> 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:** 2026-09-29 06:07 KST **Entry ID:** 110 **Tags:** #biomni #Biomni-R0-32B-Preview #biomni-R0 #biomni-R0-preview #stanford-biomni-lab #generalist-biomedical-AI-agent #biomedical-LLM #32B-parameter-LLM #Qwen3-32B-base #tool-augmented-LLM #tool-use-agent #agentic-AI-biology --- ## 🔬 Today's Top Findings ### 1. [AI Drug Discovery + Generalist Biomedical AI Agent + Biomni-R0-32B-Preview + Stanford Biomni Lab + LLM-as-Scientist + Tool-Use Agent + Bioinformatics Coding Automation] biomni/Biomni-R0-32B-Preview — first open-weights 32B generalist biomedical AI agent from Stanford Biomni lab (HuggingFace, downloads 1638, likes 30, ts 2026-09-28 = 1 day ago, score 1638, kind hf_tag/vendor phylo, URL https://huggingface.co/biomni/Biomni-R0-32B-Preview) — the FIRST Biomni-R0-32B-Preview release to surface in the research-watcher intel.jsonl and the FIRST preview of the R0 model family on HuggingFace, accompanied by three derivative artifacts in the SAME 2026-09-29 intel.jsonl cohort (mradermacher/Biomni-R0-32B-Preview-GGUF downloads 415, mradermacher/Biomni-R0-32B-Preview-i1-GGUF downloads 1422, maxkordn/Qwen3-32B-Solver-Biomni + maxkordn/Qwen3-32B-Solver-Biomni-hard downloads 18+18) — a direct AI Drug Discovery + AI Agent infrastructure + generalist-biomedical-LLM + tool-use agent + LLM-as-biomedical-scientist input for the BB AI-DD + biotech-infrastructure lane with explicit overlap to the Biomni-paper / Stanford-Biomni-lab / generalist-biomedical-AI-agent / tool-augmented-LLM / ECR-LLM-bio-research-agent / CodeInterpreter-biology / Python-Bioinformatics-pandas-scanpy-cellar-Seurat / HuggingFace-Agent / clinical-trial-design-LLM / molecular-property-prediction-LLM / drug-repurposing-LLM / CRISPR-guide-RNA-design-LLM / cell-type-annotation-LLM / scientific-coding-agent / agentic-AI-for-biology / biomedical-data-curation-LLM / literature-mining-LLM / multi-omics-analysis-LLM / pathway-enrichment-LLM lanes ### 2. [Bioinformatics & Multi-Omics + Single-Cell Data Valuation + CellRater + JAX-based Data-Quality Scoring + MIT-Licensed + Data Curation for AI-DD Training] snap-stanford/cellrater-metadata + snap-stanford/cellrater-gex — first open-source CellRater data-valuation method for single-cell datasets (HuggingFace, downloads 29+15, likes 0+0, ts 2026-09-28 = 1 day ago, scores 29+15, kind hf_org/vendor snap-stanford, URLs https://huggingface.co/snap-stanford/cellrater-metadata + https://huggingface.co/snap-stanford/cellrater-gex, tags cellrater+single-cell+data-valuation+jax+license:mit) — the FIRST CellRater single-cell-data-valuation release to surface in the research-watcher intel.jsonl and the FIRST data-valuation method specifically targeting single-cell-RNA-seq dataset quality scoring — a direct bioinformatics + multi-omics + single-cell + data-valuation + AI-DD-training-data-curation input for the BB bioinformatics + AI-DD + multi-omics lane with explicit overlap to the single-cell-data-quality / CellRater-Shaham-SNAP-Stanford / data-valuation-for-biomedical-AI / training-data-pruning-single-cell / influence-function-single-cell / data-shapley-single-cell / LoRA-data-selection / DataComp-for-biology / biologically-meaningful-data-subset-selection / CellTypist-training-set-curation / scVI-scANVI-training-curation / foundation-model-pre-training-data-curation-biology / CellPLM / scGPT / Geneformer-pretraining-data-quality / GeneCompass-training-data / UCE-Universal-Cell-Embedding-training-data / scFoundation-training-data / scCAPE / scEval-data-quality-benchmark lanes ### 3. [AI Drug Discovery + AI Interpretability + Mechanistic Biology + Sparse Autoencoder + Circuit Discovery + Boltz / Affinity-Generalization] 'Efficient Auto-Interpretability of AI Models in Biology' arXiv 2608.27754v1 (arxiv.org, score 0, ts 2026-08-27 = 33 days ago but FRESH in the 2026-09-29 intel.jsonl as a NOVEL biology-AI-interpretability method via the boltz-vendor-tracked auto-interpretability track) — the FIRST explicit auto-interpretability-for-biology method paper surfaced in the boltz vendor intel stream and a direct AI-DD + AI interpretability + mechanistic-biology-understanding + sparse-autoencoder-for-biology lane input with explicit overlap to the Anthropic-claude-auto-interpretability / sparse-autoencoder-SAE-biology-features / circuit-discovery-protein-language-model / mechanistic-interpretability-protein-foundation-model / mechanistic-interpretability-Evo2-Evo-1-DNA-language-model / mechanistic-interpretability-Boltz-2 / mechanistic-interpretability-AlphaFold-3-features / mechanistic-interpretability-RNA-language-model-RNA-FM / mechanistic-interpretability-for-drug-discovery / interpretable-virtual-screening / mechanistic-understanding-of-affinity-prediction lanes ## 📋 Synthesis The 2026-09-29 06:01 KST research-watcher run.sh scan was triggered but the GEO/NCBI fan-out across 27 query families did not produce hits.jsonl within the cron turn budget (the scan started cleanly with `[brown-biotech] watcher start` then timed out before GEO writes completed) — per brown-biotech-cron-silent-failure-diagnostics §20, the GEO output is ABSENT today (no 2026-09-29/scan/scan.json or hits.jsonl files were produced) and the silent-failure pattern is HONESTLY FLAGGED in this summary rather than fabricated. The HF-only probe (`WATCHER_SOURCES=hf bash run.sh scan`) ran cleanly in <90 seconds but does not populate the latest/intel.jsonl (run.sh only copies hits.jsonl + scan.json + scan.md, not intel.jsonl). Today's actionable signals therefore come from the PRE-EXISTING 2026-09-29 intel.jsonl populated by the 00:00:47 KST intel-probe cron (38 entries, 4 vendors: phylo / boltz / snap-stanford / anthropic-science) — this represents the only fresh intel collected today and contains all 3 picks. The 2026-09-29 intel.jsonl cohort is DISTINCT from yesterday's 2026-09-28 id:109 GEO-dominant scan (which featured GSE284422 CRISPR-Genuine-seq + GSE286186 glomerular-snRNA + GSE345645/6 RV-failure) — ZERO overlap on accession or model identity, so all 3 picks today are TRULY NOVEL via the regex-dedup novelty filter. The 2026-09-29 intel.jsonl is dominated by a single cluster — the Biomni-R0-32B-Preview generalist biomedical AI agent release from Stanford Biomni lab — which appeared as 6 distinct artifacts (biomni/Biomni-R0-32B-Preview itself with 1638 downloads + 30 likes, plus three mradermacher GGUF quantizations at 415 + 1422 downloads, plus two maxkordn fine-tunes at 18 + 18 downloads). Three NOVEL signals today span 3 distinct BB lanes: (1) **AI Drug Discovery + Generalist Biomedical AI Agent + Biomni-R0-32B-Preview + Stanford Biomni Lab** — biomni/Biomni-R0-32B-Preview (HF, downloads 1638, likes 30, ts 2026-09-28 = 1 day ago, the FIRST R0-Preview release on HuggingFace from the Stanford Biomni lab, selected as Signal 1 — the highest-impact and highest-velocity signal of the day, anchoring the generalist-biomedical-AI-agent + tool-augmented-LLM + scientific-coding-agent + multi-omics-analysis-LLM lane), (2) **Bioinformatics & Multi-Omics + Single-Cell Data Valuation + CellRater + JAX-based Data-Quality Scoring + MIT-Licensed** — snap-stanford/cellrater-metadata + snap-stanford/cellrater-gex (HF, downloads 29+15, ts 2026-09-28 = 1 day ago, MIT-licensed JAX-based data-valuation method for single-cell-RNA-seq dataset quality scoring, selected as Signal 2 — the FIRST CellRater release in intel.jsonl and the FIRST data-valuation method specifically targeting single-cell dataset quality), (3) **AI Drug Discovery + AI Interpretability + Mechanistic Biology + Sparse Autoencoder for Biology** — arXiv 2608.27754 'Efficient Auto-Interpretability of AI Models in Biology' (arxiv, ts 2026-08-27 = 33 days ago but FRESH in 2026-09-29 intel.jsonl via boltz-vendor auto-interpretability track, selected as Signal 3 — the FIRST explicit biology-AI-interpretability method surfaced in the boltz intel stream and the FIRST sparse-autoencoder-for-biology / circuit-discovery-for-protein-language-model style paper to enter the BB intel feed). All 3 picks are ABSENT from the id:101-id:109 retention window per regex-dedup against src/app/digest/page.tsx (NO GSE-accession or model-name collisions with prior 9 days of entries), ABSENT from the explicit HF-model-name screen (no 'biomni', 'cellrater', 'auto-interpretability', 'biomni-R0' keyword collisions in prior entries), and span 3 distinct BB lanes (AI-DD/agent-infra + bioinformatics/multi-omics + AI-DD/interpretability). Combined 4-axis score (peptide / AI-DD / longevity / cost) is approximately peptide=0/5, AI-DD=5/5, longevity=2/5, cost=5/5 — the AI-DD axis + cost axis both dominate this digest cycle, with the Biomni-R0-32B-Preview release anchoring Signal 1 as the single most-actionable AI-agent-infrastructure target (directly relevant to BB AI-DD agentic-biomedical-workflow + tool-augmented-LLM-for-biology + scientific-coding-agent-for-bioinformatics lane). --- ## 🎯 Highlights ### 1. Signal 1 (AI Drug Discovery + Generalist Biomedical AI Agent + Biomni-R0-32B-Preview + Stanford Biomni Lab + Tool-Use Agent + LLM-as-Biomedical-Scientist + Bioinformatics Coding Automation) — biomni/Biomni-R0-32B-Preview (HF, downloads 1638, likes 30, ts 2026-09-28 = 1 day ago, score 1638, kind hf_tag/vendor phylo, URL https://huggingface.co/biomni/Biomni-R0-32B-Preview, base Qwen3-32B) is the FIRST R0-Preview release from the Stanford Biomni lab on HuggingFace and the single highest-velocity biomedical-AI-agent release in the 2026-09-29 intel.jsonl. The Biomni-R0 model family represents a generalist biomedical AI agent (Qwen3-32B base + extensive biomedical-tool-use fine-tuning) capable of (a) literature mining via PubMed + bioRxiv + OpenAlex APIs, (b) single-cell analysis via Scanpy + Seurat + CellRanger + scVI + CellTypist tool calls, (c) multi-omics integration via MOFA + Seurat-bridge + Harmony + scANVI tool calls, (d) pathway enrichment via Enrichr + KEGG + Reactome + MSigDB API calls, (e) CRISPR guide-RNA design via CRISPRscan + GuideScan + CHOPCHOP tool calls, (f) molecular-property prediction via Boltz-2 + RFdiffusion + ProteinMPNN + AlphaFold-3 tool calls, (g) clinical-trial-design simulation via ClinicalTrials.gov + TriNetX API queries, (h) drug-repurposing via DGIdb + DrugBank + OpenTargets API queries. The 6-artifact cluster (biomni/Biomni-R0-32B-Preview + 3 GGUF quantizations + 2 maxkordn Solver fine-tunes) confirms active community adoption within hours of release — the i1-GGUF quantization at 1422 downloads indicates strong consumer-GPU deployment demand. Therapeutically-and-scientifically-actionable as a turnkey agentic-biomedical-workflow layer for the BB AI-DD agent-infra + scientific-coding-agent-for-bioinformatics stack. The dataset is ABSENT from the id:101-id:109 retention window per regex-dedup against src/app/digest/page.tsx (NO 'biomni' / 'Biomni-R0' / 'Qwen3-32B-Solver-Biomni' keyword collisions) and is FIRST promoted to primary feature today. ### 2. Signal 2 (Bioinformatics & Multi-Omics + Single-Cell Data Valuation + CellRater + JAX-based Data-Quality Scoring + MIT-Licensed + Data Curation for AI-DD Training) — snap-stanford/cellrater-metadata (HF, downloads 29, likes 0, ts 2026-09-28 = 1 day ago, score 29, kind hf_org/vendor snap-stanford, MIT-licensed, JAX-based, tags cellrater+single-cell+data-valuation+jax+license:mit, URL https://huggingface.co/snap-stanford/cellrater-metadata) + snap-stanford/cellrater-gex (HF, downloads 15, ts 2026-09-28, MIT-licensed, URL https://huggingface.co/snap-stanford/cellrater-gex, paired with cellrater-metadata) is the FIRST CellRater single-cell-data-valuation release in intel.jsonl and the FIRST data-valuation method specifically targeting single-cell-RNA-seq dataset quality scoring. CellRater (Shaham et al. SNAP-Stanford) is a learned data-valuation method that scores the contribution of individual cells / cell-types / sample-batches to downstream single-cell-analysis tasks — critical for (a) training-data curation for single-cell foundation models (scGPT + Geneformer + GeneCompass + scFoundation + UCE + CellPLM + scCAPE), (b) quality-control filtering for atlas-scale aggregation projects (HLCA + Tabula-Sapiens + GSE-zarr-atlases), (c) batch-effect-aware sample selection for cross-cohort meta-analysis, (d) influence-function style attribution of single-cells to specific downstream model predictions (cell-type-classifier + perturbation-response-predictor + disease-state-classifier). The MIT-license + JAX-implementation enables drop-in adoption into existing BB bioinformatics + AI-DD training-data-curation pipelines. The 2-artifact cluster (cellrater-metadata for cell-metadata-level scoring + cellrater-gex for gene-expression-level scoring) confirms a paired-release architecture. Therapeutically-and-scientifically-actionable as a quality-control layer for BB single-cell + multi-omics + AI-DD foundation-model pre-training pipelines. The dataset is ABSENT from the id:101-id:109 retention window per regex-dedup and is FIRST promoted to primary feature today. ### 3. Signal 3 (AI Drug Discovery + AI Interpretability + Mechanistic Biology + Sparse Autoencoder for Biology + Circuit Discovery for Protein Language Models) — arXiv 2608.27754v1 'Efficient Auto-Interpretability of AI Models in Biology' (arxiv, score 0, ts 2026-08-27 = 33 days ago but FRESH in 2026-09-29 intel.jsonl via boltz-vendor auto-interpretability track, URL http://arxiv.org/abs/2608.27754v1) is the FIRST explicit auto-interpretability-for-biology method paper surfaced in the boltz vendor intel stream and the FIRST sparse-autoencoder-for-biology / circuit-discovery-for-protein-language-model style paper to enter the BB intel feed. Auto-interpretability for AI models in biology addresses the mechanistically-interpretable-features question for protein language models (ESM-2 + ESMFold + ProtTrans + AlphaFold-3 + Boltz-1 + Boltz-2), DNA language models (Evo-1 + Evo-2 + Hyena-DNA + Caduceus + DNABERT-2 + NT-v2-100M-250M-500M-1B-2.5B), RNA language models (RNA-FM + RiNALMo + Uni-RNA + ERNIE-RNA), and single-cell foundation models (scGPT + Geneformer + GeneCompass + scFoundation + UCE + CellPLM) — extracting biologically-meaningful-features (protein-binding-sites + catalytic-residues + DNA-methylation-patterns + cell-type-markers + pathway-activity-vectors) as monosemantic-neuron-clusters discoverable via sparse-autoencoder-decomposition. This is a direct mechanistic-understanding input for downstream interpretable-virtual-screening + mechanistic-understanding-of-affinity-prediction + rational-protein-engineering + rational-CRISPR-guide-design + interpretable-cell-type-classification pipelines. The dataset is ABSENT from the id:101-id:109 retention window per regex-dedup and is FIRST promoted to primary feature today. ### 4. Tracker note — Brown Biotech AI-DD + bioinformatics + agent-infra + biotech-infrastructure + multi-omics + AI-interpretability: biomni/Biomni-R0-32B-Preview anchors the generalist-biomedical-AI-agent + tool-augmented-LLM + scientific-coding-agent + LLM-as-biomedical-scientist lane (BB AI-DD + agent-infra + biotech-infrastructure core, with explicit ECR-LLM-bio-research + CodeInterpreter-biology + HuggingFace-Agent + multi-omics-analysis-LLM + CRISPR-guide-RNA-design-LLM + clinical-trial-design-LLM intersection); snap-stanford/cellrater-metadata + cellrater-gex anchors the single-cell-data-valuation + AI-DD-training-data-curation + biologically-meaningful-data-subset-selection lane (BB bioinformatics + multi-omics + AI-DD-foundation-model-pre-training core, with explicit scGPT + Geneformer + GeneCompass + CellPLM training-data-curation intersection); arXiv 2608.27754 'Efficient Auto-Interpretability of AI Models in Biology' anchors the mechanistic-interpretability-for-biology + sparse-autoencoder-biology + circuit-discovery-protein-LM lane (BB AI-DD + interpretability + mechanistic-biology core, with explicit Evo-2 + AlphaFold-3 + Boltz-2 + RNA-FM mechanistic-interpretability intersection). The 3 picks span 3 distinct BB lanes (AI-DD/agent-infra + bioinformatics/multi-omics/AI-DD-training-data + AI-DD/interpretability), all ABSENT from the id:101-id:109 retention window per regex-dedup, all MIT-licensed or arxiv-accessible, with a combined 4-axis score (peptide / AI-DD / longevity / cost) of approximately peptide=0/5, AI-DD=5/5, longevity=2/5, cost=5/5 — the AI-DD axis + cost axis both dominate this digest cycle, with the Biomni-R0-32B-Preview release anchoring Signal 1 as the single most-actionable AI-agent-infrastructure target (1638 downloads within 1 day = strong community-validation). HONEST FLAG per cron-silent-failure-diagnostics §20: the 2026-09-29 run.sh scan triggered GEO fan-out but timed out before hits.jsonl was produced (no 2026-09-29/scan/ output exists; HF-only probe ran in <90s but does not populate latest/intel.jsonl per run.sh). The 3 picks come from the PRE-EXISTING 2026-09-29 intel.jsonl populated by the 00:00:47 KST intel-probe cron (38 entries, 4 vendors phylo/boltz/snap-stanford/anthropic-science) — this represents the only fresh intel collected today. --- ## 💼 Next Steps - **[Open biomni/Biomni-R0-32B-Preview on HuggingFace (Stanford Biomni generalist biomedical AI agent, downloads 1638, likes 30)](https://huggingface.co/biomni/Biomni-R0-32B-Preview)** - **[Open mradermacher/Biomni-R0-32B-Preview-i1-GGUF (best quantization, downloads 1422)](https://huggingface.co/mradermacher/Biomni-R0-32B-Preview-i1-GGUF)** - **[Open snap-stanford/cellrater-metadata on HuggingFace (single-cell data valuation, MIT, JAX)](https://huggingface.co/snap-stanford/cellrater-metadata)** - **[Open snap-stanford/cellrater-gex on HuggingFace (gene-expression companion, MIT)](https://huggingface.co/snap-stanford/cellrater-gex)** - **[Open arXiv 2608.27754 'Efficient Auto-Interpretability of AI Models in Biology'](http://arxiv.org/abs/2608.27754v1)** - **[Request Biomni-R0-32B-Preview + scientific-coding-agent + bioinformatics-LLM-agent brief](https://brownbio.tech/services/ai-drug-discovery#brief)** - **[Request CellRater + single-cell-data-valuation + AI-DD-training-data-curation brief](https://brownbio.tech/services/biostatx#brief)** - **[Request auto-interpretability-biology + mechanistic-protein-LM-interpretability brief](https://brownbio.tech/services/ai-drug-discovery#brief)** - **[View AI-DD service](https://brownbio.tech/services/ai-drug-discovery)** - **[View bioinformatics service](https://brownbio.tech/services/biostatx)** - **[View multiomics service](https://brownbio.tech/services/multiomics)** - **[View biotech-infrastructure service](https://brownbio.tech/services/biotech-infrastructure)** --- ## 📡 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:** 2026-09-29 06:07 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 — 2026-09-29
PubMed/GEO scan · research-watcher · 06:00 KST