Alex Zhavoronkov
@elonverse.bsky.social
My only objective function is to extend productive longevity for everyone on the planet. Here to share research papers. https://x.com/biogerontology
Agentic AI closes the experimental loop - forming hypotheses, designing experiments, running analysis, revising from results. Our Cell perspective (Guo LY, Ting DSJ, Su X, Aliper A, Zhavoronkov A, Ting DSW): upside and risks are inseparable. www.cell.com/cell/abstrac...
September 23, 2026 at 1:08 AM
Agentic AI closes the experimental loop - forming hypotheses, designing experiments, running analysis, revising from results. Our Cell perspective (Guo LY, Ting DSJ, Su X, Aliper A, Zhavoronkov A, Ting DSW): upside and risks are inseparable. www.cell.com/cell/abstrac...
Co-Scientist, Robin, Biomni: agentic AI that forms hypotheses, runs experiments, and revises from results. New Cell perspective I co-authored on what this means, and why hallucination and accountability now matter more than capability. www.cell.com/cell/abstrac...
September 23, 2026 at 12:58 AM
Co-Scientist, Robin, Biomni: agentic AI that forms hypotheses, runs experiments, and revises from results. New Cell perspective I co-authored on what this means, and why hallucination and accountability now matter more than capability. www.cell.com/cell/abstrac...
A 9B model just beat GPT-5, Gemini 3.1, Grok, and Claude Opus at aging biology. Cover of Cell this week.
LongevityBench: 17 tasks, 18 frontier LLMs. Scale didn't sort them - specialized training did.
Models + benchmark public at longevitybenchmarks.org
www.cell.com/cell/fulltex...
LongevityBench: 17 tasks, 18 frontier LLMs. Scale didn't sort them - specialized training did.
Models + benchmark public at longevitybenchmarks.org
www.cell.com/cell/fulltex...
LongevityBench & Longevity-LLMs — AI for Aging Biology
The first comprehensive, multi-task, multi-omic benchmark for LLMs in aging biology. Five compact Longevity-LLMs from Insilico Medicine beat frontier models on aging tasks.
longevitybenchmarks.org
September 17, 2026 at 10:43 PM
A 9B model just beat GPT-5, Gemini 3.1, Grok, and Claude Opus at aging biology. Cover of Cell this week.
LongevityBench: 17 tasks, 18 frontier LLMs. Scale didn't sort them - specialized training did.
Models + benchmark public at longevitybenchmarks.org
www.cell.com/cell/fulltex...
LongevityBench: 17 tasks, 18 frontier LLMs. Scale didn't sort them - specialized training did.
Models + benchmark public at longevitybenchmarks.org
www.cell.com/cell/fulltex...
Every MS/MS run detects thousands of unknowns. Existing AI requires the molecular formula first — an oracle that doesn’t exist for novel compounds.
MARLIN: formula-free de novo structure elucidation. SOTA on NPLIB1 benchmark.
arxiv.org/abs/2607.04774
MARLIN: formula-free de novo structure elucidation. SOTA on NPLIB1 benchmark.
arxiv.org/abs/2607.04774
July 13, 2026 at 1:18 PM
Every MS/MS run detects thousands of unknowns. Existing AI requires the molecular formula first — an oracle that doesn’t exist for novel compounds.
MARLIN: formula-free de novo structure elucidation. SOTA on NPLIB1 benchmark.
arxiv.org/abs/2607.04774
MARLIN: formula-free de novo structure elucidation. SOTA on NPLIB1 benchmark.
arxiv.org/abs/2607.04774
Day 71 #ScienceAIBench: Dopamine receptors D1-D4. GPT 5.5 wins 3/4 subtypes, peak Spearman 0.627 on D2. D4 hardest (max 0.228). Frontier captures dopamine SAR partly — subtype gaps remain. #DrugDiscovery #InsilicoBench
July 13, 2026 at 1:05 PM
Day 71 #ScienceAIBench: Dopamine receptors D1-D4. GPT 5.5 wins 3/4 subtypes, peak Spearman 0.627 on D2. D4 hardest (max 0.228). Frontier captures dopamine SAR partly — subtype gaps remain. #DrugDiscovery #InsilicoBench
DrugGen-2 makes molecule generation disease-aware: conditioning on disease ontology + target sequence, not only the protein (GPT-2 + GRPO RL).
On diabetic nephropathy targets its ligands out-dock enalapril (-9.9 vs -8.3). Docking is a hypothesis, but the direction is right.
arxiv.org/abs/2607.08404
On diabetic nephropathy targets its ligands out-dock enalapril (-9.9 vs -8.3). Docking is a hypothesis, but the direction is right.
arxiv.org/abs/2607.08404
July 12, 2026 at 3:31 PM
DrugGen-2 makes molecule generation disease-aware: conditioning on disease ontology + target sequence, not only the protein (GPT-2 + GRPO RL).
On diabetic nephropathy targets its ligands out-dock enalapril (-9.9 vs -8.3). Docking is a hypothesis, but the direction is right.
arxiv.org/abs/2607.08404
On diabetic nephropathy targets its ligands out-dock enalapril (-9.9 vs -8.3). Docking is a hypothesis, but the direction is right.
arxiv.org/abs/2607.08404
Day 70 #ScienceAIBench: antibody polyreactivity. GPT 5.5 leads (peak 0.152 Ovalbumin). Predicting nonspecific antibody binding stays a near-zero blind spot for every model. #DrugDiscovery #InsilicoBench
July 12, 2026 at 2:38 PM
Day 70 #ScienceAIBench: antibody polyreactivity. GPT 5.5 leads (peak 0.152 Ovalbumin). Predicting nonspecific antibody binding stays a near-zero blind spot for every model. #DrugDiscovery #InsilicoBench
Day 69 #ScienceAIBench: Lipophilicity & Solubility (ADMET). GPT 5.5 wins 4/4, peaking 0.925 on LogP. Kinetic Solubility hardest (max 0.314). LogP nearly solved; aqueous solubility still meaningfully harder. #DrugDiscovery #InsilicoBench
July 11, 2026 at 5:57 PM
Day 69 #ScienceAIBench: Lipophilicity & Solubility (ADMET). GPT 5.5 wins 4/4, peaking 0.925 on LogP. Kinetic Solubility hardest (max 0.314). LogP nearly solved; aqueous solubility still meaningfully harder. #DrugDiscovery #InsilicoBench
Day 68 of #ScienceAIBench: subtype selectivity on opioid receptors (Delta/Kappa/Mu). GPT 5.5 wins all 3, peaking 0.639 on OPRD1; Gemini 3.1 Pro 2nd. Mu (OPRM1) stays hardest - no model clears 0.330. Grok 4.3 trails, dips negative.\n\n#DrugDiscovery #InsilicoBench
July 10, 2026 at 3:48 PM
Day 68 of #ScienceAIBench: subtype selectivity on opioid receptors (Delta/Kappa/Mu). GPT 5.5 wins all 3, peaking 0.639 on OPRD1; Gemini 3.1 Pro 2nd. Mu (OPRM1) stays hardest - no model clears 0.330. Grok 4.3 trails, dips negative.\n\n#DrugDiscovery #InsilicoBench
Longevity medicine starts too late - aging trajectories may be set at or before birth. Our new Nature Health paper launches PROSPER, a life-course consortium taking gerodiagnostics from preconception to adulthood.
www.nature.com/articles/s44...
www.nature.com/articles/s44...
July 10, 2026 at 12:12 PM
Longevity medicine starts too late - aging trajectories may be set at or before birth. Our new Nature Health paper launches PROSPER, a life-course consortium taking gerodiagnostics from preconception to adulthood.
www.nature.com/articles/s44...
www.nature.com/articles/s44...
The enemy of every "AI scientist" is the context window. AstraZeneca's Ensemble QSP fixes it with 3-layer hierarchical memory: bounded context, 100% PD model selection vs 50-63% for Copilot, autonomous PK/PD. Memory > bigger model. arxiv.org/abs/2607.07666
July 9, 2026 at 1:29 PM
The enemy of every "AI scientist" is the context window. AstraZeneca's Ensemble QSP fixes it with 3-layer hierarchical memory: bounded context, 100% PD model selection vs 50-63% for Copilot, autonomous PK/PD. Memory > bigger model. arxiv.org/abs/2607.07666
New PNAS paper on why exercise protects aging muscle: exercise -> FOXO -> suppresses DEAF1 -> normalizes overactive mTORC1. DEAF1 binds the mTOR promoter directly. Block FOXO and the benefit disappears. A druggable anti-sarcopenia axis.
www.pnas.org/doi/10.1073/...
www.pnas.org/doi/10.1073/...
July 9, 2026 at 2:28 AM
New PNAS paper on why exercise protects aging muscle: exercise -> FOXO -> suppresses DEAF1 -> normalizes overactive mTORC1. DEAF1 binds the mTOR promoter directly. Block FOXO and the benefit disappears. A druggable anti-sarcopenia axis.
www.pnas.org/doi/10.1073/...
www.pnas.org/doi/10.1073/...
MARLIN: de novo molecular structure from tandem mass spec with NO ground-truth formula - the oracle almost every rival method quietly assumes but that does not exist for novel compounds. SOTA on NPLIB1.
arxiv.org/abs/2607.04774
arxiv.org/abs/2607.04774
July 8, 2026 at 1:30 PM
MARLIN: de novo molecular structure from tandem mass spec with NO ground-truth formula - the oracle almost every rival method quietly assumes but that does not exist for novel compounds. SOTA on NPLIB1.
arxiv.org/abs/2607.04774
arxiv.org/abs/2607.04774
Most AI drug-discovery agents forget everything after each task and repay the full search cost.
DrugSAGE gives them cross-task memory: verified skills + error->fix logs, reused. #1 of 9 SOTA agents on 33 tasks; big gains at near-zero test-time search.
arxiv.org/abs/2605.15461
DrugSAGE gives them cross-task memory: verified skills + error->fix logs, reused. #1 of 9 SOTA agents on 33 tasks; big gains at near-zero test-time search.
arxiv.org/abs/2605.15461
July 8, 2026 at 3:41 AM
Most AI drug-discovery agents forget everything after each task and repay the full search cost.
DrugSAGE gives them cross-task memory: verified skills + error->fix logs, reused. #1 of 9 SOTA agents on 33 tasks; big gains at near-zero test-time search.
arxiv.org/abs/2605.15461
DrugSAGE gives them cross-task memory: verified skills + error->fix logs, reused. #1 of 9 SOTA agents on 33 tasks; big gains at near-zero test-time search.
arxiv.org/abs/2605.15461
The artery doesn't age as one tissue - it ages as a committee of cells that stop agreeing.
New Aging Cell review (scRNA-seq): ECs, SMCs, fibroblasts + immune cells each reprogram toward inflammatory/senescent states.
Annotation gaps = where AI multi-omics earns its keep.
doi.org/10.1111/acel...
New Aging Cell review (scRNA-seq): ECs, SMCs, fibroblasts + immune cells each reprogram toward inflammatory/senescent states.
Annotation gaps = where AI multi-omics earns its keep.
doi.org/10.1111/acel...
July 5, 2026 at 10:25 AM
The artery doesn't age as one tissue - it ages as a committee of cells that stop agreeing.
New Aging Cell review (scRNA-seq): ECs, SMCs, fibroblasts + immune cells each reprogram toward inflammatory/senescent states.
Annotation gaps = where AI multi-omics earns its keep.
doi.org/10.1111/acel...
New Aging Cell review (scRNA-seq): ECs, SMCs, fibroblasts + immune cells each reprogram toward inflammatory/senescent states.
Annotation gaps = where AI multi-omics earns its keep.
doi.org/10.1111/acel...
BCIs are the future of humanity - and my academic publishing started there. In 2011 we published “Brain-Computer Interface Based on Generation of Visual Images” in PLOS ONE: decoding EEG as someone imagines a face vs a house, ~68% accuracy. Patent granted 2013.
doi.org/10.1371/journal.pone.0020674
doi.org/10.1371/journal.pone.0020674
July 3, 2026 at 4:31 AM
BCIs are the future of humanity - and my academic publishing started there. In 2011 we published “Brain-Computer Interface Based on Generation of Visual Images” in PLOS ONE: decoding EEG as someone imagines a face vs a house, ~68% accuracy. Patent granted 2013.
doi.org/10.1371/journal.pone.0020674
doi.org/10.1371/journal.pone.0020674
New inflammaging mechanism (Nature Aging, Hao et al.): senescent cells export R-loops to the cytoplasm via DDX1/XPO1, firing cGAS-STING and the SASP. Blocking XPO1 with FDA-approved KPT-330 cut inflammaging and extended lifespan in aged mice.
www.nature.com/articles/s43...
www.nature.com/articles/s43...
July 2, 2026 at 2:05 PM
New inflammaging mechanism (Nature Aging, Hao et al.): senescent cells export R-loops to the cytoplasm via DDX1/XPO1, firing cGAS-STING and the SASP. Blocking XPO1 with FDA-approved KPT-330 cut inflammaging and extended lifespan in aged mice.
www.nature.com/articles/s43...
www.nature.com/articles/s43...
At UBS Asian Investment Conference 2026 (Hong Kong), my message to investors on AI drug discovery: ignore the noise, look at the metrics. Insilico now has 13 IND approvals for AI drugs.
sponsored.bloomberg.com/article/ubs/...
sponsored.bloomberg.com/article/ubs/...
July 2, 2026 at 10:15 AM
At UBS Asian Investment Conference 2026 (Hong Kong), my message to investors on AI drug discovery: ignore the noise, look at the metrics. Insilico now has 13 IND approvals for AI drugs.
sponsored.bloomberg.com/article/ubs/...
sponsored.bloomberg.com/article/ubs/...
Cryopreservation still runs on two cryoprotectants found by accident - glycerol (1949) and DMSO (1959).
Our new Advanced Science Perspective: design the next generation like drugs. The bottleneck is open data, not hardware.
doi.org/10.1002/advs...
Our new Advanced Science Perspective: design the next generation like drugs. The bottleneck is open data, not hardware.
doi.org/10.1002/advs...
July 2, 2026 at 8:10 AM
Cryopreservation still runs on two cryoprotectants found by accident - glycerol (1949) and DMSO (1959).
Our new Advanced Science Perspective: design the next generation like drugs. The bottleneck is open data, not hardware.
doi.org/10.1002/advs...
Our new Advanced Science Perspective: design the next generation like drugs. The bottleneck is open data, not hardware.
doi.org/10.1002/advs...
GEN covered Insilico Medicine's new collaboration with SK Biopharmaceuticals to discover AI-designed drug candidates for neuroimmune disorders of the CNS.
Neuroimmunology is hard: safety, brain penetration, and translational biology all have to work. Now the real work begins.
Neuroimmunology is hard: safety, brain penetration, and translational biology all have to work. Now the real work begins.
June 29, 2026 at 1:05 AM
GEN covered Insilico Medicine's new collaboration with SK Biopharmaceuticals to discover AI-designed drug candidates for neuroimmune disorders of the CNS.
Neuroimmunology is hard: safety, brain penetration, and translational biology all have to work. Now the real work begins.
Neuroimmunology is hard: safety, brain penetration, and translational biology all have to work. Now the real work begins.
New review in Ageing Research Reviews: AI across the entire aging continuum - from hallmark-grounded biomarkers to Phase 2a trials.
Happy to see many of our seminal longevity-AI papers cited. I configured my LongevityClaw to follow the authors' research.
doi.org/10.1016/j.ar...
Happy to see many of our seminal longevity-AI papers cited. I configured my LongevityClaw to follow the authors' research.
doi.org/10.1016/j.ar...
Redirecting
doi.org
June 12, 2026 at 7:02 PM
New review in Ageing Research Reviews: AI across the entire aging continuum - from hallmark-grounded biomarkers to Phase 2a trials.
Happy to see many of our seminal longevity-AI papers cited. I configured my LongevityClaw to follow the authors' research.
doi.org/10.1016/j.ar...
Happy to see many of our seminal longevity-AI papers cited. I configured my LongevityClaw to follow the authors' research.
doi.org/10.1016/j.ar...
Our very important finding - NLRP3 in diabetic foot ulcers
diabetesjournals.org/diabetes/art...
diabetesjournals.org/diabetes/art...
1155-OR: ISM5059, a Novel Peripherally Restricted NLRP3 Inhibitor, Accelerates Healing of Diabetic Foot Ulcers
Introduction and Objective: Diabetic foot ulcers (DFUs) are chronic, slow-healing wounds characterized by unresolved inflammation affecting a significant p
diabetesjournals.org
June 12, 2026 at 6:20 PM
Our very important finding - NLRP3 in diabetic foot ulcers
diabetesjournals.org/diabetes/art...
diabetesjournals.org/diabetes/art...
The aging gut mapped at single-nucleus resolution in primates. NCoR1 decline drives barrier failure, inflammation, stem cell lineage shift. Metformin restores NCoR1 and delays intestinal aging in NHPs.
Nature Aging 2026
doi.org/10.1038/s43587-026-01131-0
Nature Aging 2026
doi.org/10.1038/s43587-026-01131-0
Single-nucleus interrogation of primate small intestinal aging reveals NCoR1 decline as a conserved feature that is reversed by metformin - Nature Aging
Profiling the aging nonhuman primate small intestine at single-nucleus resolution, Li et al. identify barrier dysfunction, inflammation, stem cell fate bias and a decline in NCoR1 levels. Metformin tr...
doi.org
June 11, 2026 at 11:06 PM
The aging gut mapped at single-nucleus resolution in primates. NCoR1 decline drives barrier failure, inflammation, stem cell lineage shift. Metformin restores NCoR1 and delays intestinal aging in NHPs.
Nature Aging 2026
doi.org/10.1038/s43587-026-01131-0
Nature Aging 2026
doi.org/10.1038/s43587-026-01131-0
PKMYT1: from niche kinase to 5 clinical candidates in 4 years. Our patent review maps the inhibitor + degrader explosion (2021-2025). PoC via RP-6306 in CCNE1/FBXW7-altered cancers.
doi.org/10.1080/13543776.2026.2674071
doi.org/10.1080/13543776.2026.2674071
Recent advances in the discovery of PKMYT1-targeting drugs: a patent review (2021–2025)
PKMYT1, a member of the WEE family of kinases that regulates mitotic entry by phosphorylating CDK1 at Thr14, has recently gained prominence as a promising synthetic lethal target for oncology. Sinc...
doi.org
June 11, 2026 at 10:40 PM
PKMYT1: from niche kinase to 5 clinical candidates in 4 years. Our patent review maps the inhibitor + degrader explosion (2021-2025). PoC via RP-6306 in CCNE1/FBXW7-altered cancers.
doi.org/10.1080/13543776.2026.2674071
doi.org/10.1080/13543776.2026.2674071
Useful aging target-discovery paper: glucose-lowering drug target genes + tissue-specific eQTLs + DNA methylation + aging phenotypes. The authors prioritize 9 genes, with notable CNS signals. Good map for validation. link.springer.com/article/10.1...
The tissue-specific effects of glucose-lowering drug targets on aging mediated through DNA methylation: a multi-omics genetic study - BMC Medicine
Background DNA methylation plays a key role in mediating the anti-aging effects of glucose-lowering drugs. This study aims to systematically explore the potential anti-aging effects of target genes of...
link.springer.com
June 9, 2026 at 2:13 PM
Useful aging target-discovery paper: glucose-lowering drug target genes + tissue-specific eQTLs + DNA methylation + aging phenotypes. The authors prioritize 9 genes, with notable CNS signals. Good map for validation. link.springer.com/article/10.1...