48–72h+
Physical Cell Verifier
Compiler: ~10ms; PDOs & mice take weeks
−91% / +48%
DeepMind & Virtual Biotech
Vorinostat microHO; 37k-agent cell-type lift
+46% / 92.1%
BioDiscovery & TxAgent
Non-essential CRISPR; 211+ ToolRAG tools
1,000× / 2×
Boltz-2 & IsoDDE
18s FEP-grade affinity; 2× AF3 OOD docking
3-Stage
AI → Longevity Arc
VERVE-102 (−62% LDL) · Life Bio FDA ’26
1. SYSTEM ANALOGY: WET BIOLOGY AS 3-BILLION-YEAR LEGACY CODE
5 CORE PRIMITIVES
GENOME (3.2B bp DNA)
Monolithic repo: 1.5% coding exons, 98.5% undocumented switches (GWAS).
EVO 2 / ALPHAGENOME
1M-bp static analyzer
PAM CUT
CRISPR / BASE EDITOR
Regex-targeted binary patch: 20-nt gRNA address; single typo ruins 4 wks.
CRISPR-GPT / VERVE
FSM + in-vivo A→G edit
PERTURB-SEQ / scRNA
Runtime stack trace across ~20,000 genes/cell; sparse & noisy dropouts.
ARC STATE / BIODISC.
100M+ cell state rollouts
CO-FOLDING & AFFINITY
3D API key-in-lock physics; classical FEP+ simulations cost hours/ligand.
ISODDE / BOLTZ-2
2× AF3 / 18s (1,000×)
PATIENT ORGANOID (PDO)
3D living mini-organ staging; ~90% drugs pass 2D cells but fail in-vivo.
ORIGENE / CO-SCI
−91% liver microHO fib.
2. WHY SWARMS EVOLVED HERE: 8-ORDER-OF-MAGNITUDE VERIFIER LATENCY CHASM
LOG10 TIME SCALE
1ms
1s
1m
1h
72h
3wk
1yr+
Lean / Rust AST
No. 1–2 Math (exact)
~10 ms · 100% deterministic
MuJoCo / Waymo Sim
No. 6 Robotics (rigid)
~25 ms · rigid kinematics
Boltz-2 / Evo 2 / State
In-silico priors (noisy)
18s (1,000× speedup vs FEP+)
2D CRISPR / NK Cells
AutoScreen / BioDiscovery
48–72 hrs ($2k+)
3D PDO / Aged Mouse
OriGene / ClockBase
1–6 wks
WET-LAB WALL: NO BRUTE-FORCE RL ROLLOUTS
ARCHITECTURAL CONSEQUENCE: PRE-COMMIT CRITICS & SURROGATES
Math swarms fire 10,000 Lean checks/min. Bio agents get ONE physical batch per week.
→ Winning stacks filter 10^8 candidates via Neural Simulators + Swarm Critics first.
TOOLRAG / MCP DIRECT QUERY BYPASS
(TxAgent, Biomni, AutoScreen query DBs live)
FOUNDRY TRAINING FLYWHEEL (MO–YR)
(Perturb-seq plates & trials rebuild atlases)
OUTER VERIFIER: WET LABS, XR FOUNDRIES & HUMAN TRIALS
48H – 18 MO GROUND TRUTH
XR / PLATE
LabOS XR & Pilots
Smart glasses check pipettes
Lila / Emerald / Opentrons
Robotic NK & microHO plates
VERVE-102 / Life Bio ’26
−62% LDL-C · Reprogram IND
Physical reality barrier: executes ranked edits and returns sparse, noisy wet-lab readouts to swarms.
DISPATCH TOP 10^2 → 1 ASSAYS
ACTIVE-LEARNING READOUT (48h–2wk)
CENTRAL CONTROL HUB: SWARM ORCHESTRATION, CRITICS & MCP BUS
ROUTER BETWEEN DRY & WET
DEBATE + MCP
CRISPR-GPT / AutoScreen
AI Co-Sci (Elo) / OriGene
Kosmos KG / Robin / Clock
TxAgent (211+) & BioDisc.
Virtual Biotech (37k Fleet)
Paper2Agent / Biomni
Coordinates FSM, Pairwise Elo, KG memory, and fast oracle filtering prior to wet plates.
FAST SIM LOOP (<1s): ZERO-SHOT SCORES
CO-FOLD LOOP (18s): DOCK & AFFINITY
DNA & VIRTUAL CELL SIMULATORS
PEAKS/CELL
Evo 2 (40B/9.3T nt) · OpenCRISPR
AlphaGenome & AG2 (1M-bp / 9B)
Arc State (100M+) · Precious-3
Seq→
3D CO-FOLDING & ENZYME ORACLES
3D POCKET
IsoDDE (2× AF3 & Cryptic Pockets)
Boltz-2 (18s FEP+ · 1,000×) + Flow
VENUSRAR Cas12i3 · Chai-2
PRETRAIN: 9.3T nt & 167M CELLS
PRETRAIN: PDB, BINDING & AFFINITY
DATA SUBSTRATES: PERTURB-SEQ ATLASES, GEO OMICS & FDA GRAPH
ATLAS GRID
Tahoe-100M (100M cells)
Xaira / Recursion / Parse
ClockBase GEO (40+ Clks)
Open Targets / 26+ Bio DBs
openFDA ’39+ & 56k Trials
ChEMBL / HPO / PubMed
Fast Swarms Meet Slow Cells: Unlike Lean theorem provers returning pass/fail in 10ms, living human cells and 3D organoids ignore infinite context windows and demand days of bench reality.
STANFORD / ARC
& HARVARD LABS
STACK DEPTH
CRISPR-GPT / AutoScreen
MUC1 #118→#5 · 35.5m
Virtual Biotech (37k)
+48% Ph.IV · Science
TxAgent / BioDiscovery
211+ Tools · +46% hits
Biomni / Paper2Agent
Nature ’26 MCP wraps
OPEN INFRA STANDARD
Zou, Cong, Leskovec,
Zitnik & Regev define
FSM, ToolRAG & MCP
DEEPMIND &
ISOMORPHIC LABS
MODALITIES + COFOLD
IsoDDE ($2.1B Ser. B)
2× AF3 docking rate
AlphaGenome & AG2
1M-bp / 11 tracks / 9B
AI Co-Scientist
−91% Vorinostat · AML
TxGemma (2B/9B/27B)
Open weights + Agentic
TEST-TIME ELO + ORACLE
Pairs pairwise tournament
debate with closed 1-bp
& cryptic-pocket physics
FUTUREHOUSE &
EDISON SCIENTIFIC
WORLD MODEL SCALE
Robin (Nature May ’26)
Ripasudil dry AMD RPE
Kosmos World Model
1.5k papers · 42k LoC
PaperQA2 / LitQA2
79.4% verified factuals
Incyte & Pfizer ’26
Commercial KG deploys
ANTI-ROT MEMORY
Solves 20-turn ReAct rot
via central structured KG
& code-traceable claims
MIT / RECURSION /
BROAD / SJTU
FAST AFFINITY + CLOCK
Boltz-2 (+ NVIDIA)
1,000× affinity (~18s)
ClockBase Agent
40+ clocks · 500+ hits
VENUSRAR & Chai-2
5.05× Cas12i3 · de novo
OriGene (600+ MCP)
GPR160 / ARG2 PDOs
OPEN WEIGHTS FLYWHEEL
Boltz-2 steers SynFlowNet
RL; ClockBase re-mines
public GEO aging omics
OPENAI / ANTHROPIC
& RETRO / NEWLIMIT
REPROGRAMMING CAPEX
GPT-4b micro + Retro
>50× Yamanaka markers
NewLimit ($435M Ser C)
Ambrosia hepatocyte reset
Claude for Life Sci.
Benchling · PubMed MCP
Amodei 150-Yr Thesis
100 yrs bio → 5–10 yrs
FOUNDATION + LAB
LLM labs pair enterprise
MCP pipes with LNP-mRNA
cellular reprogramming
INSILICO / LILA /
XAIRA / SHIFT BIO
AUTOMATED FACTORIES
Precious-3GPT & cmRNA
Longevity Vaccines
Lila AI Factories
Closed-loop robotic lab
Shift Bio (SB000)
Safe virtual-cell reset
LabOS / STELLA / SC
Self-evolving tools + XR
PHYSICAL FLYWHEELS
Data-first biotechs turn
robotic assay telemetry
straight into priors
CLINICAL PIONEERS
2025–2026 HUMAN DATA
HUMAN TRANSLATION
VERVE-102 (Lilly)
−88% PCSK9 · −62% LDL
Life Biosciences ’26
1st FDA Epigenetic Trial
Retro RTR242 (Ph. 1)
Autophagy small mol.
Altos Labs ($3B)
Cellular resilience
TRIAL INFLECTION
Base-edited PCSK9 & FDA
reprogramming move AI
longevity into patients
BEYOND FREE-FORM REACT:
FOUR SWARM TOPOLOGIES DEFEAT HORIZON COLLAPSE
AUTOSCREEN · AI CO-SCIENTIST · KOSMOS · TXAGENT · VIRTUAL BIOTECH
1. RIGID STATE MACHINE + ANCHORED DB CRITIC
CRISPR-GPT · AUTOSCREEN · CELLAGENT
1. DEEP LIT
Mine assays &
CRISPR context
2. RESTRUCTURE
Typed JSON
schema gates
3. 26+ DB QUERY
GWAS, pathways,
PPI priors
4. SYNTHESIS
Anchored gene IDs
(zero free halluc.)
5. CRITIC VETO
Audit provenance
& reject noise
SCHEMA / PAM / OFF-TARGET REJECTION LOOP
SIGNAL RESCUE IN NOISY SCREENS
AutoScreen (Leukemia NK Cell Killing):
Raw Wet Screen
MUC1 buried at Rank #118 (dropout noise)
Swarm Post-Screen Re-Rank
MUC1 RESCUED → RANK #5 (NK Validated)
WHY RIGID FSM + SCHEMA WINS HERE:
• CRISPR-GPT: 4-gene KO + 2-gene dCas9 ON
• CellAgent / scAgents: sc-Omni single-cell pipeline
• Primer3 & CRISPResso2 prevent 4-wk oligo loss
AUTOSCREEN BENCHMARK BARS
Prospective Hit Recovery (FDR<0.1)
77.1%
Top-100 Hit Lift (320 Screens)
+19% vs best agent
Library Design Runtime (vs ~3 days)
35.54 min (1.2× smaller lib)
AUTOSCREEN RESOURCE HUB
500+ pre-built screen/GWAS reports
anchored gene IDs across 26+ bio DBs
2. EVOLUTIONARY PAIRWISE ELO TOURNAMENT
AI CO-SCIENTIST · ORIGENE · VENUSRAR
HEAD-TO-HEAD HYPOTHESIS MATCHMAKING (ZERO 1–10 LLM DRIFT)
H1: KIRA6 (AML)
H2: Alt Kinase
H3: Vorinostat
H4: Duplicate
REFLECTION CRITIC
Safety + Mechanism Audit
PROXIMITY GRAPH
Prunes idea collapse
RANKING AGENT (TOP ELO)
Evolution fuses & mutates top-Elo
winners; Meta-Review updates rubric.
WET-LAB VALIDATED WINNERS
VORINOSTAT (LIVER FIBROSIS)
−91% TGFβ remodeling in liver microHOs
KIRA6 (ACUTE MYELOID LEUKEMIA)
Selective killing in MOLM13 & KG-1a AML lines
Spared healthy marrow in concentration assays
cf-PICI “TAIL PIRACY” (BLIND TEST)
Recapitulated ~10 yrs of Imperial wet work
in days before manuscript publication
ORIGENE (600+ MCP TOOLS)
Self-evolving templates → GPR160 (liver)
& ARG2 (colorectal) verified in patient PDOs
3. CENTRAL STRUCTURED WORLD MODEL + CODE-TRACEABLE KG
KOSMOS · ROBIN · EDISON
SEPARATING WORKERS FROM SHARED STATE
WORLD MODEL KG
Verified Claims +
Code/Paper Provenance
Lit Agents (1,500 papers)
Code Agents (42k LoC)
ROBIN (Crow/Falcon)
PaperQA2 LitQA2
FINCH CODE LOOP
RNA-seq → ABCA1
HORIZON SCALING (12H RUN = ~6 MO HUMAN)
Linear Chat (Rot >20 turns)
Kosmos KG (200 rollouts · 79.4% Acc)
Cycles: 1
20 cycles / 12 hrs
PROVABLE VALIDATIONS: BLINDNESS REPURPOSING & PHARMA DEPLOYS
• Robin (Nature May ’26): Nominated ripasudil & KL001 in dry AMD; Finch RNA-seq pinned ABCA1 clearance pump.
• Provenance Rule: Every quantitative claim MUST point to exact executed script lines or paper passage (zero cherry-picks).
• Edison Scientific (2026): Deployed Kosmos into Incyte & Pfizer oncology/immunology proprietary world models.
4. TOOLRAG ROUTING + MASSIVE MAP-REDUCE PHARMA SWARM
TXAGENT · BIOMNI · VIRTUAL BIOTECH · PAPER2AGENT
DYNAMIC TOOLRAG & MCP COMPILER
TOOL OCEAN (211–600+ Bio APIs)
openFDA ’39+ · Open Targets · HPO · Pertpy
ToolRAG Select
TxAgent (Harvard)
92.1% DrugPC acc.
Biomni (Science)
150 tools / 59 DBs
Paper2Agent (Nature Sept ’26):
Auto-compiles AlphaGenome & Scanpy into MCP
THE VIRTUAL BIOTECH (SCIENCE ’26)
CSO AGENT ORCHESTRATOR
Genetics
& CRISPR
scRNA
Specificity
ADC /
Chem Mod.
37,000+ CLINICAL-TRIALIST SWARM
Audited 55,984 clinical trials vs scRNA:
+48% Ph. IV Approval for Cell-Specific Targets
TWO-REGIME SCALING: 5-AGENT DEBATE VS 37K MAP-REDUCE
• Small Creative Team (Virtual Lab, Nature ’25): PI + Critics designed 92 nanobodies binding JN.1/KP.3 spikes.
• Massive Fleet (Virtual Biotech): Proposed B7-H3 ADC for lung cancer & ran ulcerative colitis post-mortems.
• Rule of Thumb: Past ~100 bio tools, naive prompt schemas fail → ToolRAG + Code-as-Action takes over.
Unbeaten Until Round 1: Pairwise Elo tournaments sweep every LLM rubric in the gym—then living patient organoids deliver the real knockout.
FOUR HORIZON FAILURES & ARCHITECTURAL FIXES
1. CONTEXT ROT (>20t)
Logs drown chat
→
KOSMOS EXTERNAL KG
79.4% acc @ 200 rollouts
2. MODE COLLAPSE
ECHO-CHAMBER 9/10s
→
PROXIMITY + ELO
Head-to-head ranking tree
3. FAKE GENE IDS
Speculative symbols
→
ANCHORED EMISSION
DB lock + code provenance
4. SCHEMA BLOAT
600 tools break prompts
→
TOOLRAG + STELLA
Tool Ocean grows on demand
VERIFIED SWARM METRICS (2024–2026)
TxAgent (DrugPC Open Reasoning)
92.1%
Kosmos Independent Fact Accuracy
79.4%
AutoScreen Prospective Hit Recovery
77.1%
STELLA LAB-Bench LitQA (vs ~57%)
~63%
Virtual Biotech Approval Odds Lift
+48% Ph.IV
BioDiscoveryAgent Non-Ess. Gene Hits
+46% vs BO
VENUSRAR: 46.7% (14/30) Cas12i3 wet-lab hit rate
MUC1 (#118 → #5)
AUTOSCREEN
Target: Leukemia NK immune evasion
Assay: Individual KO in primary NK cells
RESCUED FROM SCREEN NOISE & VALIDATED
VORINOSTAT & KIRA6
AI CO-SCIENTIST
Targets: Liver fibrosis (HDAC) & AML blasts
Assay: Liver microHO + MOLM13/KG-1a lines
−91% TGFβ FIBROSIS + SELECTIVE AML KILL
RIPASUDIL & KL001
ROBIN (NATURE ’26)
Target: Dry AMD blindness via RPE clearance
Assay: Human RPE RNA-seq + phagocytosis
ABCA1 UPREGULATION RESTORES RPE CLEARANCE
B7-H3 ADC + +48% RULE
VIRTUAL BIOTECH
Target: Lung cancer cell-specific surface antigen
Evidence: 55,984 trials × scRNA-seq maps
+48% PH. IV ODDS FOR CELL-SPECIFIC TARGETS
GPR160 & ARG2 IN PDOs
ORIGENE
Targets: Liver (GPR160) & colorectal (ARG2)
Assay: Patient-derived tumoroids & fragments
ANTI-TUMOR ACTIVITY VALIDATED EX VIVO
CAS12i3 & 92 NANOBODIES
VENUSRAR / V-LAB
Compact Cas12i3 editing + JN.1/KP.3 spike lock
Assay: N=30 mutants; 92 synthesized binders
5.05× EDITING CUT + VARIANT SPIKE BINDING
COMPILING DNA & CELLS:
FROM CAS12i3 & PERTURB-SEQ TO 1M-BP DARK GENOMES
EVO 2 · ISODDE · BOLTZ-2 · ALPHAGENOME · ARC STATE · PAPER2AGENT
DNA PATCH & PERTURB-SEQ DROPLET WORKFLOW
5′ → 3′ CHROMATIN + BARCODE
20-nt gRNA
PAM
×
Cas9 / Cas12i3 DSB or Non-DSB Base Edit (A•T → G•C)
Primer3 PCR + CRISPResso2 sequencing quantifies edit % vs indels
SINGLE-CELL PERTURB-SEQ
Readout: ~20k genes/cell
Sequential active-learning
cuts wet sequencing spend
2M+ tissue cells mapped (SpatialAgent & CellAgent)
CRISPR-GPT (Nat. Biomed. Eng. ’25)
4-agent deterministic FSM: Modality → Delivery →
gRNA → Primer3 → CRISPResso2 verification.
Novices knocked out 4 genes (Cas12a) + 2 dCas9 ON
VERVE-102 Clinical Echo (NEJM ’26):
Single-dose base edit of PCSK9:
−88% PCSK9 · −62% LDL-C sustained to 18 mos
BIODISCOVERYAGENT (Marson / Leskovec)
Closed-loop genetic perturbation designer selecting
128 genes/round without black-box GP surrogates.
Overall Screen Hit Lift vs Bayesian Opt:
+21–23%
Hard Non-Essential Genes Lift:
+46%
Yielded ~17 extra causal hits/dataset across primary T cells
VENUSRAR & OPENCRISPR-1
Agents & pLMs engineering CRISPR scissors themselves:
VENUSRAR (arXiv:2602.00197 · Cas12i3)
ProteinGym ρ=0.551 · 14/30 (46.7%) N=30 wet hits
4.23× & 5.05× higher DNA cutting activity
OpenCRISPR-1 (Profluent, Nature ’25)
Mined 26TB / 1M+ operons; 400+ aa from SpCas9,
lower off-targets in human cells (open-source IP)
PERTURBOAGENT & CELLAGENT
Closing single-cell & Perturb-seq experiment cycles:
PerTurboAgent (Genentech / Regev)
Sequential active-learning over pathway priors;
0.27 vs 0.23 sparse hit ratio vs active baselines
CellAgent & scAgents (ICLR / sc-Omni)
Planner + Executor + Evaluator automating
scRNA-seq, scATAC-seq & spatial tissue clustering
1. CO-FOLDING & AFFINITY ORACLES (2025–2026)
ISODDE (Isomorphic Feb ’26)
Runs N’ Poses OOD ligand generalization:
1.0× AF3
2.0× AF3 SUCCESS RATE
Unifies docking + affinity + cryptic pockets.
$2.1B Series B (May ’26) + Chai-2 Abs
BOLTZ-2 (MIT / Recursion / NV)
Structure + FEP+-grade affinity compute:
Physics FEP+: Hours / ligand
~18s on 1 GPU (1,000× faster)
Steers SynFlowNet GFlowNet molecular
agents open-source across active sites
2. EVO 2 (NATURE MAR ’26): 1M-BP DNA COMPILER ACROSS ALL LIFE
40B
StripedHyena 2 Params
~9T nt
OpenGenome2 All Domains
1 token = 1 base (1M ctx)
CO-DESIGNED OPERONS & DE NOVO PHAGES
• Generates Cas + gRNA + flanks in one pass
• De novo functional bacteriophages (ΦX174)
• Outperformed wild phages in lysis assays
• Zero-shot coding & non-coding variant scores
3. ARC STATE VIRTUAL CELL + STELLA TEST-TIME EVOLUTION
ARC STATE + TAHOE-100M
1. State Embedding (SE): 167M obs. cells
2. State Transition (ST): 100M+ perturbed
+30–50% Perturb & >2× DE Gene Acc.
50 tumor models × ~1,100 drugs; energy-dist loss
Zero-shot transfer to unseen T-cell contexts
STELLA TOOL OCEAN (arXiv:2507)
26%
~54%
~63%
HLE Bio
DBQA
LitQA
Agents write & test new tools on the fly
Illumined 98.5%: Most GWAS disease variants sit in regulatory enhancers, not protein exons; AlphaGenome + Evo 2 expose their switches before agents edit.
1,000,000 BP (1 Mb) REGULATORY CONTEXT WINDOW
QUERY SNV VARIANT: REF vs ALT (ΔAVI SCORE)
ENHANCER SNV
GENOME LOCUS
EXON 1 — EXON 2 — EXON 3 — TARGET GENE PROMOTER / EXONS
1. RNA-SEQ
Expression Δ
ALT drops exon RNA
2. SPLICING
Donor / Exon Skip
Cryptic Exon-Skip Junction Predicted
3. ATAC / DNase
Open Chromatin
Δ Enhancer Peak Disrupt
4. HISTONE ChIP
H3K27ac / H3K4me3
5. TF BINDING
CTCF / Motifs
Motif binding lost at non-coding SNV
6. 3D CONTACT
Hi-C DNA Looping
3D Chromatin Loop Links Distal Enhancer ↔ Target Gene
1 PETABYTE ATLAS (~9B SNVs)
Pre-computes AVI impact across all ~9B
single-letter human mutations = O(1) lookup
ALPHAGENOME 2 (100M+ CELLS)
Single-cell resolution sc-eQTLs unmask
cell-type variants washed out in bulk assays
PAPER2AGENT (NATURE SEPT ’26)
Wraps AlphaGenome & Scanpy repos as live
tested MCP servers for downstream agents
1. TXAGENT & TOOLUNIVERSE
92.1%
DrugPC Open Reasoning Acc.
• Harvard Zitnik Lab (Mar ’25)
• ToolRAG indexes 211+ tools
• openFDA (1939+), Open Targets,
HPO clinical phenotype graph
• Multi-step drug repurpose
BEATS GPT-4o & R1
on contraindication, renal/age
dosing & adverse event chains
2. PAPER2AGENT (NATURE ’26)
Paper +
GitHub Repo
→
Live MCP
Tested Server
• Stanford Miao & James Zou
• Turns static biology papers into
“virtual corresponding authors”
• Demonstrated wrapping Alpha-
Genome, Scanpy & TISSUE
ZERO MANUAL WRAPPERS
Auto-generates unit tests +
executable MCP tool schemas
3. BIOMNI + CLAUDE MCP
150
Bio Tools
105
Pkgs
59
Databases
• Biomni (Science Aug ’26): Action
Discovery across 25 domains
• Code-as-Action: scverse, Pertpy
• Claude Life Sci: Benchling, 10x
Genomics, PubMed & BioRender
UNIFIED WET/DRY ENV
Writes Python/R pipelines zero-shot;
BMS, Generate & BigHat MCPs
4. SPATIALAGENT & PROVER
2M+ Cells
Tissue Niches
• Wang, Leskovec, Regev (bioRxiv)
• Designs spatial gene panels &
decodes cell-cell interaction
• PROVER claim-centric auditor
catches post-hoc storytelling
EXPERT-GRADE NICHES
Brain, fetal heart, tonsil, tumor;
matches or beats human experts
5. PANTHEONOS & TXGEMMA
Pantheon-Evolve (Stanford)
Mutates batch-correction & panel code
• Uncovered Cer1–Nodal in 3D
mouse embryo & cardiac programs
• TxGemma (2B/9B/27B) open
weights score ADMET toxicity
• Pairs w/ Lila & Opentrons robots
CODE + WEIGHTS SYNERGY
Evolving bio algorithms meet
automated robotic wet execution
COMPRESSING A CENTURY OF BIOLOGY:
THREE LONGEVITY STAGES INSIDE THE WET VERIFIER
VERVE-102 · NEWLIMIT · CLOCKBASE · LIFE BIO ’26 · LABOS XR
HEALTHSPAN ESCAPE CURVE & AMODEI COMPRESSION
“Machines of Loving Grace”: 50–100 yrs bio → 5–10 yrs → ~150 yr target
100%
50%
0%
YOUTH / ORGAN SURVIVAL
Age 30
Age 75
Age 100
Age 150 Horizon
Baseline (Killers)
+ STAGE 1 & 2 PLATEAU
STAGE 3: CLOCK RESET PULSES
3-STAGE BIO MECHANISMS
1. BASE EDIT
PCSK9 A•T→G•C
2. LNP PURGE
Clear aged T-cells
3. RESET CLK
Rewind DNAm age
CRITICAL THESIS: SWARMS ATTACK ALL 3 LONGEVITY BOTTLENECKS
1. Acute killers removed via base editing & organoid-screened drugs (VERVE-102).
2. Immune / tissue decay reversed via cmRNA senescent vaccines & NewLimit LNP.
3. Biological age reset & verified across 40+ clocks + Life Bio FDA ’26 human trial.
STAGE 1 · CONQUER MAJOR KILLER DISEASES (HEART, CANCER, FIBROSIS, BLINDNESS)
IN CLINIC / PDO VALIDATED
HEART: VERVE-102 TRIAL (NEJM ’26)
Single-infusion hepatic PCSK9 gene editing
from Verve / Eli Lilly clears plaque risk:
−88% PCSK9 · −62% LDL-C (18m)
CANCER: IMMUNE & ORGANOID HITS
• AutoScreen MUC1 (#5 NK) · V-Bio (+48%)
• OriGene GPR160 & ARG2 hit colon PDOs
• AI Co-Sci KIRA6 selectivity kills AML
FIBROSIS, BLINDNESS & AUTOPHAGY
• Vorinostat slashes liver microHOs −91%
• Robin: ripasudil & KL001 treat dry AMD
• Retro RTR242 autophagy Ph. 1 human trial
STAGE 2 · ENHANCE IMMUNITY, CLEAR SENESCENT CELLS & BOOST SELF-HEALING
INFLAMMAGING & REGENERATION
INSILICO CM-RNA AGING VACCINES
Insilico single-shot circular mRNA shots
purge senescent lymphocytes & autoreactive
fibrotic drivers of systemic inflammaging
NEWLIMIT ($435M C · $3.1B VAL)
Ambrosia AI screening designs LNP-mRNA TF
payloads rejuvenating human hepatocytes,
aged kidney endothelium & human T cells
CLOCKBASE OUABAIN & T-CELL HITS
• ClockBase ouabain cuts microglial brain
inflammation, slows frailty & boosts heart
• BioDiscovery: +46% non-essential T hits
STAGE 3 · REWIND & VERIFY THE AGING CLOCK (EPIGENETIC RESET)
CLOCK AUDIT → HUMAN IND ’26
CLOCKBASE AGENT (HARVARD MED)
Re-mines decades of GEO cohorts via 40+
aging clocks to find 500+ missed reversal
hits; loss-of-func beats gain-of-func
GPT-4b MICRO & PRECIOUS-3GPT
• OpenAI+RetroSOX/KLF engineered factors:
>50× stem reprogramming & DNA repair lift
• Precious-3GPT multi-omics aging clocks
LIFE BIO FDA ’26 · SHIFT · ALTOS
• Life Bio: 1st FDA human clock-reset IND
• Shift SB000 resets age, keeps cell ID
• Altos Labs ($3B) cellular resilience R&D
LOG-FUNNEL FILTERS EXPENSIVE WET BENCH TESTS
10^8 IN-SILICO (<sec–18s)
Evo 2 · AlphaGenome 9B · Boltz-2 · IsoDDE
10^4 SWARM TRIAGE (hrs)
AI Co-Sci Elo · Kosmos · TxAgent · V-Biotech
10^2 CELL SCREEN (48–72h)
AutoScreen · BioDiscovery · VENUSRAR
10^1 PDO & MOUSE (2–6 wks)
microHO · RPE · ClockBase · OriGene PDOs
1 HUMAN EDIT / IND
VERVE-102 (−62% LDL)
Life Bio ’26 Reprogram
96-WELL
3D ORGANOID
LABOS XR
DOSE / IND
CLOSING THE BENCH GAP: LABOS XR & FOUNDRIES
• LabOS (arXiv:2510.14861): STELLA 98+ tools + XR smart
glasses (LabSuperVision ) catch pipetting typos live.
• Robotic Foundries: Lila Factories, Emerald Cloud Lab,
Opentrons execute exact JSON protocols remotely.
Turn-key error prevention stops 3-week stem-cell waste.
G1
DNA LOCK
DNA SYNTHESIS & CUSTOMER SCREENING GATE
Evo 2 & protein weights are open, yet commercial oligo
vendors enforce dual-use sequence & customer KYC audits.
GATE 1 · HARD PHYSICAL OLIGO-SYNTHESIS BOTTLE-NECK
G2
FSM GATE
STATIC OFF-TARGET & ADMET SAFETY GUARD
AlphaGenome & PAM filters block promiscuous gRNA cuts,
while TxGemma & TxAgent audit liver tox & drug risk.
GATE 2 · PROVER TRACE & TOOL-CALL AUDIT REQUIRED
G3
XR HUMAN
DRY-TO-WET AIR-GAP & LABOS XR KILL-SWITCH
Zero autonomous pathogen passage; trained technicians
wearing LabOS XR glasses hold the final pipette veto.
GATE 3 · HUMAN APPROVAL BEFORE BIOSAFETY BENCH
G4
CLINICAL
REVERSIBILITY & FDA LONGEVITY IND SHIELD
Longevity swarms favor transient LNP-mRNA or base edits
(VERVE-102); SB000 screens teratoma drift pre-FDA IND.
GATE 4 · ZERO CELL-IDENTITY OR TERATOMA DRIFT RISK
From Mercury to Clock Reset: Emperor Qin Shi Huang hunted toxic mercury elixirs in vain for eternity; AI bio-swarms now map all three steps—curing killers, renewing immunity, and rewinding DNA clocks.
1. AUTONOMY DEPTH × BIOLOGICAL SCALE MATRIX (25 BENCHMARK SYSTEMS)
EMPTY QUADRANT: ORACLE BRIDGE GAP
STATIC
1-shot
×
SWARMS UNLOCK TISSUE
Need Elo + MCP + PDO loop
BIOLOGICAL SCALE (DNA → HUMAN TRIAL)
Human
Tissue
Cell
DNA/3D
← ORACLE / COPILOT / SINGLE-LOOP
CLOSED-LOOP SWARM / CLINICAL EXECUTION →
Boltz-2 (18s)
IsoDDE (2×)
OpenCRISPR-1
Evo 2 (40B)
AlphaGenome / AG2
Arc State (100M)
VENUSRAR
CRISPR-GPT
GPT-4b micro
BioDiscovery (+46%)
AutoScreen
PerTurbo / LabOS
TxAgent (211+)
Paper2Agent MCP
Biomni / CellAgent
Kosmos KG
Robin / OriGene
AI Co-Scientist
Virtual Biotech (37k)
ClockBase / Shift
NewLimit / Insilico
2026 CLINICAL LANDING
• VERVE-102 (−62% LDL-C, 18 mo)
• Life Bio FDA Epigenetic Trial
• Retro RTR242 Phase 1
2. FOUR LAWS FOR BIO-AGENT ENGINEERS
LAW 1 · RIGID FSMs BEAT FREE-FORM REACT LOOPS
Brittle oligos (CRISPR-GPT) & noisy screens (AutoScreen) require
typed state gates + anchored DB IDs rather than raw prose chains.
FSM GATE ✓
LAW 2 · WORLD MODEL KGs BEAT 1M-TOKEN CHAT LOGS
Kosmos (79.4% acc) & ToolRAG (TxAgent 211+ tools) store verified
code/paper provenance externally, defeating 20-turn rot.
KG + MCP
LAW 3 · PAIRWISE ELO + CRITICS PREVENT MODE COLLAPSE
Head-to-head tournaments (AI Co-Scientist, VENUSRAR) & Proximity
Graphs replace noisy 1–10 rubric scores with sharp signal.
#1
ELO vs COLL.
LAW 4 · SIMULATORS TRIAGE; WET CELLS DECIDE
Boltz-2 (18s), IsoDDE & AlphaGenome cut 10^8 down — yet 72h NK
assays, PDOs & human clocks remain final non-negotiable truth.
18s
GPU → CELL
3. AI WEEKEND SERIES THREAD: HOW THE VERIFIER SHAPES AGENT ARCHITECTURE (VOL. 1 NO. 1 → NO. 7)
NO. 1–2 · MATH SWARMS
Instant ms compilers (Lean)
enable 10,000× tree search
⊢ 10ms
NO. 3 · VERIFIER CAGE
Observable unit checks invite
eval-awareness & spec gaming
EVAL SPY
NO. 4–5 · CODE HARNESS
ToolRAG, MCP & memory rules
prevent silent horizon drift
HARNESS
NO. 6 · PHYSICAL AI
Sim world models + 3-tier CBF
shields across bots & cars
3-CBF
NO. 7 · LIVING VERIFIER
Cells & clocks replace AST:
swarms compress aging & R&D
72h BIO
