Research Pulse — 2026-09-24 | Brown Biotech

# Research Pulse — 2026-09-24
_Generated 2026-09-24 07:33 KST · AI × Bioresearch Daily synthesis_

Brown Biotech daily research pulse — multi-front synthesis across CRISPR /
gene editing, AI drug discovery, protein AI, single-cell foundation models,
FDA/regulatory, and spatial transcriptomics. Sources: company filings,
peer-reviewed journals, FDA guidance, conference disclosures.

---

### Post 1: 오늘의 전체 흐름

🧬 **AI × Bioresearch Daily — 9월 24일 (목요일) · Scaling(확장) Day**

9/24(확장) → 9/23(진전) → 9/22(집적화) → 9/21(정상화) → 9/20(실체화) → 9/19(제도화) → 9/18(산업화) → 9/17(확산). Advancing(9/23)된 기술들이 **다양한 조직·적응증·표적 집단으로의 확장(Scaling) 단계로 이전하며, 기술이 다수의 연구실·기관·적응증에서 재현되고 검증되는 날** — Tao et al.으로 AI-guided reverse transcriptase redesign — PE8 redesigned RTs 30-163 aa substitutions + up to 2.9-fold higher editing in vivo (Nature Biotechnology 2026), Sakai et al.으로 PE-PRISM 2,858 RNA motifs screen — 90%+ pathogenic ClinVar variants에서 efficiency 향상 + in vivo mouse brain/liver 적용 확인(Nature Biotechnology 2026), Meng et al.으로 ContactSeek AF3-based base editor specificity — Cas9-TadA off-target detection + high-fidelity variant 개발 (Nature 2026), DenAdel et al.으로 single-cell FM pretraining dataset scaling — 22.2M cells 400 models 6,400 experiments + plateau at fraction of current corpora (Nature Methods 2026), Hills et al.으로 FDA-EMA 10 Guiding Principles for AI in oncology drug development — 2026년 1월 joint publication operational guidance (AI in Precision Oncology 2026), Yiu et al.으로 IN-DEPTH same-slide spatial multiomics — protein-first workflow + SGCC framework + DLBCL EBV/LMP1 tumor microenvironment (Cancer Discovery 2026)이 동시에 진행된다.

Scaling은 Advancing의 다음 단계다. 오늘 6개 전선은 진전된 기술들이 **다양한 적응증·조직·표적 집단으로 확장되며 재현성과 범용성이 검증되는 날**이다.

*출처: [Tao AI-Guided PE Reverse Transcriptase Nature Biotechnology 2026](https://consensus.app/papers/details/46dfa4adac1755e3b9dbb7bbd18e294c/), [Sakai PE-PRISM RNA Motifs Nature Biotechnology 2026](https://consensus.app/papers/details/990c012ae5ab5d8587b26fa402926108/), [Meng ContactSeek AlphaFold3 Base Editor Nature 2026](https://consensus.app/papers/details/573169c799595fcca1f1656322b57bae/), [DenAdel Single-Cell FM Scaling Nature Methods 2026](https://consensus.app/papers/details/f0f6bddc16985563ad8478918950e0f4/), [Hills FDA-EMA AI Guiding Principles AI in Precision Oncology 2026](https://consensus.app/papers/details/28de9bf5ec6a56c9af30a59ee808c24c/), [Yiu IN-DEPTH Spatial Multiomics Cancer Discovery 2026](https://consensus.app/papers/details/9aa0458a622d5ff6901ed5525d1f2b39/)*

*#ScalingDay #Tao #NatureBiotechnology #PrimeEditing #AIRedesign #RTDomain #29Fold #Sakai #PEPRISM #RNAMotifs #NatureBiotech #847Variants #90Percent #Meng #ContactSeek #AlphaFold3 #BaseEditor #OffTarget #Specificity #DenAdel #NatureMethods #SingleCell #FoundationModel #Scaling #Hills #FDAEMA #GuidingPrinciples #AIPrecisionOncology #Yiu #INDEPTH #SpatialMultiomics #CancerDiscovery #DLBCL #EBV #Scaling #Reproducibility #BiomedicalAI #DailyReport*

---

### Post 2: CRISPR & Gene Editing — PE-PRISM RNA Motifs + ContactSeek AF3 Specificity의 Scaling: 90%+ Variants + 2.9-Fold

🪄 **[CRISPR/Gene Editing] Sakai et al.으로 PE-PRISM 2,858 RNA motifs screen — 847개 pathogenic ClinVar variants에서 90%+ efficiency 향상 + in vivo mouse brain/liver 적용 확인(Nature Biotechnology 2026) + Tao et al.으로 AI-guided reverse transcriptase redesign — PE8 redesigned RTs 30-163 aa substitutions + up to 2.9-fold higher in vivo editing (Nature Biotechnology 2026) + Meng et al.으로 ContactSeek AlphaFold3-based base editor specificity — Cas9-TadA off-target detection + consensus contact regions + high-fidelity variant 개발 (Nature 2026): CRISPR Scaling(확장) 동시 확인**

CRISPR prime/base editing의**다양한 pathogenic variants로의 확장 + specificity 확장(Scale)**이 출처에서 동시에 확인된다[1][2][3].

**PE-PRISM: 2,858 RNA Motifs Screen + 90%+ Pathogenic ClinVar Variants Efficiency 향상[1]**
H. Sakai et al.(Nature Biotechnology 2026)의 PE-PRISM은 prime editing 3' RNA motifs의 고속 스카이닝을 수행한다[1]. **핵심 성과 1 — 2,858 RNA motifs 평가**: 4개 iterative libraries — natural/-engineered pseudoknots, G-quadruplexes, RT recruitment elements. **핵심 성과 2 — 847개 pathogenic ClinVar variants screen**: disease-associated mutations. **핵심 성과 3 — 90%+ efficiency 향상**: top-performing motifs이 847개 pathogenic ClinVar variants 중 90% 이상에서 기존 tevopreQ1 motif 대비 향상. **핵심 성과 4 — In vivo 적용 확인**: primary human cells + mouse brain/liver 모두에서 efficiency 향상 확인. **핵심 성과 5 — tevo2.0, eHAV, eSBRMV1-A**: engineered pseudoknot variants. **핵심**: PE RNA motifs optimization의 확장 검증 — CRISPR Scaling Consolidating.

**AI-Guided Reverse Transcriptase Redesign: PE8 + 2.9-Fold In Vivo[2]**
Y. Tao et al.(Nature Biotechnology 2026)의 연구는 AI-guided inverse-folding network(ProteinMPNN)로 prime editor reverse transcriptase를 재설계한다[2]. **핵심 성과 1 — 30-163 aa substitutions**: 설계된 RT domains. **핵심 성과 2 — Enhanced folding stability + soluble expression**: 단백질 안정성 및 발현 향상. **핵심 성과 3 — Up to 2.9-fold higher intracellular protein levels**: mRNA delivery 후 세포 내 PE 단백질 수준. **핵심 성과 4 — 2.9-fold higher in vivo editing**: mouse model에서 PE6/PE7/PEmax 대비. **핵심 성과 5 — Ex vivo human primary cells**: 여러 인간 일차 세포 유형에서 efficiency 향상. **핵심**: AI-guided PE RT redesign의 확장 검증 — CRISPR Scaling Consolidating.

**ContactSeek: AF3-Based Base Editor Specificity + High-Fidelity Variant[3]**
H. Meng et al.(Nature 2026)의 ContactSeek는 AlphaFold3 기반 contact probability로 base editor specificity를 향상시킨다[3]. **핵심 성과 1 — AF3 contact probability**: Cas9-TadA base editor의 off-target DNA sequences 분석. **핵심 성과 2 — Consensus contact regions同定**: neighboring Cas residues with consistent contact changes. **핵심 성과 3 — Specificity-determining residues同定**: contact probability 기반. **핵심 성과 4 — High-fidelity variant**: Cas9 + TadA8e mutations 결합 — 기존 high-fidelity ABE보다 superior. **핵심 성과 5 — Modular 적용 가능**: Cas12a cytosine base editors로 확장 가능 확인. **핵심**: AF3-driven base editor specificity 향상 — CRISPR Scaling Consolidating.

**9월 23일 Advancing vs 오늘의 차별점**: 9/23이 **Lushington first-in-human 기능 회복 + Yu 80% rare disease mutation 표적 가능**(Advancing)을, **Ji BBB penetration 5가지 전략**(Advancing)을 제시했다면, 오늘은 **Sakai Nature Biotech**으로 **847개 pathogenic ClinVar variants에서 90%+ efficiency 향상 구체적 수치**(Scaling — 이전 PE motif 최적화의 확장 검증)를(new angle), **Tao Nature Biotech**으로 **AI-guided RT redesign 2.9-fold in vivo 수치**(Scaling — AI-guided PE 최적화의 확장 검증)를(new angle), **Meng Nature**으로 **AF3-based specificity framework의 확장 적용**(Scaling — base editor specificity framework의 확장)을 동시에 제시한다. Advancing(9/23) → CRISPR Scaling(9/24).

실무 함의: 유전체 편집 플랫폼 팀에서 Sakai et al.의 PE-PRISM 결과를 참조하여 다양한 pathogenic mutations에 대한 PE optimization pipeline을 설계해야 한다.

*출처: [Sakai PE-PRISM RNA Motifs Nature Biotechnology 2026](https://consensus.app/papers/details/990c012ae5ab5d8587b26fa402926108/), [Tao AI-Guided PE Reverse Transcriptase Nature Biotechnology 2026](https://consensus.app/papers/details/46dfa4adac1755e3b9dbb7bbd18e294c/), [Meng ContactSeek AlphaFold3 Base Editor Nature 2026](https://consensus.app/papers/details/573169c799595fcca1f1656322b57bae/)*

*#CRISPR #PrimeEditing #BaseEditing #Sakai #NatureBiotechnology #PEPRISM #RNAMotifs #Pseudoknot #847Variants #90Percent #Tao #AIRedesign #ProteinMPNN #ReverseTranscriptase #29Fold #InVivo #Meng #ContactSeek #AlphaFold3 #BaseEditor #OffTarget #Specificity #Cas9 #TadA #HighFidelity #ScalingDay #PathogenicVariants #Reproducibility #BiomedicalAI #DailyReport*

---

### Post 3: AI 신약 — FDA-EMA 10 Guiding Principles Operational Guidance + Spatial Proteogenomics Scaling

💊 **[AI 신약] Hills et al.으로 FDA-EMA 10 Guiding Principles for AI in oncology drug development — 2026년 1월 joint publication operational guidance + Good AI Practice (AI in Precision Oncology 2026) + Wang Yida et al.으로 spatial proteogenomics comprehensive review — CITE-seq RNA-protein co-localization + MALDI metabolomics + DVP AI-guided microdissection + KRONOS/HEIST multimodal AI (Advanced Science 2026) + Fay Lin으로 Pharma AI 2026 Deals actual execution — Lilly·Novo Nordisk·AstraZeneca(GEN Biotechnology 2026): AI Drug Discovery Scaling(확장) 동시 확인**

AI-Driven Drug Discovery의**규제 운영 체계 확장과 공간적 통합 확장(Scaling)**이 출처에서 동시에 확인된다[4][5][6].

**FDA-EMA Joint Guiding Principles: 10 Guiding Principles + Operational Guidance[4]**
Bethany Hills et al.(AI in Precision Oncology 2026)의 기사는 2026년 1월 FDA와 EMA가 공동 발표한 "Guiding Principles of Good AI Practice in Drug Development"를 운영 지침으로 번역한다[4]. **핵심 성과 1 — 10 Guiding Principles**: AI 시스템의 책임 있는 설계·검증·감독을 위한 shared transatlantic framework. **핵심 성과 2 — Oncology-specific operational guidance**: 종양학 연구자·임상의·기관 리더를 위한 구체적 실행 지침. **핵심 성과 3 — Good AI Practice**: AI 시스템 전生命周期에 걸친 올바른 실행 기준. **핵심 성과 4 — January 2026 joint publication**: FDA-EMA 공동 발표. **핵심**: FDA-EMA AI 규제 운영 체계의 확장 — AI drug discovery Scaling Consolidating.

**Spatial Proteogenomics: CITE-Seq + MALDI + DVP + KRONOS/HEIST multimodal AI[5]**
Yida Wang et al.(Advanced Science 2026)의 review는 spatial proteogenomics의 전체 그림을 체계적으로 분석한다[5]. **핵심 성과 1 — RNA-protein co-localization**: Spatial CITE-seq로 immune microenvironment patterns + neoantigen distribution. **핵심 성과 2 — Spatial Proteomics + Metabolomics**: MALDI(Matrix-Assisted Laser Desorption/Ionization) imaging으로 대사 재프로그래밍 캡처. **핵심 성과 3 — Deep Visual Proteomics(DVP)**: AI-guided microdissection으로 무偏見 spatial analysis. **핵심 성과 4 — NanoPOTS + PiMS**: trace samples processing + proteoform imaging mass spectrometry. **핵심 성과 5 — Multimodal foundation AI models**: KRONOS + HEIST — multiple data modalities 통합, 진단 정밀도 및 치료 예측 향상. **핵심 성과 6 — Dark proteome exploration**: low abundance/stable proteins 탐색. **핵심**: spatial proteogenomics의 확장 — AI integration Scaling Consolidating.

**Pharma AI 2026 Deals: Lilly·Novo Nordisk·AstraZeneca Actual Execution[6]**
Fay Lin(GEN Biotechnology 2026)의 기사는 2026년 빅파마 AI 거래 실제 이행을 분석한다[6]. **핵심 성과 1 — Lilly AI Deal**: 구체적 파트너십 이행. **핵심 성과 2 — Novo Nordisk AI Deal**: 구체적 파트너십 이행. **핵심 성과 3 — AstraZeneca AI Deal**: 구체적 파트너십 이행. **핵심 성과 4 — Industry validation**: AI drug discovery의 산업적 실현 가능성 확인. **핵심**: 2026 Pharma AI Deal 실제 이행 — AI drug discovery Scaling Consolidating.

**9월 23일 Advancing vs 오늘의 차별점**: 9/23이 **Fay 2026 Deals actual execution + Seyhan 0건 완전 AI 마케팅 승인**(Advancing)을 제시했다면, 오늘은 **Hills AI in Precision Oncology**으로 **FDA-EMA 10 Guiding Principles의 구체적 operational guidance**(Scaling — 규제 체계의 구체적 실행 지침으로 확장)를(new angle), **Wang Yida Advanced Science**으로 **spatial proteogenomics의 CITE-seq/MALDI/DVP/KRONOS/HEIST 전체 방법론 확장**(Scaling — 공간적 통합의 확장)을(new angle), **Fay Deals**으로 **Lilly/Novo/Astra 구체적 Deal 이행 확대**(Scaling)를 동시에 제시한다. Advancing(9/23) → AI drug discovery Scaling(9/24).

실무 함의: AI drug discovery 팀에서 Hills et al.의 FDA-EMA operational guidance를 참조하여 AI validation protocol을 수립해야 한다.

*출처: [Hills FDA-EMA AI Guiding Principles AI in Precision Oncology 2026](https://consensus.app/papers/details/28de9bf5ec6a56c9af30a59ee808c24c/), [Wang Yida Spatial Proteogenomics Advanced Science 2026](https://consensus.app/papers/details/b4757de1ad385d27875a882dd8e8071e/), [Fay Pharma AI 2026 Deals GEN Biotechnology 2026](https://consensus.app/papers/details/bc6e9ea3ec2e5025af707131d2fa0ca7/)*

*#AI #DrugDiscovery #Hills #AIPrecisionOncology #FDAEMA #GuidingPrinciples #GoodAIPractice #WangYida #AdvancedScience #SpatialProteogenomics #CITEseq #MALDI #DVP #DeepVisual #NanoPOTS #PiMS #KRONOS #HEIST #MultimodalAI #Fay #GENBiotechnology #PharmaAI #2026Deals #Lilly #NovoNordisk #AstraZeneca #ScalingDay #OperationalGuidance #RegulatoryFramework #BiomedicalAI #DailyReport*

---

### Post 4: 단백질 AI — AlphaFold2/3/ESMFold 88%/88%/76% Benchmark + OpenFold Systematic Evaluation의 Scaling

🔬 **[단백질 AI] Mahtha et al.으로 ProModEv portal AlphaFold2 88%/AlphaFold3 88%/ESMFold 76% monomer benchmark — 1,666 monomer + 994 dimer challenging targets 체계적 비교(NAR Genomics Bioinformatics 2026) + Díaz et al.으로 OpenFold/AlphaFold2/AlphaFold3/ColabFold systematic evaluation — Bland-Altman agreement analysis + per-residue divergence + UMAP visualization(Briefings in Bioinformatics 2026) + García et al.으로 de novo protein design zero-shot discrimination — ESMFold pLDDT best single predictor + topology-dependent + experimental success rate limited(Protein Science 2026): 단백질 AI Scaling(확장) 동시 확인**

단백질 AI의**벤치마크 체계 확장 + De novo design 적용 범위(Scale)**가 출처에서 동시에 확인된다[7][8][9].

**AlphaFold2/3/ESMFold Benchmark ProModEv: 88%/88%/76% vs 77%/77%/41% + 1,666/994 Challenging Targets[7]**
Sanjeet Kumar Mahtha et al.(NAR Genomics and Bioinformatics 2026)의 ProModEv portal 연구는 주요 단백질 구조 예측 모델의 체계적 벤치마크를 제시한다[7]. **핵심 성과 1 — 1,666 monomer + 994 dimer challenging targets**: PDB 2022-2024 homology-redundant 제거. **핵심 성과 2 — Monomer benchmark**: AlphaFold2 88%, AlphaFold3 88%, ESMFold 76% correct. **핵심 성과 3 — Dimer benchmark**: AlphaFold2 77%, AlphaFold3 77%, ESMFold 41%. **핵심 성과 4 — X-ray/cryo-EM only**: NMR 제외 시 AF 95%, ESMFold 83%. **핵심 성과 5 — ProModEv portal**: interactive web portal 공개(http://pdbi.nii.ac.in/ProModEv/). **핵심**: AlphaFold2/3/ESMFold 수치 기반 벤치마크 체계 확장 — 단백질 AI Scaling Consolidating.

**OpenFold/AlphaFold2/AlphaFold3/ColabFold Systematic Evaluation: Bland-Altman + Per-Residue Divergence[8]**
Florencia R. Díaz et al.(Briefings in Bioinformatics 2026)의 연구는 여러 AF 변형의 체계적 평가를 수행한다[8]. **핵심 성과 1 — 10 proteins + 222 single amino acid variants**: 체계적 평가. **핵심 성과 2 — Global metrics insufficient**: mean pLDDT, pTM-score, RMSD가 종종 유사해 보이나 실제 상이. **핵심 성과 3 — Bland-Altman agreement analysis**: per-residue Cα-confidence 차이 평가. **핵심 성과 4 — UMAP visualization**: conformational shifts 시각화. **핵심 성과 5 — Flexible/disordered regions**: largest conformational shifts — predictor choice + single-residue substitution 모두에서. **핵심 성과 6 — Reduced homology database**: predicted structural quality에 minimal impact — computationally efficient alternative. **핵심**: AF variant 체계적 평가의 확장 — 단백질 AI Scaling Consolidating.

**De Novo Protein Design Zero-Shot Discrimination: ESMFold pLDDT Best + Topology-Dependent[9]**
Mario García et al.(Protein Science 2026)의 연구는 de novo designed proteins에 대한 구조 예측 모델의 zero-shot discrimination 능력을 평가한다[9]. **핵심 성과 1 — 614 de novo designed monomers**: 11개 design studies 2012-2021. **핵심 성과 2 — Moderate discrimination**: 성공적/실패적 designs 모두 moderate discrimination. **핵심 성과 3 — ESMFold pLDDT best single predictor**: 가장 나은 개별 성능. **핵심 성과 4 — Topology-dependent**: 성공률 topology에 의존. **핵심 성과 5 — Logistic regression combination**: 모든 metrics 결합해도 modest improvement만. **핵심 성과 6 — Task-specific training 필요**: experimental validation 전 filtering 전략으로만 유용. **핵심**: De novo protein design에서의 모델 적용 범위 한계 확장 확인 — 단백질 AI Scaling Consolidating.

**9월 23일 Advancing vs 오늘의 차별점**: 9/23이 **Mahtha ProModEv 88%/88%/76% + Maddipatla AF3 experiment-guided + Gu DFG-motif 75% success rate**(Advancing)을 제시했다면, 오늘은 **Díaz Briefings in Bioinformatics**으로 **OpenFold/ColabFold 체계적 평가 + Bland-Altman + UMAP**(Scaling — 모델 평가 방법론 확장)를(new angle), **García Protein Science**으로 **de novo protein design에서 ESMFold pLDDT의 topology-dependent discrimination 한계**(Scaling — 모델 적용 범위의 구체적 한계 확장)를(new angle) 동시에 제시한다. Advancing(9/23) → 단백질 AI Scaling(9/24).

실무 함의: 단백질 AI팀에서 Mahtha et al.의 ProModEv benchmark 수치를 참조하여 de novo protein design quality control protocol을 수립해야 한다.

*출처: [Mahtha AlphaFold Benchmark ProModEv NAR Genomics Bioinformatics 2026](https://consensus.app/papers/details/11a60df3938051cc87f7e295ef5e73f2/), [Díaz OpenFold Systematic Evaluation Briefings in Bioinformatics 2026](https://consensus.app/papers/details/20ce9a43389d5e01991087d591f9097b/), [García De Novo Protein Design Zero-Shot Protein Science 2026](https://consensus.app/papers/details/f9c8c442c8065c0688de2538546747db/)*

*#ProteinAI #AlphaFold #ProModEv #Mahtha #NARGenoBioinformatics #Benchmark #88Percent #76Percent #Díaz #BriefingsInBioinformatics #OpenFold #ColabFold #BlandAltman #PerResidue #UMAP #DisorderedRegions #García #ProteinScience #DeNovo #ProteinDesign #ZeroShot #ESMFold #pLDDT #TopologyDependent #ScalingDay #BenchmarkScaling #BiomedicalAI #DailyReport*

---

### Post 5: Single-Cell Multi-Omics Foundation Model — CAPTAIN 382 + Zero-Shot + Data Scaling Plateau의 Scaling

🧫 **[Single-Cell Multi-Omics] Ji et al.으로 CAPTAIN multimodal FM — scRNA + 382 proteins + 4M cells + zero-shot + COVID-19 immune dynamics(Nature Communications 2026) + DenAdel et al.으로 pretraining dataset scaling — 22.2M cells 400 models 6,400 experiments + plateau at fraction of current corpora + no clear scaling laws (Nature Methods 2026) + Gaballa et al.으로 zero-shot FM benchmarking — scGPT/SCimilarity/UCE/Transcriptformer embeddings don't consistently outperform baselines + baseline stronger in cell type annotation + protein prediction FM better (bioRxiv 2026): Single-Cell Scaling(확장) 동시 확인**

Single-Cell Multi-Omics의**data scaling 한계 확인 + zero-shot 적용 범위(Scale)**가 출처에서 동시에 확인된다[10][11][12].

**SCMT Bench: 22.2M Cells 400 Models 6,400 Experiments + Plateau at Fraction of Current Corpora[11]**
Alan DenAdel et al.(Nature Methods 2026)의 연구는 단일세포 foundation model의 pretraining dataset size/diversity 효과를 대규모로 분석한다[11]. **핵심 성과 1 — 22.2M cells corpus**: 대규모 pretraining corpus. **핵심 성과 2 — 400 models trained**: 체계적 scaling experiment. **핵심 성과 3 — 6,400 experiments 평가**: 광범위한 평가. **핵심 성과 4 — Performance plateau**: current methods tend to plateau at only a fraction of the size of current training corpora. **핵심 성과 5 — No clear scaling laws**: LLM과 달리 명확한 scaling law 없음. **핵심 성과 6 — Dataset size+diversity+model capacity balance**: indiscriminately increasing all three보다 균형 중요. **핵심**: Single-cell FM data scaling 한계 확인 — Single-Cell Scaling Consolidating.

**Zero-Shot FM Benchmarking: scGPT/SCimilarity/UCE/Transcriptformer Don't Consistently Outperform Baselines[12]**
Yasmine Gaballa et al.(bioRxiv 2026)의 연구는 zero-shot 설정에서 foundation model embeddings를 전통적 방법과 비교한다[12]. **핵심 성과 1 — Cell type annotation**: baseline methods strongest in most datasets — FM embeddings 미달. **핵심 성과 2 — Protein expression prediction**: FM embeddings generally more accurate than baseline — SCimilarity lowest error, Transcriptformer highest correlation. **핵심 성과 3 — Data integration**: moderate results for all FMs — scVI baseline strongest. **핵심 성과 4 — Task-dependent effectiveness**: FM의 효과가 적용 및 평가 설정에 따라 상이. **핵심 성과 5 — Universal replacement 아님**: task-specific methods의 대체재 아님. **핵심**: Zero-shot FM benchmarking의 한계 확인 — Single-Cell Scaling Consolidating.

**SCMBench: CAPTAIN 382 Proteins + Zero-Shot Generalization[10]**
Boya Ji et al.(Nature Communications 2026)의 CAPTAIN은 단일세포 multimodal foundation model이다[10]. **핵심 성과 1 — 4M+ cells**: 대규모 training. **핵심 성과 2 — 382 proteins**: 동시 측정 표면 단백질. **핵심 성과 3 — Unified multimodal representations**: cross-modality dependencies 모델링. **핵심 성과 4 — Zero-shot generalization**:未见过的 task에서도 generalization. **핵심 성과 5 — COVID-19 immune dynamics**: severity 관련 immune patterns 발견. **핵심**: CAPTAIN multimodal FM의 zero-shot 확장 확인 — Single-Cell Scaling Consolidating.

**9월 23일 Advancing vs 오늘의 차별점**: 9/23이 **Ji CAPTAIN 382 proteins + zero-shot breakthrough**(Advancing)을 제시했다면, 오늘은 **DenAdel Nature Methods**으로 **22.2M cells 400 models scaling에서 plateau 발견 + scaling law 부재**(Scaling — FM pretraining의 확장 한계 실증)를(new angle), **Gaballa bioRxiv**으로 **zero-shot에서 FM embeddings가 baseline consistently outperform 실패 확인**(Scaling — zero-shot 적용 범위의 구체적 한계)를(new angle) 동시에 제시한다. Advancing(9/23) → Single-Cell Scaling(9/24).

실무 함의: 단일세포 분석팀에서 Gaballa et al.의 FM vs baseline 결과를 참조하여 task-specific method selection criteria를 수립해야 한다.

*출처: [Ji CAPTAIN Nature Communications 2026](https://consensus.app/papers/details/40361e55d4815c1dae3d5f6a96d4468a/), [DenAdel Single-Cell FM Scaling Nature Methods 2026](https://consensus.app/papers/details/f0f6bddc16985563ad8478918950e0f4/), [Gaballa Zero-Shot FM Benchmark bioRxiv 2026](https://consensus.app/papers/details/8ff15b82c5df52ca9d62383904277dd9/)*

*#SingleCell #FoundationModel #CAPTAIN #NatureCommunications #Ji #382Proteins #ZeroShot #COVID19 #DenAdel #NatureMethods #Pretraining #Scaling #22MillionCells #400Models #6400Experiments #Plateau #NoScalingLaws #Gaballa #ZeroShotBenchmarking #scGPT #SCimilarity #UCE #Transcriptformer #Baseline #CellTypeAnnotation #ScalingDay #FoundationModel #Limitations #BiomedicalAI #DailyReport*

---

### Post 6: FDA/규제 — FDA-EMA 10 Guiding Principles Operational + Oncology AI Prospective Testing 5% Evidence Gap의 Scaling

🏛️ **[FDA/규제] Hills et al.으로 FDA-EMA 10 Guiding Principles for AI in oncology drug development — 2026년 1월 joint publication + operational guidance for researchers/clinicians/institutional leaders (AI in Precision Oncology 2026) + Litt et al.으로 FDA oncology AI 149개 authorization — prospective testing 5% + clinician-in-the-loop 21% + evidence calibration 필요 (J Cancer Policy 2026) + Kunanayagam et al.으로 FDA 44 CGTs + 14 pediatric clinical pharmacology 체계화 (Pediatric Research 2026): FDA/규제 Scaling(확장) 동시 확인**

FDA/규제 영역의**FDA-EMA 공동 운영 체계 확장과 evidence gap 실증(Scale)**이 출처에서 동시에 확인된다[4][13][14].

**FDA-EMA Joint Guiding Principles: 10 Principles + Operational Guidance for Oncology[4]**
Bethany Hills et al.(AI in Precision Oncology 2026)의 기사는 FDA-EMA 2026년 1월 공동 발표한 10개 Guiding Principles를 종양학 맥락에서 운영 지침으로 번역한다[4]. **핵심 성과 1 — 10 Guiding Principles**: Good AI Practice in Drug Development. **핵심 성과 2 — Joint publication January 2026**: FDA + EMA 공동. **핵심 성과 3 — Operational guidance**: oncology researchers, clinicians, institutional leaders를 위한 구체적 실행 지침. **핵심 성과 4 — Transatlantic framework**: 미국-유럽 공동 규제 체계. **핵심 성과 5 — AI lifecycle coverage**: 설계·검증·감독 전生命周期. **핵심**: FDA-EMA 공동 규제 체계의 운영 체계로 확장 — FDA Scaling Consolidating.

**FDA Oncology AI 149 Authorization: Prospective Testing 5% + Evidence Calibration 필요[13]**
H. Litt et al.(J Cancer Policy 2026)의 분석은 FDA oncology AI의 evidence 품질 격차를 실증한다[13]. **핵심 성과 1 — 149 oncology-specific devices**: 전체 1,008개 중 15%. **핵심 성과 2 — Radiology 46% + Radiation oncology 38%**: imaging/radiation 집중. **핵심 성과 3 — Clinical testing 76%**: patient-derived data. **핵심 성과 4 — Clinician-in-the-loop 21%**: 31/149. **핵심 성과 5 — Prospective testing 5%**: 7/149만. **핵심 성과 6 — Evidence calibration 필요**: device function과 clinical risk에 따른 차등 evidence 기준 권고. **핵심**: FDA oncology AI evidence gap의 확장 실증 — FDA Scaling Consolidating.

**FDA 44 CGTs + 14 Pediatric Clinical Pharmacology[14]**
S. Kunanayagam et al.(Pediatric Research 2026)의 review는 FDA CGT의 소아 적용을 분석한다[14]. **핵심 성과 1 — 44 CGTs 승인(2025-05 기준)**: cord blood 제외. **핵심 성과 2 — 14 pediatric 사용**: 44개 중 14개. **핵심 성과 3 — PK considerations**: vector shedding, biodistribution. **핵심 성과 4 — Cellular kinetics**: expansion, persistence. **핵심 성과 5 — Preclinical models 의존**: dose selection, safety, efficacy. **핵심**: FDA CGT 소아 적용의 확장 — FDA Scaling Consolidating.

**9월 23일 Advancing vs 오늘의 차별점**: 9/23이 **Litt prospective testing 5% evidence gap**(Advancing)을 제시했다면, 오늘은 **Hills AI in Precision Oncology**으로 **FDA-EMA 10 Guiding Principles의 구체적 operational guidance**(Scaling — 규제 체계의 구체적 실행 체계로 확장)를(new angle), **Kunanayagam Pediatric Research**으로 **44 CGTs + 14 pediatric의 구체적 승인 수치**(Scaling — CGT 소아 적용의 구체적 수치 확장)를 동시에 제시한다. Advancing(9/23) → FDA/규제 Scaling(9/24).

실무 함의: 규제 기획팀에서 Hills et al.의 FDA-EMA operational guidance를 참조하여 oncology AI regulatory strategy를 수립해야 한다.

*출처: [Hills FDA-EMA AI Guiding Principles AI in Precision Oncology 2026](https://consensus.app/papers/details/28de9bf5ec6a56c9af30a59ee808c24c/), [Litt FDA Oncology AI J Cancer Policy 2026](https://consensus.app/papers/details/2f1d83bb38045de991b9cdb33bb7d624/), [Kunanayagam FDA CGT Pediatrics Pediatric Research 2026](https://consensus.app/papers/details/1c13548ad5be537a8451eed2c85ecd75/)*

*#FDA #Regulation #Hills #AIPrecisionOncology #FDAEMA #GuidingPrinciples #GoodAIPractice #OperationalGuidance #Transatlantic #Litt #JCancerPolicy #OncologyAI #ProspectiveTesting #5Percent #ClinicianInTheLoop #21Percent #EvidenceGap #Calibration #Kunanayagam #PediatricResearch #CGT #44Products #14Pediatric #ScalingDay #RegulatoryFramework #OperationalGuidance #BiomedicalAI #DailyReport*

---

### Post 7: Spatial Transcriptomics/Cancer Atlas — IN-DEPTH Same-Slide + Spatial Proteogenomics + Spatial Omics TME의 Scaling

🗺️ **[Spatial Transcriptomics/Cancer Atlas] Yiu et al.으로 IN-DEPTH same-slide spatial multiomics — protein-first workflow + SGCC framework + DLBCL EBV/LMP1 tumor microenvironment reorganization (Cancer Discovery 2026) + Wang Yida et al.으로 spatial proteogenomics comprehensive review — CITE-seq/MALDI/DVP/KRONOS/HEIST AI multimodal (Advanced Science 2026) + Nguyen et al.으로 spatial omics TME review — AI integration promise + cross-platform standardization gap (Clinical & Translational Immunology 2026): Spatial Transcriptomics Scaling(확장) 동시 확인**

Spatial Transcriptomics/Cancer Atlas의**same-slide multiomics 확장 + AI 통합 확장(Scaling)**이 출처에서 동시에 확인된다[5][15][16].

**IN-DEPTH: Same-Slide Spatial Multiomics + SGCC + DLBCL EBV/LMP1 Remodeling[15]**
S. Yiu et al.(Cancer Discovery 2026)의 IN-DEPTH는 동일 슬라이드에서 spatial proteomics로 transcriptomics를Guide하는 새로운 방법을 제시한다[15]. **핵심 성과 1 — Protein-first workflow**: protein epitopes → RNA quality/tissue integrity 보존. **핵심 성과 2 — Commercially compatible**: 기존 상용 플랫폼과 호환. **핵심 성과 3 — No RNA signal loss**: 기존 방법의 한계 극복. **핵심 성과 4 — SGCC(Spectral Graph Cross-Correlation)**: proteomic-transcriptomic framework — spatially coordinated functional state changes resolving. **핵심 성과 5 — DLBCL 적용**: EBV-positive vs EBV-negative tumor comparison. **핵심 성과 6 — C1Q macrophage enrichment**: immunosuppressive macrophage patterns. **핵심 성과 7 — CD4 T-cell dysfunction**: EBV/LMP1-driven. **핵심 성과 8 — IL27-STAT3 axis**: candidate signaling axis discovery. **핵심**: Same-slide spatial multiomics의 확장 — Spatial Transcriptomics Scaling Consolidating.

**Spatial Proteogenomics: CITE-Seq + MALDI + DVP + KRONOS/HEIST AI[5]**
Yida Wang et al.(Advanced Science 2026)의 review는 spatial proteogenomics의 전체 그림을 제시한다[5]. **핵심 성과 1 — RNA-protein co-localization**: Spatial CITE-seq. **핵심 성과 2 — MALDI metabolomics**: 대사 재프로그래밍 캡처. **핵심 성과 3 — DVP AI-guided microdissection**: 무偏見 spatial analysis. **핵심 성과 4 — NanoPOTS + PiMS**: trace samples processing. **핵심 성과 5 — KRONOS + HEIST**: multimodal foundation AI models — 진단 정밀도 및 치료 예측 향상. **핵심 성과 6 — Dark proteome**: low abundance/stable proteins 탐색. **핵심**: Spatial proteogenomics 방법론 확장 — Spatial Transcriptomics Scaling Consolidating.

**Spatial Omics TME Review: AI Integration Promise + Standardization Gap[16]**
Hao Nguyen et al.(Clinical & Translational Immunology 2026)의 review는 spatial omics의 종양 미세환경 적용을 종합한다[16]. **핵심 성과 1 — Cross-platform integration gap**: 주요 격차. **핵심 성과 2 — Data standardization gap**: 주요 격차. **핵심 성과 3 — Computational scalability**: 주요 격차. **핵심 성과 4 — AI integration promise**: biological/translational applications potential. **핵심 성과 5 — Standardized workflows 필요**: AI practical application 전제 조건. **핵심**: Spatial omics TME의 도전과 AI 통합 확장 — Spatial Transcriptomics Scaling Consolidating.

**9월 23일 Advancing vs 오늘의 차별점**: 9/23이 **Li Niche_4/Niche_11 biomarkers + Zhang SCLC 600K cells + Gao gastric cancer LAR**(Advancing)을 제시했다면, 오늘은 **Yiu Cancer Discovery**으로 **IN-DEPTH same-slide spatial multiomics의 구체적 workflow + SGCC framework**(Scaling — 같은 슬라이드 내 multimodal 통합의 확장)를(new angle), **Wang Yida Advanced Science**으로 **spatial proteogenomics의 KRONOS/HEIST multimodal AI model 확장**(Scaling — AI 통합 spatial proteogenomics의 확장)을(new angle), **Nguyen Clinical Immunology**으로 **spatial omics의 standardization gap + AI integration promise**(Scaling — 공간적 통합의 도전과 가능성 확장)을 동시에 제시한다. Advancing(9/23) → Spatial Transcriptomics Scaling(9/24).

실무 함의: 면역종양학팀에서 Yiu et al.의 IN-DEPTH workflow를 참조하여 same-slide spatial multiomics analysis protocol을 설계해야 한다.

*출처: [Yiu IN-DEPTH Same-Slide Spatial Multiomics Cancer Discovery 2026](https://consensus.app/papers/details/9aa0458a622d5ff6901ed5525d1f2b39/), [Wang Yida Spatial Proteogenomics Advanced Science 2026](https://consensus.app/papers/details/b4757de1ad385d27875a882dd8e8071e/), [Nguyen Spatial Omics TME Clinical & Translational Immunology 2026](https://consensus.app/papers/details/1b906e5abaaf5e789aaabde458127265/)*

*#SpatialTranscriptomics #CancerAtlas #Yiu #CancerDiscovery #INDEPTH #SameSlide #SpatialMultiomics #SGCC #ProteinFirst #DLBCL #EBV #LMP1 #C1QMacrophage #CD4TCell #IL27STAT3 #WangYida #AdvancedScience #SpatialProteogenomics #CITEseq #MALDI #DVP #NanoPOTS #PiMS #KRONOS #HEIST #MultimodalAI #Nguyen #ClinicalTranslationalImmunology #SpatialOmics #TME #CrossPlatform #Standardization #ScalingDay #SameSlide #Multiomics #BiomedicalAI #DailyReport*

---

### Post 8: 오늘의 요약 — Scaling Day 핵심 성과

🧬 **오늘 Scaling Day 요약 — 6개 전선 확장 검증**

**① CRISPR — PE-PRISM 847개 Variants 90%+ Efficiency + AI-Guided RT Redesign 2.9-Fold[1][2][3]**
Sakai et al.(Nature Biotech 2026)의 **PE-PRISM으로 2,858 RNA motifs screen → 847개 pathogenic ClinVar variants 중 90%+에서 efficiency 향상** 확인. Tao et al.(Nature Biotech 2026)의 **AI-guided reverse transcriptase redesign → 2.9-fold higher in vivo editing**. Meng et al.(Nature 2026)의 **ContactSeek AlphaFold3-based base editor specificity → high-fidelity variant 개발**. **핵심 수치**: 90%+ variants, 2.9-fold, 30-163 aa substitutions.

**② AI 신약 — FDA-EMA 10 Guiding Principles Operational + Spatial Proteogenomics AI[4][5][6]**
Hills et al.(AI in Precision Oncology 2026)의 **FDA-EMA 10 Guiding Principles January 2026 joint publication + operational guidance for oncology**. Wang Yida et al.(Advanced Science 2026)의 **spatial proteogenomics CITE-seq/MALDI/DVP/KRONOS/HEIST multimodal AI 확장**. Fay Lin(GEN Biotechnology 2026)의 **Lilly/Novo Nordisk/AstraZeneca 2026 Deal 실제 이행**. **핵심 수치**: 10 principles, January 2026.

**③ 단백질 AI — ProModEv 88%/88%/76% + De Novo Design Topology-Dependent 한계[7][8][9]**
Mahtha et al.(NAR Genomics 2026)의 **ProModEv portal AlphaFold2/3 88%/88% + ESMFold 76% monomer 체계적 비교**. Díaz et al.(Briefings in Bioinformatics 2026)의 **OpenFold/ColabFold Bland-Altman per-residue evaluation**. García et al.(Protein Science 2026)의 **de novo protein design zero-shot discrimination — topology-dependent + ESMFold pLDDT best single**. **핵심 수치**: 88%/88%/76%, topology-dependent.

**④ 단일세포 — 22.2M Cells Plateau + Zero-Shot FM Baseline 미달[10][11][12]**
DenAdel et al.(Nature Methods 2026)의 **22.2M cells 400 models 6,400 experiments → plateau at fraction of current corpora + no scaling laws**. Gaballa et al.(bioRxiv 2026)의 **zero-shot에서 FM embeddings가 baseline consistently outperform 실패 — cell type annotation에서 baseline 우위**. Ji et al.(Nature Communications 2026)의 **CAPTAIN 382 proteins + zero-shot generalization**. **핵심 수치**: plateau, no scaling laws, baseline stronger in cell type.

**⑤ FDA/규제 — FDA-EMA Operational Guidance + Oncology AI Prospective 5%[4][13][14]**
Hills et al.(AI in Precision Oncology 2026)의 **FDA-EMA 10 Guiding Principles operational guidance**. Litt et al.(J Cancer Policy 2026)의 **FDA oncology AI 149개 prospective testing 5% evidence gap**. Kunanayagam et al.(Pediatric Research 2026)의 **44 CGTs + 14 pediatric 구체적 수치**. **핵심 수치**: 10 principles, 5% prospective, 44 products, 14 pediatric.

**⑥ Spatial — IN-DEPTH Same-Slide + Spatial Proteogenomics AI Multimodal[5][15][16]**
Yiu et al.(Cancer Discovery 2026)의 **IN-DEPTH same-slide spatial multiomics + SGCC framework + DLBCL EBV/LMP1 C1Q macrophage/CD4 T-cell/IL27-STAT3 발견**. Wang Yida et al.(Advanced Science 2026)의 **spatial proteogenomics KRONOS/HEIST multimodal AI 확장**. Nguyen et al.(Clinical & Translational Immunology 2026)의 **spatial omics TME cross-platform standardization gap + AI integration promise**. **핵심 수치**: protein-first workflow, SGCC, C1Q, IL27-STAT3.

**Advancing(9/23) → Scaling(9/24) 핵심 전환**: 임상 성과 측정(Advancing) → **다양한 pathogenic variants/적응증/조직으로의 확장 검증(Scale)**. PE-PRISM 847개 pathogenic variants에서 90%+ efficiency, FDA-EMA 10 Guiding Principles operational guidance, 22.2M cells pretraining plateau, IN-DEPTH same-slide spatial multiomics 동시 확인.

*#ScalingDay #Summary #PEPRISM #847Variants #90Percent #Tao #AIRedesign #29Fold #ContactSeek #AlphaFold3 #Hills #FDAEMA #GuidingPrinciples #Operational #WangYida #SpatialProteogenomics #KRONOS #HEIST #Mahtha #ProModEv #88Percent #Díaz #OpenFold #García #DeNovo #TopologyDependent #DenAdel #Plateau #NoScalingLaws #Gaballa #ZeroShot #Baseline #Ji #CAPTAIN #382Proteins #Litt #OncologyAI #5Percent #Kunanayagam #CGT #14Pediatric #Yiu #INDEPTH #SameSlide #SGCC #Nguyen #Standardization #Scaling #Reproducibility #BiomedicalAI #DailyReport*

---

## Sources (1차 출처, 10개 이상 URL)

[1] [Sakai et al. — Directed evolution of small RNA-stabilizing motifs that improve prime-editing efficiency, Nature Biotechnology 2026](https://consensus.app/papers/details/990c012ae5ab5d8587b26fa402926108/)

[2] [Tao et al. — AI-guided redesign of laboratory-evolved reverse transcriptases enhances prime editing, Nature Biotechnology 2026](https://consensus.app/papers/details/46dfa4adac1755e3b9dbb7bbd18e294c/)

[3] [Meng et al. — Precise DNA base editing using AlphaFold3-based contact modelling, Nature 2026](https://consensus.app/papers/details/573169c799595fcca1f1656322b57bae/)

[4] [Hills et al. — From Principles to Practice: Operationalizing the FDA and EMA Guiding Principles for AI in Oncology Drug Development, AI in Precision Oncology 2026](https://consensus.app/papers/details/28de9bf5ec6a56c9af30a59ee808c24c/)

[5] [Wang Yida et al. — Integrating Spatial Proteogenomics in Cancer Research, Advanced Science 2026](https://consensus.app/papers/details/b4757de1ad385d27875a882dd8e8071e/)

[6] [Fay Lin — Pharma Bets Big on AI Platforms with Flurry of 2026 Deals, GEN Biotechnology 2026](https://consensus.app/papers/details/bc6e9ea3ec2e5025af707131d2fa0ca7/)

[7] [Mahtha et al. — Comparative evaluation of the prediction accuracy of AlphaFold and ESMFold for monomeric and dimeric proteins, NAR Genomics and Bioinformatics 2026](https://consensus.app/papers/details/11a60df3938051cc87f7e295ef5e73f2/)

[8] [Díaz et al. — Comprehensive evaluation of AlphaFold/OpenFold prediction of experimentally unresolved proteins through novel metrics, Briefings in Bioinformatics 2026](https://consensus.app/papers/details/20ce9a43389d5e01991087d591f9097b/)

[9] [García et al. — Evaluating zero-shot prediction of monomeric protein design success by AlphaFold, ESMFold, and ProteinMPNN, Protein Science 2026](https://consensus.app/papers/details/f9c8c442c8065c0688de2538546747db/)

[10] [Ji et al. — CAPTAIN: a multimodal foundation model pretrained on co-assayed single-cell RNA and protein, Nature Communications 2026](https://consensus.app/papers/details/40361e55d4815c1dae3d5f6a96d4468a/)

[11] [DenAdel et al. — Evaluating the role of pre-training dataset size and diversity on single-cell foundation model performance, Nature Methods 2026](https://consensus.app/papers/details/f0f6bddc16985563ad8478918950e0f4/)

[12] [Gaballa et al. — Benchmarking single-cell foundation models in a zero-shot setting, bioRxiv 2026](https://consensus.app/papers/details/8ff15b82c5df52ca9d62383904277dd9/)

[13] [Litt et al. — FDA-authorized oncology artificial intelligence and machine learning devices and their clinical evidence: A cross-sectional analysis, Journal of Cancer Policy 2026](https://consensus.app/papers/details/2f1d83bb38045de991b9cdb33bb7d624/)

[14] [Kunanayagam et al. — Clinical pharmacology insights from recent cell and gene therapy approvals relevant to pediatrics, Pediatric Research 2026](https://consensus.app/papers/details/1c13548ad5be537a8451eed2c85ecd75/)

[15] [Yiu et al. — Same-Slide Spatial Multi-Omics Integration With IN-DEPTH Reveals Tumor Virus-Linked Spatial Reorganization of the Tumor Microenvironment, Cancer Discovery 2026](https://consensus.app/papers/details/9aa0458a622d5ff6901ed5525d1f2b39/)

[16] [Nguyen et al. — Spatial omics for profiling the dynamic tumor microenvironment, Clinical & Translational Immunology 2026](https://consensus.app/papers/details/1b906e5abaaf5e789aaabde458127265/)