Skip to content

Coming up: Sat 31 Oct · Frontier Paper Club: DeepSeek-R1: reasoning learned through reinforcement learning; Wed 4 Nov · Build Night: Agents for Science: Build a literature agent that cites its sources; Sat 7 Nov · Frontier Paper Club: AI Co-Scientist: multi-agent hypothesis generation; Sat 14 Nov · Frontier Paper Club: Language agents that synthesise scientific knowledge; Sun 15 Nov · GPU 101: From Notebook to Accelerator: Profile a PyTorch training loop on a free GPU; Wed 18 Nov · Build Night: Agents for Science: Score genetic variants with Evo 2

Research insights

What we are reading, building on, and thinking about across AI for science, biology foundation models, and clinical AI. Every item links to its source.

Lab notes

  • Verification is the bottleneck now, not generation

    Co-Scientist, Robin, and AlphaEvolve all succeed where an automatic check, an assay, or a clinician can judge the output. For a small lab the lesson is to spend effort on evaluators, citation checks, and validation plans before building more agents.

  • Open weights changed who can do this research

    Between DeepSeek-R1, gpt-oss, Evo 2, Boltz, and ESM, the models that matter for AI-for-biology can be downloaded, studied, and fine-tuned. That moves the advantage from owning compute to asking good questions and measuring carefully.

  • Cheaper tokens favour patient, long-running work

    Each NVIDIA generation from Blackwell to Rubin lowers the cost of inference. Agents that read for hours or re-check their own citations become affordable, which suits careful scientific work more than quick chat.

Written by the ASL Life Sciences research team.

Papers we’re reading

Several are Paper Club picks this season
  1. Agents for sciencearXiv 2502.18864 · 2025
    Accelerating scientific discovery with Co-Scientist

    Gottweis et al., Google

    A generate, debate, and evolve loop of agents that proposes hypotheses. The interesting part is the wet-lab validation, not the benchmark.

  2. Agents for sciencearXiv 2505.13400 · 2025
    Robin: A multi-agent system for automating scientific discovery

    Ghareeb et al., FutureHouse

    Closes the loop from literature to experiment design to data analysis. Worth reading for where it still needed people.

  3. Agents for sciencearXiv 2409.13740 · 2024
    Language agents achieve superhuman synthesis of scientific knowledge

    Skarlinski et al., FutureHouse

    PaperQA2. Cited, checkable answers over the literature. The evaluation design is as useful as the system.

  4. Agents for sciencearXiv 2408.06292 · 2024
    The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

    Lu et al., Sakana AI

    An end-to-end pipeline that writes its own papers. Read it with its limitations section open beside it.

  5. Agents for sciencearXiv 2506.13131 · 2025
    AlphaEvolve: A coding agent for scientific and algorithmic discovery

    Novikov et al., Google DeepMind

    Evolutionary search over code, scored by automatic evaluators. Shows what agents can do when the result is verifiable.

  6. ReasoningarXiv 2501.12948 · 2025
    DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning

    DeepSeek-AI

    Open weights and an open recipe for reasoning. The reason small labs can now study reasoning models directly.

Latest on arXiv

Newest AI preprints in biomolecules and genomics (q-bio.BM, q-bio.GN), pulled from the arXiv API on 2026-10-28. Preprints are not yet peer-reviewed.

  1. 2026-10-19 · cs.AIMagnolia: A Persistent-Memory Agent Framework for Computational ChemistryTao Jiang, Benjamin Navet, Yousouf Taghzouti et al.
  2. 2026-10-17 · q-bio.BMAptamers as a sandbox for lab-in-the-loop agentic molecular discoveryZachary Cutts, John Coller, Rachel Chen et al.
  3. 2026-10-15 · q-bio.GNHarness-Driven Agentic Operation of Optical Pooled Screen AnalysisMatteo Di Bernardo, Claire Peterson, Ege Topkoc et al.
  4. 2026-10-15 · q-bio.BMEvolutionary Informed Translation from Protein to Nucleic Acid SequenceAnna Hart, Mohammed El-Kebir, Huimin Zhao et al.
  5. 2026-10-15 · q-bio.GNPanoramic modelling of single-cell transcriptomes at scale with CellNetXiaoqing Lian, Pei Liu, Jiahui Tan et al.
  6. 2026-10-13 · q-bio.MNDeep learning-based virtual perturbation reveals increasing heterogeneity of gene regulatory responses with agingYongjian He, Julia Hanna Haddad Guez, Xu-Wen Wang et al.
  7. 2026-10-12 · cs.DBCellChain: Benchmarking Cross-Scale Evidence Use for Perturbation ReasoningXue Xia, Yue Wang, Shuizhou Chen et al.
  8. 2026-10-11 · cs.CRTraining-Data Extraction from Genomic Language ModelsTianhao Mao, Zihao Wang, Dongfang Zhao et al.

Open-source tools we build on

Open models and code from other labs that our programs use. Credit and licences belong to their authors.

Read these with us at Frontier Paper Club, on Saturdays this season.

See the calendar