Stefan N.

sharing my findings and thoughts
MLE @InSilicoMeds

[Stefan N. @stefan_neymark]

Antibody developability is the unglamorous gate every candidate has to clear.
GPT-6 Astra is now the best frontier model at calling it — which matters more than one more binder generator. #AI4DD #AntibodyEngineering

[Andrew Aiginin @andrewaiginin]

GPT-6 Astra is now the best-performing frontier model for antibody developability prediction in our benchmark. 🧬

Finding an antibody that binds is only the beginning. Will it aggregate? Will it stay stable? Can it actually behave like a drug?

GPT-6 Astra predicts these developability properties better than the other frontier models we tested. 🤯

And yes, the interactive visualization showing how an antibody actually works was also built with Astra in about 1 hour.

This model looks seriously powerful.

Congrats @gbt @sama
#insilicoSOTAFM

[Stefan N. @stefan_neymark]

🎯 Potency is not a property of a drug. It belongs to a drug and 1 target.

IC50 is how little it takes to halve that target, measured target by target.

InsilicoMMAI-Chem-Kinases, Insilico's kinase domain specialist: 30 SOTA models. 📉

#insilicoSOTAFM #AI4DD

[Stefan N. @stefan_neymark]

Most consumer DNA apps sell certainty they don't have.
This one ships uncertainty on purpose — genotype, evidence, and how much we actually know, per marker.
The honest version of exploring your own biology with AI. #Genomics #AI4DD

[Vladimir Aladinskiy @VladAladin]

Overnight, I used GPT-6 Astra (high reasoning effort) to analyze raw data from my old DNA test, performed on an Illumina chip covering ~660,000 genetic variants.

Then, in 40 minutes, I built an interactive body atlas to explore selected associations involving taste, eye pigmentation, lactose digestion, muscle proteins, and metabolism.

The app brings together 25 selected markers from my file, with recorded genotypes, supporting evidence, and explicit uncertainty. DNA imports are processed locally in the browser.

My most personal data project with Astra. 🧬

Next, I plan to get fresh tests from several labs, compare the results, and learn more about my biology using Frontier AI.

@DeryaTR_ @gdb @thekaransinghal

[Stefan N. @stefan_neymark]

Two problems a new Cell paper highlights about how we study cells:
Most Perturb-seq data comes from cancer-like immortalized lines. It misses biology unique to healthy cell types — and biology that only shows up when a cell is activated.
Worth a read 👇
https://t.co/53zSGJEuDz

[Stefan N. @stefan_neymark]

@qoffee1337 that’s huge!

[Stefan N. @stefan_neymark]

Six aging clocks. Six independent groups. All read 3–4 years younger after 12 weeks.
Rentosertib — the first AI-designed drug — just showed clinical biological-age reversal.
The convergence across independent clocks is what makes this hard to dismiss. #AI4DD #Longevity

[Insilico Medicine @InSilicoMeds]

🧬 Rentosertib just turned back biological age in people.

⏱️ 6 aging clocks. 6 independent groups.
After 12 weeks, all 6 clocks read them 3 to 4 years younger.

🚀 The first AI-designed drug. The first clinical demonstration of 3–4 year biological age reversal.

The first experiment is done. 🔬 Now run your own with Insilico’s SOTA models.
#InsilicoSOTAFM

[Stefan N. @stefan_neymark]

an exploded view for MD is such a good idea — half of understanding a simulation is seeing the pieces separately. kind of surprised this wasn't already standard.
#StructuralBiology #AI4DD

[Evgeny Kirilin @EvgenyKirilin]

71,492 atoms🤯
Pull the view apart, and the molecules keep moving!

I've always wanted an exploded view for atomistic simulations so I asked GPT-6 Astra to build one.

Watch a real GPCR in its membrane become a grid of individual molecules. Stay for the lipid close-up.

Thank you @andrewaiginin for the inspiration from your cell view, and @DeryaTR_ - this is how we should be exploring our biomolecules.

[Stefan N. @stefan_neymark]

honestly this is why the retrosynthesis benchmarks matter so much. it's not a toy task — it's the thing standing between a drawing and a real compound.
#AI4DD #Retrosynthesis

[Andrew Aiginin @andrewaiginin]

🧬 Building a molecule is like LEGO, but much, much harder.

There are millions of theoretical molecules, but only a fraction are actually buildable.

And synthesis is expensive. A single complex molecule can cost $10,000+ to make. 💰

That’s why retrosynthesis matters. It helps predict how to actually make a molecule from available starting materials. 🔬
#insilicoSOTAFM

[Stefan N. @stefan_neymark]

can't wait for the benchmarks. "ridiculously powerful" is a vibe until it meets a held-out task — that's when we actually find out.
#AI4DD

[Andrew Aiginin @andrewaiginin]

Built an insane model of the human cell in just 30 minutes 🤯

GPT-6 Astra is ridiculously powerful. Can’t wait to put it to the test on biological tasks.

@RimShayakhmetov @DeryaTR_ take a look!

[Stefan N. @stefan_neymark]

@andrewaiginin @RimShayakhmetov @DeryaTR_ crazy 😳

[Stefan N. @stefan_neymark]

SOTA single-step retrosynthesis, from 2.6B parameters.
Small enough to run local, trained sharp enough to beat the specialists.
The moat keeps turning out to be the gym, not the size. #AI4DD #Retrosynthesis

[Insilico Medicine @InSilicoMeds]

Can a language model become SOTA in chemical synthesis? 🧪🔥

We put a 2.6B-parameter @liquidai model into MMAI Gym.

It now sets the state of the art in single-step retrosynthesis. 🚀

#insilicoSOTAFM

[Stefan N. @stefan_neymark]

@VladAladin been watching general LLMs creep up on the specialist tools all year. seeing them actually stand up to the baselines feels like the moment that finally lands.

[Stefan N. @stefan_neymark]

🔀 A cell passes a message inward by flipping switches. Kinases are those switches.

In cancer one is stuck on, so the instruction to grow never stops.

InsilicoMMAI-Chem-Kinases, Insilico's kinase domain specialist: 31 SOTA models. 🧬

#insilicoSOTAFM #AI4DD

[Stefan N. @stefan_neymark]

Set the age, add a compound, read the omics back. A dial for time and a knob for intervention in the same box. If it holds up, the wet lab just got a rehearsal #AI4DD

[Insilico Medicine @InSilicoMeds]

What if you could watch a cell age in real time, and see the molecular response as it happens? ⏳

Set the age, add a compound, then read the transcriptome, proteome and methylome that come back. Each profile comes with a predicted biological age.

Check out our Virtual Aging Cell: https://t.co/T9NoKpIdOn

[Stefan N. @stefan_neymark]

Docking scores are noise until you add the one constraint that matters.
For CDK2, that's the hinge H-bonds. One MCP applied it automatically and 7x'd the hits.
Context beats compute. #AI4DD #AgenticAI

[Andrew Aiginin @andrewaiginin]

One MCP. 7× Higher Hit Rate 🚀

What does Agentic Drug Discovery look like?

PLI Mapper MCP, part of the upcoming Chemistry42 MCP Hub, allows an agent to filter molecules based on their interactions with a protein.

Same CDK2 DUD-E set. Same docking. Same top-50 cutoff.

Adding Leu83 + Glu81 hinge H-bonds as a filter took the hit rate from 4% to 28%. 📈

One MCP. One additional layer of information.

[Stefan N. @stefan_neymark]

Ornith-1.5 proposes its own tasks, builds its own scaffolds, and generates its own RL rollouts — a model creating the experiences it learns from.
Open weights, MIT, Opus-4.8-comparable. #LLMEvals #OpenWeights

[Ornith @ornith_]

Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.

It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
✅Terminal-Bench 2.1 (86.1)
✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
✅DeepSWE (56)
✅HLE (44.6)
✅ClawEval (81.4)
✅Tool Decathlon (71.2)

Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.

All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
📘Tech Blog: https://t.co/OZ63scRWLB
🤗Huggingface: https://t.co/mGJLwhrQOM

[Stefan N. @stefan_neymark]

Most models pick one scale of biology and hope it generalizes. The Virtual Aging Cell is trying to hold 6 at once — and keep them consistent when you perturb one.
That coherence is the hard part, and the whole point. #AI4DD #virtualcell

[Insilico LAIR @precious_gpt]

Virtual Aging Cell in action: emulate a therapy and control the patient's organism at 6 levels of biology
↓↓↓
Check it out at https://t.co/ok63PHxhpb

[Stefan N. @stefan_neymark]

@EvgenyKirilin @InSilicoMeds @Nature A molecule took years to learn to hit TNIK. Your legs do it every summer morning. Same destination, older technology.

[Stefan N. @stefan_neymark]

The result that matters from #Bench3DFit: models satisfy the 3D constraints you specify — but satisfying them doesn't guarantee a physically plausible pose.
Instruction-following and physical realism are two different skills. We just watched them split. #AI4DD #DrugDiscovery

[Maksim Kuznetsov @Max__Kuznetsov]

Pocket-conditioned ligand generation is like finding the right key for a lock. 🔐

The protein pocket is the lock 🔒, and the ligand is the key 🔑. The key needs the right shape and size to fit into the pocket without colliding with its walls.

But fitting the pocket is only the first part of the challenge. 📐 The teeth of the key need to line up with the right parts of the lock - just like a ligand needs to position the right functional groups at the right locations to make specific interactions with protein residues.

In practice, chemists specify these spatial requirements using different abstractions. An anchor fragment 🧩 is like a part of the molecule that you already know has to be there. A pharmacophore point 🎯 marks where a particular type of chemical feature needs to be positioned. An interaction ⚡ represents a specific contact the ligand needs to make with the pocket.

Diffusion models 🌊 have traditionally dominated 3D molecular generation, but integrating multiple heterogeneous conditions is often non-trivial. LLMs 🤖 offer a compelling alternative: they can naturally combine these diverse 3D instructions through a common language interface.

We explore this in 3D-Fit 🧬, a benchmark for 3D ligand generation under realistic spatial constraints. The results show that modern foundation models can follow local 3D constraints surprisingly well, but satisfying these constraints does not always translate into physically plausible poses or strong docking scores.

Read more in the #Bench3DFit paper: https://t.co/yTs2OFstV8
Explore recent results in #DDDBench: https://t.co/oYrb7wpEcS

[Stefan N. @stefan_neymark]

A cell you can age, perturb, and watch respond — that's a wind tunnel for longevity. Test the intervention before it touches a human. #AI4DD

[Insilico Medicine @InSilicoMeds]

Insilico Medicine introduces the Virtual Aging Cell

We think of age as a number. 🔢 Biology doesn’t. 🧬

The Virtual Aging Cell (VAC) is built to explore how cellular states change across time ⏳ and how those changes shape response to intervention.

The goal isn’t just to measure aging. It’s to predict what comes next 🔬
🧵1/5