Denis Sidorenko

Machine learning for drug discovery @InSilicoMeds • Views my own

[Denis Sidorenko @qoffee1337]

@VladAladin Very cool

[Denis Sidorenko @qoffee1337]

@Hyde_ai3 Amazing

[Denis Sidorenko @qoffee1337]

Small model, seven tasks, SOTA+. Nice work team.

[Insilico Medicine @InSilicoMeds]

🧬 Rentosertib went from an AI-designed molecule to patients and recently showed shifts toward younger predicted biological age across 6 proteomic aging clocks.

The first experiment is done. Now you can try our new SOTA+ drug discovery models yourself.

InsilicoMMAI-Chem-ADMET achieves SOTA+ performance on 7 drug safety and metabolism prediction tasks. 🚀

It’s a 4B language model trained on 140K+ datapoints in our MMAI Gym for Science.
#insilicoSOTAFM

[Denis Sidorenko @qoffee1337]

@stefan_neymark Thanks! 🙌

[Denis Sidorenko @qoffee1337]

Proud to have been part of this one. The method matters more than the headline here — you can measure geroprotection inside a normal disease trial, without waiting decades to notice by accident.

[Alex Zhavoronkov, PhD (aka Aleksandrs Zavoronkovs) @biogerontology]

My dear friends, I am happy to report the publication of the most important paper in my life to date (we have several great papers coming out but this is very special). Tomorrow, I will present this paper for the first time at the Nature AI Healthcare in Paris and will post a longer post on this story and its broader implications for how to conduct clinical trials. Please read it and comment on it.
Many thanks to the great co-authors of the study and everyone who contributed. Many thanks to the many reviewers (friendly and unfriendly) for spending so much time and helping make it better.
Link in the comments.

[Denis Sidorenko @qoffee1337]

@andrewaiginin Thanks Andrew! 🙏

[Denis Sidorenko @qoffee1337]

@vovalive First in Nature Biotechnology, yes. But not my first in the Nature portfolio — Precious2GPT came out in npj Aging back in 2024. You were on that one too 🙂
https://t.co/ErnvTnNjAT

[ashe @ashebytes]

since you guys loved the exploding tesla..

I used GPT-6 Astra to create a 3D website that pulls apart the male anatomy into 2,234 modeled pieces!

we are in a renaissance of learning

[Emanuel AOF @Emanuel_Andre7]

Okay, this is genuinely impressive.

I asked GPT-6 Astra to help me understand my own ankle pain. It gave me a full interactive 3D atlas — bones, ligaments, tendons, real motion axes, sliders for plantarflexion and inversion, and a live readout of what each ligament is doing.

One session.

[Greg Brockman @gdb]

Astra for visualizing how a drug works:

[Andrew Aiginin @andrewaiginin]

Yesterday, I built a cell model. 🧬 Today, it showed me how Ozempic works.

I asked it to explain the mechanism.

1 hour later... THIS. 🤯

GPT-6 Astra is insanely powerful. 🚀

I can’t stop @DeryaTR_ @gdb

Now I’m wondering... can it model the rentosertib mechanism? 👀

[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

[elvis @omarsar0]

Finally, a good paper testing if graph memory actually beats flat retrieval for long-term agents.

(bookmark this one)

Researchers extract each conversational turn into typed nodes and attributed edges, answer from a two-hop subgraph, and hold the candidate-generation budget fixed at five retrieval roots.

On LongMemEval the graph gets token F1 0.42 against 0.47 for a flat vector baseline, and a paired bootstrap over 500 questions puts the gap at -0.050 (95% CI -0.085 to -0.016).

The damage concentrates on questions that require recalling a specific prior assistant turn, where judged correctness falls from 0.911 to 0.607. Splitting a turn into entities discards the surface form those questions depend on.

The forgetting module fares much better. One pruning pass over a persistent 27,021-node graph, scored on recency, access frequency, degree centrality and age, removes 9.8% of nodes and 9.5% of stored bytes with token F1 unchanged.

Paper: https://t.co/KDUecWNGTH

Chat with Paper: https://t.co/b661ajV2ri

[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

[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.

[Insilico Medicine @InSilicoMeds]

9 in 10 drugs fail. Often, because the target is wrong.

🎯 We introduce SOTA models for target identification. TargetPro picks the right target 83% of the time.
The best frontier model manages just 34%.

That’s 2.4× better and a new SOTA across all 8 disease areas.
#insilicoSOTAFM

[Alex Zhavoronkov, PhD (aka Aleksandrs Zavoronkovs) @biogerontology]

🧪 NEW SOTA in chemical synthesis

Building molecules is hard. There are countless possible routes, but only some actually work. We trained a 2.6B LFM model by @liquidAI in #MMAIGym to find them.
📈 +7% vs. the best specialist tool
🚀 +27% vs. frontier LLMs
💥 2.6B punches way above its weight

#insilicoSOTAFM

[Andrew Aiginin @andrewaiginin]

The Diels–Alder reaction was discovered in 1927, and nearly 100 years later, chemists still use it to build complex molecules. 🧪

I remember being amazed at university when my professor showed me how chemical synthesis works.

One of the practical tasks he gave me was to demonstrate how nucleotides could have formed under the conditions described by the Oparin–Haldane hypothesis, often referred to as the primordial soup. I struggled a lot with this task 😅

Today, I’m even more excited to see AI learning to reason about this process.

Insilico trained the @liquidai LFM, a 2.6B-parameter language model, on single-step retrosynthesis, using 45.6M verified reactions and rigorous out-of-distribution evaluation.

The result: SOTA performance, outperforming dedicated chemistry tools and frontier models.
#insilicoSOTAFM

[Alex Zhavoronkov, PhD (aka Aleksandrs Zavoronkovs) @biogerontology]

And here is the moment many of you were waiting for - Insilico launches series of SOTA foundation models for drug discovery. Partners can deploy individually, orchestrate via MCP, distill them or train with them in #MMAIGym

[Maksim Kuznetsov @Max__Kuznetsov]

We’ve updated the #DDDBench with our own models, including LFM2-MMAI-Chem-SSRS-2.0, our single-step retrosynthesis specialist and the current SOTA on #ChemCensor.