models continued to display inconsistencies in long-horizon planning and biological context integration.
ELISA Framework for Single-Cell RNA Analysis: The Embedding-Linked Interactive Single-cell Agent (ELISA) connects generative pretrained transformer embeddings with semantic retrieval for interactive transcriptomic data analysis. The framework routes user requests to gene marker scoring, semantic matching, or reciprocal rank fusion to run pathway analyses, ligand-receptor interactions, and cell proportion estimates directly on embedded data. Across six test datasets, the system exceeded the cell-type retrieval accuracy of BM25 and CellWhisperer.
GEOMeta Pipeline for Automated Metadata Curation: GEOMeta uses a sequence of task-specialized agents to standardize records in the Gene Expression Omnibus repository. The workflow separates tasks into discrete stages: record retrieval, extraction, normalization, ontology mapping, and validation checks. Researchers used the system to annotate roughly 600,000 human bulk RNA-seq samples and evaluated 22 models on prospective record processing.
Educational AI Philosophy: Researchers proposed using Daoist philosophy as a conceptual framework for guiding artificial intelligence applications in educational settings.
Adverse Drug Reaction Identification: Researchers created AIPharm, a computational system designed to detect adverse drug reactions and classify their clinical severity levels.

Sources

La Agente Óptima: Towards Agentic Self-Driving Laboratories
Self-driving laboratories (SDLs) combine automated experimentation with adaptive decision-making to accelerate scientific discovery. Their operation nevertheless often depends on human specialists who translate scientific objectives into executable closed-loop campaigns. Specialists adjust them as data and operating conditions change. Here, we present La Agente Óptima, an agentic framework that constructs and supervises Bayesian optimization campaigns across computational and experimental systems while maintaining a persistent optimization state. By separating large language model (LLM)…
Agentic BAIM-LLM Evaluation (ABLE): Benchmarking LLM Use of Protein Design Tools
We introduce ABLE, a benchmark for evaluating LLM agents' ability to use biological AI models (BAIMs), such as ProteinMPNN and AlphaFold3, in dual-use protein design workflows. ABLE assesses agent performance through a set of tasks spanning structure retrieval, sequence generation, and design validation. We evaluate 15 frontier models and find that seven refuse all tasks, while the remaining models exhibit substantial performance differences. Claude Sonnet 4 and Gemini 3 Pro achieve the highest scores across information retrieval, tool selection, and tool use. We further compare model…
ELISA (Embedding-Linked Interactive Single-cell Agent): an interpretable hybrid generative Artificial Intelligence agent for expression-grounded discovery in single-cell genomics
Translating single-cell RNA sequencing (scRNA-seq) data into mechanistic biological hypotheses remains a critical bottleneck, as agentic AI systems lack direct access to transcriptomic representations while expression foundation models remain opaque to natural language. Here, we introduce ELISA (Embedding-Linked Interactive Single-cell Agent), an interpretable framework that unifies single-cell generative pretrained transformer expression embeddings with biomedical bidirectional encoder representations from transformers-based semantic retrieval and large-language model (LLM)-mediated…
Automating scientific annotations for open transcriptomic profiles via multi-stage agents
Abstract Public transcriptomic repositories contain millions of samples, yet their large-scale reuse is hindered by heterogeneous and inconsistently reported metadata. In the Gene Expression Omnibus (GEO), key biological information is often distributed across study- and sample-level records, requiring context-dependent interpretation. Here we present GEOMeta, a large language model (LLM)-based multi-stage workflow with task-specialized agents for automated GEO metadata curation. The pipeline separates metadata retrieval, task-specific information extraction, field standardization, ontology…

Also this week

Full transcript

Automating laboratory experiments requires software that can detect errors and modify physical procedures on the fly. How agent architectures supervise active experimental pipelines is what we track today on Agents in Research, reporting on artificial intelligence across the sciences. Here is the latest work. The architectural split in La Agente Óptima targets a specific failure mode in self-driving laboratories. It keeps the language model out of the repetitive parameter loops entirely. The model holds the optimization state, but intervenes only when data trends demand interpretation, an operational anomaly occurs, or the protocol itself needs revision. And that division showed measurable differences in physical hardware tests—flow-chemistry synthesis and contact angle optimization. It caught measurement errors, adjusted reagent formulations, and pushed chemical yields from 30 percent to 59 percent while burning through less material than human baselines. That dynamic of language models orchestrating wet-lab software carries straight into computational biology with the ABLE benchmark. ABLE evaluates models on dual-use protein design workflows, tracking whether agents can coordinate tools like ProteinMPNN and AlphaFold3 through structure retrieval, sequence generation, and final validation. The safety boundaries showed up immediately. Seven out of fifteen frontier models refused every task outright because of the dual-use framing. And for the systems that did run, Claude Sonnet 4 and Gemini 3 Pro led on tool selection and information retrieval. But the authors highlighted a recurring ceiling: the agents reduce the friction of executing commands, yet they still break down on multistep planning and tracking biological context. Handling biological representations directly, rather than just chaining software calls, is the focus of ELISA in single-cell transcriptomics. That stands for Embedding-Linked Interactive Single-cell Agent. It links single-cell GPT expression embeddings with transformer-based semantic search and language model output. An automated classifier sorts the user's prompt into gene marker scoring, semantic matching, or reciprocal rank fusion. From there, it calculates cell proportions, runs pathway analyses, and generates ligand-receptor predictions over the raw representations. Across six reference datasets, that pipeline surpassed CellWhisperer and BM25 on cell-type retrieval accuracy while matching the biological conclusions of the original papers. Which addresses exploratory analysis, but the raw input data itself remains a bottleneck—specifically the unstandardized metadata across repositories like the Gene Expression Omnibus. That is the gap GEOMeta targets. They split the aggregation pipeline across specialized agents for metadata retrieval, targeted extraction, field normalization, ontology mapping, and quality control verification. Using that multi-agent sequence, they generated standardized records for about 600,000 human bulk RNA-seq samples, testing 22 frontier models on prospective submissions to check extraction accuracy. Two other developments appeared in this batch: a framework proposing Daoism as a philosophical basis for artificial intelligence in education, and AIPharm, designed to detect adverse drug reactions and classify their severity. We will return next week with more reporting on autonomous systems across scientific discovery. That is all for today on Agents in Research.