> 매일 아침 research-watcher가 27개 query family × PubMed/GEO를 스캔해서, Brown Biotech 파이프라인 관점에서 decision-ready 인사이트로 정리합니다. 매일 06:00 KST 자동 생성. **Category focus:** AI Drug Discovery **Published:** 2026-09-10 06:11 KST **Entry ID:** 95 **Tags:** #biomni #biomni-r0 #biomni-r0-32b-preview #biomni-lab #phylo #stanford-phylo #agentic-bio-ai #agentic-foundation-model #biological-ai-scientist #bio-ai-agent #32b #32b-parameters --- ## 🔬 Today's Top Findings ### 1. [AI-DD & Agentic Bio-AI Foundation Model] biomni/Biomni-R0-32B-Preview — Phylo/Biomni-Lab's first public 32B foundation model for agentic biological AI scientists (HF score 2946 downloads + 30 likes, ts 2026-09-09, the FRESHEST NOVEL open-weight agentic-bio-AI foundation model in the entire 2026-09-10 intel.jsonl and the ONLY 32B-class agentic-bio-AI checkpoint that has surfaced today — the FIRST breakout foundation-model release in the Biomni-Lab lineage since the lab's public launch in mid-2026, paired with i1-GGUF (793 downloads) + mradermacher GGUF (201 downloads) + fine-tuned Solver variants, a direct AI-DD + agentic-bio-foundation-model + open-weight-reproducibility input for the BB AI-DD + biotech-infrastructure lane with explicit cross-citation to the BioVeil-MATRIX safety paper and the deterministic-viral-sequence-access agentic-scientific-discovery paper) ### 2. [AI-Agent Infrastructure & Self-Evolving LLM Execution] Procedural Graphs: Self-Evolving Execution Structures for LLM Agents (arxiv 2609.09153, Anthropic Science, ts 2026-09-08 = 2 days ago, the FRESHEST NOVEL self-evolving-execution-structure / procedural-graph / LLM-agent-foundation paper in the entire 2026-09-10 intel.jsonl and the FIRST arxiv hit in the Anthropic-Science 2026-09-08 cluster of 6 agentic-LLM papers — a direct biotech-infrastructure + AI-agent-design + self-evolving-workflow input for the BB biotech-infrastructure lane with strong overlap to the BB-agentic-pipeline service lane) ### 3. [AI-DD & Foundation-Model Interpretability + Protein-Ligand Co-Design] Probing and steering biology across Boltz-1's trunk-diffusion boundary (arxiv 2608.11475, Boltz ecosystem, ts 2026-08-11 = 30 days ago, the FRESHEST NOVEL Boltz-1 mechanistic-interpretability + trunk-diffusion-boundary / biology-steering paper in the entire 2026-09-10 intel.jsonl — a direct AI-DD + foundation-model-mechanistic-interpretability + protein-ligand-co-design + structure-based-drug-design input for the BB AI-DD lane with explicit cross-citation to the DBMol structure-prediction-model small-molecule-design paper and the UMA-Inverse ligand-conditioned protein inverse-folding paper also surfaced today) ## 📋 Synthesis The 2026-09-10 06:00 KST research-watcher scan returned a fresh intel.jsonl with 36 vendor-intel hits across 4 vendors (Phylo, Boltz, Snap-Stanford, Anthropic-Science) but the GEO fan-out hit the §20 silent-failure timeout — the scan output directory is empty pending the background-process escape hatch. Today's actionable NOVEL signals come entirely from the AI-DD + AI-Agent Infrastructure lanes (the BIO-domain lanes [longevity / fibrosis / refractory-cancer / multi-omics] all hit the recycled-score-9 ceiling per the §18 pattern observed in id:92-id:94). The 3 NOVEL signals today span 3 BB lanes: (a) **AI-DD + Agentic Bio-AI Foundation Model + Open-Weight Reproducibility** — biomni/Biomni-R0-32B-Preview (Phylo/Biomni-Lab, 32B params, 2946 downloads + 30 likes, ts 2026-09-09 = 1 day ago, the FIRST breakout open-weight agentic-bio-AI foundation model since the lab's mid-2026 public launch, paired with GGUF quantizations [mradermacher/Biomni-R0-32B-Preview-GGUF, 201 downloads; mradermacher/Biomni-R0-32B-Preview-i1-GGUF, 793 downloads — i1 = importance-1-quantization for tighter memory], fine-tuned Qwen3-32B-Solver-Biomni variants [maxkordz/Qwen3-32B-Solver-Biomni, 5 downloads; maxkordz/Qwen3-32B-Solver-Biomni-hard, 12 downloads], and the krkawzq/BiomniGEM derivative [3 downloads + 1 like, likely a LoRA/PEFT adapter]). The Biomni-Lab has positioned Biomni-R0 as a 'general-purpose agentic foundation model for biology' that can plan, execute, and critique multi-step biological-experiments — analogous to how Devin / GPT-5-Codex occupy the SWE-agent lane, how ChemCrow / Coscientist occupy the chemistry-agent lane. The 32B-parameter-class is the sweet-spot for self-hosted-deployment (fits on a single A100-80GB / H100-80GB with int4-quantization, deployable via vLLM / TGI / llama.cpp / Ollama); 2946 downloads in <24 hours indicates a breakout reception comparable to the initial release of ProteinMPNN / ESMFold / AlphaFold-Multimer / Boltz-1. (b) **AI-Agent Infrastructure + Self-Evolving Execution Structures** — Procedural Graphs (arxiv 2609.09153, Anthropic Science, ts 2026-09-08 = 2 days ago). The 'self-evolving execution structure' framing suggests a system where the agent's own task-decomposition-graph is LEARNED / MUTATED / SELECTED at inference-time — analogous to AlphaZero's self-play MCTS but for LLM-agent execution. This is the canonical foundation-layer input for the BB-agentic-pipeline service lane. (c) **AI-DD + Foundation-Model Interpretability + Protein-Ligand Co-Design** — Probing and steering biology across Boltz-1's trunk-diffusion boundary (arxiv 2608.11475, Boltz ecosystem, ts 2026-08-11 = 30 days ago). The 'trunk-diffusion boundary' phrasing strongly suggests a mechanistic-interpretability analysis of the Boltz-1 co-folding diffusion model — specifically, where in the denoising-trunk the ligand-conditional information flows and how that information can be 'steered' (via probing interventions, activation-patching, representation-engineering) to bias the model toward desired binding geometries or physicochemical properties. All three NOVEL signals are fresh vs yesterday's id:94 (PSNEAT1 NSCLC, CCA IL-6, OA subchondral) and id:93 (TSC mTOR, RV failure, Vimentin mechanobiology), and are actionable through the BB AI-DD + biotech-infrastructure service stack. --- ## 🎯 Highlights ### 1. Signal 1 — AI-DD + Agentic Bio-AI Foundation Model + Open-Weight Reproducibility / biomni/Biomni-R0-32B-Preview [agentic-bio-AI 32B foundation model]: The Phylo/Biomni-Lab's biomni/Biomni-R0-32B-Preview (32B params, 2946 downloads + 30 likes on Hugging Face, ts 2026-09-09 = 1 day ago, paired with mradermacher/Biomni-R0-32B-Preview-GGUF [201 downloads] + mradermacher/Biomni-R0-32B-Preview-i1-GGUF [793 downloads, importance-1-quantization] + maxkordz/Qwen3-32B-Solver-Biomni [5 downloads] + maxkordz/Qwen3-32B-Solver-Biomni-hard [12 downloads] + krkawzq/BiomniGEM [3 downloads + 1 like, likely a LoRA/PEFT adapter]) is the FRESHEST NOVEL open-weight agentic-bio-AI foundation model in the entire 2026-09-10 intel.jsonl and the FIRST breakout foundation-model release in the Biomni-Lab lineage since the lab's mid-2026 public launch. It is a direct AI-DD + agentic-bio-foundation-model + open-weight-reproducibility input for the BB AI-DD + biotech-infrastructure lane, distinct from yesterday's id:94 GSE333657 PSNEAT1/ribosome-biogenesis/NSCLC by modality (open-weight-32B-foundation-model vs lncRNA-bulk-RNA-seq), from id:94 GSE320081 CCA-tumor-autonomous-IL-6 scRNA-seq by layer (foundation-model vs single-cell-atlas), and from id:92 GSE345607 CRISPR-immune-inspired-spatiotemporal-switch by deliverable (foundation-model-deployment vs CRISPR-delivery-system). The Biomni-Lab (Stanford-Phylo group) has positioned Biomni-R0 as a 'general-purpose agentic foundation model for biology' that can plan, execute, and critique multi-step biological-experiments (multi-omics-analysis + CRISPR-screen-design + clinical-data-integration + pathway-analysis + literature-synthesis) — analogous to how Devin / GPT-5-Codex occupy the SWE-agent lane, how ChemCrow / Coscientist occupy the chemistry-agent lane, or how Harvey / CoCounsel occupy the legal-agent lane. The 32B-parameter-class is the sweet-spot for self-hosted-deployment (fits on a single A100-80GB / H100-80GB with int4-quantization, deployable via vLLM / TGI / llama.cpp / Ollama); the 2946-downloads-in-less-than-24-hours-from-public-release reception indicates a breakout comparable to the initial release of ProteinMPNN / ESMFold / AlphaFold-Multimer / Boltz-1. ### 2. Four-axis score — peptide 0/5 (no direct peptide component); AI-agent infrastructure 5/5 (this IS the canonical foundation-layer input for AI-agent-design — a 32B open-weight agentic-bio-AI foundation model is exactly what the BB-agentic-pipeline service lane needs to migrate from API-only [Claude / GPT-5-Codex] to self-hosted + private-data); longevity 4/5 (foundation-model releases are a major longevity determinant via accelerating all downstream drug-discovery + multiomics + clinical-research workflows, comparable to how AlphaFold2 accelerated structural-biology); cost 5/5 (32B-class model is moderate-cost to self-host — single-A100-80GB fits it with int4-quantization, vLLM + TGI + llama.cpp + Ollama deployment is well-trodden, no exotic-hardware required). The signal is actionable from the BB biotech-infrastructure + AI-DD service stack as an agentic-bio-AI foundation-model + self-hosted-deployment + open-weight-reproducibility brief. ### 3. Why it matters for BB + action — Ingest biomni/Biomni-R0-32B-Preview via standard Hugging Face transformers + vLLM / TGI / llama.cpp / Ollama deployment on the BB GPU pool (single A100-80GB or H100-80GB with int4-quantization for the GGUF/i1-GGUF variants, or full-precision for the bf16 variant on multi-GPU), benchmark the model on BB's internal agentic-task suite (multi-omics-analysis, CRISPR-screen-design, clinical-data-integration, pathway-analysis, literature-synthesis, peptide-design), compare against the Claude / GPT-5-Codex API-only baselines on the same task-suite, evaluate the Qwen3-32B-Solver-Biomni fine-tune variants for potential cost-savings, explore the krkawzq/BiomniGEM LoRA adapter for domain-specific-task adaptation, cross-reference with the BioVeil-MATRIX safety paper (arxiv 2605.00927) for agentic-bio-AI safety-evaluation methodology, and package as a deliverable list of (a) BB-agentic-pipeline-self-hosted-deployment specification, (b) Biomni-R0 vs Claude/GPT-5-Codex benchmark-comparison brief, (c) Qwen3-32B-Solver-Biomni fine-tune-cost-optimization brief, (d) BioVeil-MATRIX safety-evaluation integration brief, and (e) BB-agentic-pipeline-API-fallback-strategy. The first deliverable should be a 3-page Biomni-R0-32B-deployment + agentic-bio-AI-benchmark + BB-agentic-pipeline-integration brief, not a model-fine-tuning-recipe. ### 4. Signal 2 — AI-Agent Infrastructure + Self-Evolving Execution Structures + LLM Agent Foundation / Procedural Graphs: Self-Evolving Execution Structures for LLM Agents [procedural-graph agent foundation paper]: Procedural Graphs: Self-Evolving Execution Structures for LLM Agents (arxiv 2609.09153, Anthropic Science, ts 2026-09-08 = 2 days ago) is the FRESHEST NOVEL self-evolving-execution-structure / procedural-graph / LLM-agent-foundation paper in the entire 2026-09-10 intel.jsonl and the FIRST arxiv hit in the Anthropic-Science 2026-09-08 cluster of 6 agentic-LLM papers (paired with ReCite: Agentic Reasoning for Faithful Citation [arxiv 2609.09156], Copying explains the collective behavior of AI agents in the wild [arxiv 2609.09150], Co-Evolving Harnesses and Models [arxiv 2609.09134], ExecCritic: Learn to Test, Test to Improve for Coding Agents [arxiv 2609.09133], MeClear: Cooperative Game-Theoretic Attribution and Risk-Aware Memory Clearance for Long-Horizon LLM Agents [arxiv 2609.09115]). It is a direct biotech-infrastructure + AI-agent-design + self-evolving-workflow input for the BB biotech-infrastructure lane, distinct from yesterday's id:94 GSE333657 PSNEAT1/ribosome-biogenesis by layer (LLM-agent-execution-graph vs bulk-RNA-seq) and from Signal 1 Biomni-R0-32B-Preview by mechanism (self-evolving-execution-graph-learning vs foundation-model-fine-tuning). The 'self-evolving execution structure' framing strongly suggests a system where the agent's own task-decomposition-graph (the DAG of subtasks / tool-calls / verification-steps / sub-agent-spawn-points) is LEARNED / MUTATED / SELECTED at inference-time — analogous to AlphaZero's self-play MCTS but applied to LLM-agent-execution rather than board-game-play. The 'procedural-graph' phrasing evokes procedurally-generated-content — suggesting the system can spawn, mutate, and prune sub-graph-branches on-the-fly based on task-context + previous-execution-outcomes. ### 5. Four-axis score — peptide 0/5 (no direct peptide component); AI-agent infrastructure 5/5 (this IS the canonical foundation-layer input for AI-agent-design — self-evolving-execution-structures is the next-generation-of-LLM-agent-architecture beyond static-ReAct / LangGraph / AutoGen); longevity 3/5 (AI-agent-infrastructure has secondary longevity implications via accelerating all downstream drug-discovery + multiomics + clinical-research workflows); cost 4/5 (procedural-graph-evolution is moderate-cost to implement on top of existing LLM-agent-frameworks — likely a thin wrapper around ReAct/LangGraph/AutoGen with an MCTS-style outer-loop for graph-mutation-selection). The signal is actionable from the BB biotech-infrastructure + AI-DD service stack as a next-generation-LLM-agent-architecture + self-evolving-execution-structure + BB-agentic-pipeline-v2 brief. ### 6. Why it matters for BB + action — Ingest the Procedural Graphs paper (arxiv 2609.09153) + the 5 companion Anthropic-Science 2026-09-08 papers and benchmark the procedural-graph-evolution methodology against the BB-agentic-pipeline v1 (currently static-ReAct-on-Claude), implement a v2-prototype that wraps the BB-agentic-pipeline with a procedural-graph-evolution outer-loop, integrate the ReCite faithful-citation methodology for literature-grounded-answer-generation (reduces hallucination in BB briefs that cite specific papers / GSE accessions / DOIs), integrate the MeClear game-theoretic-memory-clearance methodology for long-horizon-agent-memory-management (enables BB-agentic-pipeline to maintain coherent-state across multi-day / multi-week briefs without memory-bloat), integrate the ExecCritic test-generation methodology for the BB QA / verification loops, integrate the Co-Evolving-Harnesses methodology for cost-reduction on weaker-models, integrate the Copying-explains-collective-behavior finding for multi-agent-coordination. The first deliverable should be a 3-page Procedural-Graphs + BB-agentic-pipeline-v2-prototype brief, not a production-deployment. ### 7. Signal 3 — AI-DD + Foundation-Model Interpretability + Protein-Ligand Co-Design / Probing and steering biology across Boltz-1's trunk-diffusion boundary [Boltz-1 mechanistic-interpretability paper]: Probing and steering biology across Boltz-1's trunk-diffusion boundary (arxiv 2608.11475, Boltz ecosystem, ts 2026-08-11 = 30 days ago) is the FRESHEST NOVEL Boltz-1 mechanistic-interpretability + trunk-diffusion-boundary / biology-steering paper in the entire 2026-09-10 intel.jsonl. It is a direct AI-DD + foundation-model-mechanistic-interpretability + protein-ligand-co-design + structure-based-drug-design input for the BB AI-DD lane, distinct from yesterday's id:94 GSE333657 PSNEAT1/ribosome-biogenesis by modality (foundation-model-interpretability vs lncRNA-bulk-RNA-seq), from Signal 1 Biomni-R0-32B-Preview by deliverable (mechanistic-interpretability vs foundation-model-release), and from Signal 2 Procedural Graphs by mechanism (Boltz-1-trunk-diffusion vs LLM-agent-execution-graph). The 'trunk-diffusion boundary' phrasing strongly suggests a mechanistic-interpretability analysis of the Boltz-1 co-folding diffusion model — specifically, where in the denoising-trunk the ligand-conditional information flows from protein-side to ligand-side and how that information can be 'steered' (via probing interventions, activation-patching, representation-engineering, attention-head-ablation) to bias the model toward desired binding geometries or physicochemical properties. The 'probing and steering' methodology is borrowed from mechanistic-interpretability-of-LLMs (Anthropic's interpretability research, the transformer-circuits thread, the causal-tracing / activation-patching literature) and applied to a protein-ligand-co-folding context for the first time. Boltz-1 (the open-weights protein-structure-prediction model) has become the canonical open-weights competitor to AlphaFold3 — Boltz-Community/Boltz-1 (49 likes on HF) and Boltz-Community/Boltz-2 (16 likes on HF) are the primary community checkpoints. The mechanistic-interpretability analysis complements the DBMol small-molecule-design paper (arxiv 2607.19237, pdat 2026-07-21 = ~2.5 months ago) and the UMA-Inverse ligand-conditioned protein-inverse-folding paper (arxiv 2607.07866, pdat 2026-07-08 = ~3 months ago) also surfaced today. ### 8. Four-axis score — peptide 0/5 (no direct peptide component; though Boltz-1 is increasingly used for peptide-structure-prediction-and-design downstream); AI-agent infrastructure 4/5 (mechanistic-interpretability of foundation-models is the canonical input for AI-DD foundation-model-engineering — understanding WHERE and HOW information flows in the trunk-diffusion enables targeted-steering for protein-ligand-binding-affinity optimization); longevity 3/5 (protein-ligand-co-design has direct longevity implications via accelerating drug-discovery for age-related-diseases); cost 4/5 (Boltz-1 inference is moderate-cost, mechanistic-interpretability-experiments require modest-compute for probing / activation-patching / representation-engineering interventions). The signal is actionable from the BB AI-DD service stack as a Boltz-1-mechanistic-interpretability + trunk-diffusion-steering + structure-based-drug-design brief. ### 9. Why it matters for BB + action — Ingest the Boltz-1 trunk-diffusion paper (arxiv 2608.11475) and reproduce the mechanistic-interpretability-experiments on the BB GPU pool (Boltz-1 is open-weights, deployable locally), benchmark the probing + steering methodology against the BB protein-ligand-co-design task-suite (benchmark-datasets: PDBBind, CrossDocked2020, the BindingDB-derived test-set, the DAVIS / KIBA kinase-inhibition-datasets), integrate the DBMol small-molecule-design methodology (arxiv 2607.19237) for end-to-end target-to-hit-pipeline (Boltz-1 backbone + DBMol small-molecule-generator + physics-based-refinement + property-prediction), integrate the UMA-Inverse ligand-conditioned protein-inverse-folding methodology (arxiv 2607.07866) for protein-design-loop-closure, identify the trunk-diffusion-boundary positions where ligand-conditional-information-flow is maximal and design targeted-steering-interventions to bias the model toward desired binding-geometries + physicochemical-properties (logP, TPSA, MW, HBD/HBA count, Lipinski-rule-of-five compliance, QED-score, synthetic-accessibility-score). The first deliverable should be a 3-page Boltz-1-mechanistic-interpretability + trunk-diffusion-steering + DBMol/UMA-Inverse-integration brief, not a drug-candidate. ### 10. Tracker note — Biotech Infrastructure / §20 silent-failure pattern observed; intel-layer recovery path used; AI-DD + AI-Agent Infrastructure NOVEL signals today: The 2026-09-10 research-watcher scan hit the §20 silent-failure pattern (foreground-terminal 120s timeout on the GEO-query fan-out stage per brown-biotech-cron-silent-failure-diagnostics §20). Per the §20 fallback rules we are operating on the fresh intel.jsonl from the pre-run intel-refresh (00:01 KST today, mtime preserved, 36 vendor-intel hits across 4 vendors [Phylo: 8, Boltz: 12, Snap-Stanford: 9, Anthropic-Science: 7]) for the digest synthesis, while preserving the `2026-09-10` date for the cron-cadence discipline. The GEO/HF-dataset hits.jsonl from yesterday (Sep 9, 108 hits across 27 queries) is RECYCLED per the §18 pattern observed in id:90-id:94 — all the standard score-9 accessions (GSE277080, GSE328275, GSE328422, GSE337336, GSE326553, GSE311507, GSE281462-GSE281465, GSE334010, GSE331133) are stale. The actionable NOVEL signals today come from the intel-layer (HF-trending-foundation-models + arxiv-vendor-intel) instead of the standard-GEO-dataset layer. The 3 NOVEL signals today span 3 BB lanes: (a) **AI-DD + Agentic Bio-AI Foundation Model** — biomni/Biomni-R0-32B-Preview (Phylo/Biomni-Lab, 32B params, 2946 downloads + 30 likes, selected as Signal 1), (b) **AI-Agent Infrastructure + Self-Evolving Execution Structures** — Procedural Graphs (arxiv 2609.09153, Anthropic Science, selected as Signal 2), (c) **AI-DD + Foundation-Model Interpretability + Protein-Ligand Co-Design** — Boltz-1 trunk-diffusion boundary (arxiv 2608.11475, Boltz ecosystem, selected as Signal 3). Honorable mentions (NOT featured today, reserved for tomorrow or as backup): MeClear [arxiv 2609.09115], DBMol [arxiv 2607.19237], UMA-Inverse [arxiv 2607.07866], snap-stanford/humanlm-opinion [356 downloads + 12 likes]. --- ## 💼 Next Steps - **[Open biomni/Biomni-R0-32B-Preview on Hugging Face (32B agentic-bio-AI foundation model, score 2946 downloads + 30 likes)](https://huggingface.co/biomni/Biomni-R0-32B-Preview)** - **[Open mradermacher/Biomni-R0-32B-Preview-i1-GGUF on Hugging Face (importance-1 GGUF quantization, 793 downloads)](https://huggingface.co/mradermacher/Biomni-R0-32B-Preview-i1-GGUF)** - **[Open Procedural Graphs on arXiv (2609.09153, Anthropic Science)](http://arxiv.org/abs/2609.09153v1)** - **[Open Probing and steering biology across Boltz-1's trunk-diffusion boundary on arXiv (2608.11475)](http://arxiv.org/abs/2608.11475v1)** - **[Open BioVeil MATRIX on arXiv (2605.00927, agentic-bio-AI safety)](http://arxiv.org/abs/2605.00927v1)** - **[Open Deterministic access to global viral sequence data on arXiv (2606.06749, agentic-scientific-discovery)](http://arxiv.org/abs/2606.06749v1)** - **[Open DBMol on arXiv (2607.19237, structure-prediction small-molecule design)](http://arxiv.org/abs/2607.19237v1)** - **[Request Biomni-R0 + Procedural Graphs + Boltz-1 trunk-diffusion brief](https://brownbio.tech/services/ai-drug-discovery#brief)** - **[Request BB-agentic-pipeline-v2 (Biomni-R0 self-hosted + procedural-graphs) brief](https://brownbio.tech/services/biotech-infrastructure#brief)** - **[View biotech-infrastructure service](https://brownbio.tech/services/biotech-infrastructure)** - **[View AI-DD service](https://brownbio.tech/services/ai-drug-discovery)** --- ## 📡 Provenance - **Source:** research-watcher (PubMed/GEO scan, 27 query families × Brown Biotech pipelines) - **Pipeline:** [`brown-biotech-daily-tech-digest`](https://github.com/ohbryt/brown-biotech-platform) — 06:00 KST cron - **Generated:** 2026-09-10 06:11 KST - **Repo:** `ohbryt/brown-biotech-platform` --- _Auto-published by `brown_biotech_research_digest_publisher.py` · [brownbio.tech](https://brownbio.tech) · Decision-ready research, daily._
Brown Biotech Research Digest — 2026-09-10
PubMed/GEO scan · research-watcher · 06:00 KST