From autocomplete to autonomous colleague: LLMs in biology

This is a practitioner’s history, written for people who use these systems in biology rather than build them. For each era: what shipped, what new capability arrived, and the part usually skipped, what it unlocked for the life sciences. One scope note before the figure: this is a history of text-based language models and the agentic systems built on them. Structure-prediction models such as AlphaFold and ESM, and AI drug discovery built on other model families, are parallel and equally important lineages, but different stories.

Expand

Citations use author and date; full references for every entry are listed in the Sources section below.

GPT-1 (Radford et al., 2018) established pretrain-then-finetune. GPT-2 showed fluent open-ended generation. GPT-3 (Brown et al., 2020), at 175 billion parameters, showed that scale alone produced few-shot in-context learning. These were completion-only models: they extended text. No instruction following, no dialogue, no tools, no grounding.

Biology adapted the technique rather than the models. BioBERT (Lee et al., 2019) was the first BERT pretrained on biomedical corpora, trained for 23 days on eight V100 GPUs over PubMed abstracts and PubMed Central full text, and it substantially beat vanilla BERT on biomedical named entity recognition, relation extraction, and question answering. SciBERT trained from scratch on Semantic Scholar; PubMedBERT built its vocabulary from PubMed alone; BioGPT brought generative modeling to biomedical text. These were the workhorses of literature mining: tag genes, diseases, and chemicals; extract stated relationships. They could not plan, reason across steps, or run code.

The era’s honest summary: superb pattern recognition over text, zero autonomy.

The pivot was alignment. InstructGPT applied reinforcement learning from human feedback so models followed instructions rather than merely continuing text. That produced ChatGPT (OpenAI, 2022) and GPT-4 (OpenAI, 2023). In medicine, Med-PaLM passed a USMLE-style threshold on MedQA at 67.6 percent; Med-PaLM 2 (Singhal et al., 2023) reached 86.5 percent, and in blinded evaluation physicians preferred its long-form answers to other physicians’ on eight of nine clinical axes.

But fluency exposed the failure mode that matters most in science: fabrication. Two weeks before ChatGPT, Meta released Galactica, a model trained on 48 million papers, textbooks, and encyclopedias. The public demo launched on 15 November 2022 and was pulled on 17 November, surviving, as MIT Technology Review put it, three days online. Researchers had demonstrated it producing authoritative-sounding, confidently wrong, and sometimes biased scientific text, including invented citations.

Galactica is the most instructive failure in this history. It was not undertrained or insufficiently scientific. It was ungrounded. A model that generates a plausible reference is more dangerous to science than one that generates nothing.

The fix predated the problem. Retrieval-augmented generation was introduced by Lewis et al. (2020): couple a generator to a retriever over an external corpus so answers are conditioned on retrieved evidence and carry provenance. RAG went mainstream in 2023 and powered the literature-tool wave (Elicit, Consensus, Scite, SciSpace). The most ambitious was PaperQA2 (Skarlinski et al., 2024), which reported matching or exceeding PhD-level biologists on realistic literature tasks: question answering, summarization, and contradiction detection.

The lesson generalizes: grounding plus an agentic retrieval loop converts an unreliable generator into a verifiable instrument. Everything after this era is a variation on that theme.

Execution. Toolformer (Schick et al., 2023) showed models could learn to call APIs. OpenAI shipped function calling in June 2023, and Code Interpreter let a model write and run Python in a sandbox. For bioinformatics this was the difference between describing an analysis and performing one: drive scanpy or Seurat, run the statistics, produce the figures. ChemCrow (Bran et al., 2023) wired GPT-4 to expert chemistry tools; Coscientist (Boiko et al., 2023) combined search, code execution, and robotic lab automation to design and run experiments, including optimizing palladium-catalyzed cross-couplings.

Reasoning.By mid-2024 benchmarks were saturating and next-token scaling was returning less. The unlock was inference-time reasoning: models trained to think at length before answering. OpenAI’s o1 (2024) and o3 posted large gains on graduate-level science questions, and DeepSeek-R1 (2025) delivered comparable reasoning with open weights, which matters for labs that need local deployment over sensitive data.

The best frame for what reasoning bought is not accuracy but duration. Kwa et al. (2025), working at METR, defined a model’s task-completion time horizon and found it doubling on a schedule.

That is why the plateau story was wrong. Accuracy on fixed question sets stalled; the length of work a system could finish kept climbing. In METR’s January 2026 update, Claude 3.7 Sonnet sits at 60 minutes, o3 at 121, GPT-5 at 214, and Claude Opus 4.5 at 320. A system that reliably completes five hours of professional work is a different kind of tool than one that completes thirty seconds of it, even at identical benchmark scores. One caveat travels with the metric: it is measured on software tasks, and scientific work has failure modes that software does not.

With reasoning and tools in place, the unit of work shifted from the prompt to the loop: plan, call tools, observe, revise, persist.

Science produced a cohort of AI scientists. Sakana’s AI Scientist ran an end-to-end machine-learning research loop. FutureHouse, and its spinout Edison Scientific, shipped multi-agent research systems including Robin (Ghareeb et al., 2025) and Kosmos (Mitchener et al., 2025). The most striking demonstration came from Google’s AI co-scientist (Gottweis et al., 2025): given a short prompt, it proposed that capsid-forming phage-inducible chromosomal islands acquire tails from other phages to spread between bacterial species, reproducing in about two days a conclusion that José Penadés’s team at Imperial College had reached over roughly a decade of work on antimicrobial resistance. Penadés emailed Google to ask whether they had access to his computer. They did not. Worth stating plainly: the model worked from published literature, the team’s own hypothesis was unpublished at the time, and this was a retrospective test against an already-solved problem rather than a blind prospective discovery.

Two other unlocks belong to this era. Computer use (Anthropic, 2024) let agents operate screens and graphical interfaces, which matters enormously in science because so much tooling lives behind web portals with no API. And standards solved the problem that had quietly capped everything else.

The fragmentation problem. Biology runs on thousands of Bioconductor packages, nf-core pipelines, the scverse ecosystem, GATK, and databases from UniProt to ClinicalTrials.gov, each with its own interface. Connecting M models to N tools is an M-times-N integration explosion. The Model Context Protocol (Anthropic, 2024) reduced it to M plus N: implement the client once and the server once, and any model can call any tool. By December 2025 Anthropic reported more than 10,000 active public MCP servers and over 97 million monthly SDK downloads, and donated the protocol to the Agentic AI Foundation under the Linux Foundation, co-founded with Block and OpenAI. Agent Skills (Anthropic, 2025), published as an open standard that December, did the same for procedures rather than connections: a folder of instructions and scripts that teaches an agent how a piece of work is done, loaded only when needed.

The measurement that captures it. BixBench (Mitchener et al., 2025) tests agents on 53 real analysis scenarios and 296 open-answer questions rather than trivia. At launch the best frontier model reached 17 percent on open-answer questions, against GPT-4o’s 9 percent, with multiple choice no better than random. A year later the reported best is 80.5 percent. Two caveats belong with that number: it is self-reported by OpenAI for GPT-5.5 and is the only entry on the public leaderboard, with no independent replication and no published detail on the evaluation regime. The honest reading is that bioinformatics agents went from near-useless to genuinely capable in about a year, and that evaluation infrastructure has not kept pace with the systems it measures.

Three consequences follow, and none of them is “AI does science now.”

The bottleneck moved to validation. Hypothesis generation is approaching free. Experimental confirmation is not. A team that generates a thousand agent-proposed hypotheses without a disciplined way to prioritize and falsify them has manufactured a backlog, not a pipeline. The scarce resource is falsification capacity, and the skill that matters is choosing which of many plausible ideas deserves a wet-lab week. This is the Galactica lesson at portfolio scale, and we wrote about the practical version of it in how to validate agentic AI in bioinformatics.

The target space got wider, not just faster. The most underrated near-term effect is breadth. Agentic search lets a team explore more candidate targets, indications, and mechanisms in parallel than any human group could hold. Even with no single breakthrough, broader coverage mechanically yields more true positives.

Expertise became packageable. This is where the standards story turns practical. The judgment that separates a competent analysis from a misleading one (which threshold, which normalization, which covariate, which database version) has always lived in a few people’s heads. Skills let that judgment be written down once and executed consistently.

That is the pattern behind Hydra, our own autonomous research platform: bioinformatic expertise codified as 200+ skills over ~100 scientific databases including DepMap, TCGA, and cBioPortal, covering work like synthetic-lethality prediction, CRISPR screen analysis, and codependency interpretation. It is also something we publish on rather than only assert. In a January 2026 preprint, our team showed that open-weight language models reproduce CRISPR-screen-based predictions of synthetic lethal interactions zero-shot, that performance scales with model size, and that the best cost-efficient model reached 0.715 AUROC, which we used to screen 398,277 gene pairs across 893 clinically relevant genes in silico. In applied work with a partner, the same approach reached roughly 84 percent specificity across 37 drugs spanning more than ten disease areas; that pilot is written up in our in vivo augmentation case study.

• Static benchmarks decay fast. BixBench moved 17 to 80.5 in a year; that says as much about evaluation as about capability. Real-task measures like the time horizon age better than fixed question sets.
• Vendor-reported numbers are claims, not results. The 80.5 percent figure has no independent replication and no published evaluation regime.
• Impressive anecdotes are not general reliability. The co-scientist case was retrospective. Berkeley's A-Lab, which ran autonomously for 17 days, originally reported 41 new compounds from 58 targets; after criticism of its automated structural analysis the paper was corrected in January 2026 to 36 compounds from 57 targets, with 4 of 40 reported successes inconclusive from X-ray diffraction alone.
• Even flagship autonomy claims need reading carefully. Sakana's AI Scientist-v2 generated three papers for an ICLR workshop in 2025; one was accepted, and Sakana withdrew it before publication by prior arrangement, noting that its own reviewers judged none of the three to meet main-conference standards.
• Public data is finite and biased toward positive results. Negative results and proprietary assay data, the very data that would most improve these systems, are largely locked up.
• Judgment stays human. What matters, what is real, and who is accountable.