Research Scientist at @InSilicoMeds Opinions are my own
Montréal, Québec
Joined January 2018
- Tweets98
- Following248
- Followers234
- Likes390
Pinned Tweet
🧬LLMs are getting better at chemistry. But can they design molecular binders in 3D?
At @InSilicoMeds we test whether LLMs can generate ligands directly in a protein pocket while following spatial constraints such as fixed fragments, pharmacophores and required interactions.
🧵
LFM2 proved to be a remarkably capable foundation for our works on chemistry and drug discovery at @InSilicoMeds.
After seeing what LFM2 capable of, I’m excited to see what LFM3 makes possible. Congrats to @liquidai!
A few examples of what we built with LFM 👇
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
Our work on LFM2-based C3LM for retrosynthesis
xitter.cf/Max__Kuznetsov/status/…
🧪At #ICML2026, we presented our new paper from @InSilicoMeds:
“When a Single Answer Is Not Enough: Rethinking Single-Step Retrosynthesis Benchmarks for LLMs”
📄 Paper: arxiv.org/abs/2602.03554
🧵👇
And how we pushed its retrosynthesis capabilities further xitter.cf/Max__Kuznetsov/status/…
I’m excited to share the follow-up paper to our #ChemCensor framework for evaluating single-step retrosynthesis 🧪
Our C3LM model was already competitive with, and in some cases superior to, general-purpose foundation models, but a gap remained between LLMs and conventional retrosynthesis models.
We finally closed and surpassed that gap 🚀
The key ingredients were… 👇
@InSilicoMeds × @liquidai MMAI Gym paper is at #EMNLP2026! 🧬
Reach out to @AjMizera at the conference to chat about training state-of-the-art language models for chemistry and drug discovery.🧠
How do you make a foundation model fluent in drug discovery? 🧬💊
Excited to present our paper at #EMNLP2026:
“MMAI Gym for Science: Training Liquid Foundation Models for Drug Discovery”
400+ tasks. One goal: teach Liquid AI models the language of drug discovery.
Come chat!
More on MMAI Gym and how it works 🧵👇
xitter.cf/Max__Kuznetsov/status/…
Accepted to the #EMNLP2026 Industry Track 🎉
@InSilicoMeds built MMAI Gym for Science. @liquidai built the Liquid Foundation Model. We combined them.
400+ drug-discovery tasks. Chemistry-native tokens. SFT, then RL.
Not a bigger model. A better gym. 🏋️
#insilicoSOTAFM 🧵👇
Accepted to the #EMNLP2026 Industry Track 🎉
@InSilicoMeds built MMAI Gym for Science. @liquidai built the Liquid Foundation Model. We combined them.
400+ drug-discovery tasks. Chemistry-native tokens. SFT, then RL.
Not a bigger model. A better gym. 🏋️
#insilicoSOTAFM 🧵👇
5/6
The same model can also write 3D molecular structures as text 🧬
Nothing has to condition it: asked for a molecule and nothing else, the model writes the graph and every atom’s coordinates.
Given a protein pocket, it does the same thing on demand - the ligand’s SMILES graph first, then coordinates one atom at a time. No diffusion module.
We measure this on #Bench3DFit, our own benchmark for pocket-conditioned generation: both LFM2-MMAI variants achieve the strongest UniDock binding affinity of any language model we tested - beating several specialist diffusion baselines and approaching the strongest models overall.
The contrast is striking:
GPT-5.5 generates molecules that are almost perfectly valid and geometrically tidy - but bind poorly. Our models do the opposite.
MMAI Gym pushes the LM toward the objective that matters for structure-based drug design, rather than rewarding it for simply playing safe 🎯
The takeaway 💡
With the right data, rewards, and training recipe, a 2.6B model can handle drug-discovery tasks that frontier-scale general-purpose LLMs still miss - at a fraction of the inference cost.
The generalizable part is the gym, not the parameter count 🏋️
MMAI Gym for Science is designed to work with the causal LM you already have.
Joint work between @InSilicoMeds and @liquidai 🤝
With @zulfatmif, @RimShayakhmetov, @sumrexromanus, @ThomasMM17, @mathieu_reymond, @mihirbafna14, @kaelikl, @eugenebabin, @ramin_m_h, @xanamini, @VladAladin, @AlexAliper and @biogerontology.
📄 Paper:
arxiv.org/abs/2603.03517v2
🤗 Discuss on Hugging Face:
huggingface.co/papers/2603.0…
🏋️ MMAI Gym for Science:
insilico.com/mmai
🔑 Request model access:
liquid.ai/request-access-mma…
📊 Explore MMAI specialists on #DDDBench:
dddbench.insilico.com
LFM2-2.6B proved remarkably efficient for its size, providing a strong foundation for training a SOTA-level specialist on our tasks. Small model, serious chemistry.
It was a pleasure to work with #LFM by @liquidai @ramin_m_h to train our new #insilicoSOTAFM model for single-step retrosynthesis 🧪on ~46M reactions! It provides plenty of diverse reactions that only partially intersect with the reaction space by previous SOTA small models! 🔎👀
LFM2 is an impressive LLM - great work, @liquidai, @ramin_m_h, and @xanamini!
From my own experience, LFM2.5 is even more impressive and performs better on chemistry tasks.
🪶 What can a 2.6B-parameter language model achieve in retrosynthesis?
With a leak-proof, broad-coverage drug discovery benchmark #DDDBench, it becomes easier to evolve general-purpose models into SOTA-level specialists.
#insilicoSOTAFM
🚀 5 new drug-discovery specialist LLMs. SOTA-level performance across 70+ benchmark tasks.
Insilico Medicine has released a series of frontier Specialist Language Models for chemistry and biology, trained through MMAI Gym.
They cover drug safety, potency prediction, chemical synthesis, and biology.
🧬 From general-purpose models to scientific specialists.
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.
We’ve already shared some of the recipe behind this SOTA-level performance.
Find them in our recent preprint: arxiv.org/abs/2608.18940
Without access to the specialized tools, computational models, and web search used in @AnthropicAI protein binders design case study, Opus 5.0 also generates non-binding molecules in many cases, placing them outside the binding pocket, failing to fill it adequately, or missing the key interactions required for binding. There is still a long way to go before it can compete with specialized approaches such as diffusion models.
Promising results, but what about the most common drugs, small molecules?
#Bench3DFit from @InSilicoMeds challenges models to generate 3D molecule binders for proteins.
Opus-5 makes mistakes like clashes, small ligands and under-filling the pocket, but it’s come a long way!
I’m excited to share the follow-up paper to our #ChemCensor framework for evaluating single-step retrosynthesis 🧪
Our C3LM model was already competitive with, and in some cases superior to, general-purpose foundation models, but a gap remained between LLMs and conventional retrosynthesis models.
We finally closed and surpassed that gap 🚀
The key ingredients were… 👇
3/n
Together, these two ingredients enable C3LM to compete with and surpass strong conventional, non-LLM retrosynthesis models on the challenging URSA-expert-2026 test set 🏆
Our analysis shows that C3LM and conventional retrosynthesis models explore complementary reaction spaces 🧩
Each approach generates plausible reactions that the other may miss, suggesting that hybrid ensembles could provide broader chemical coverage and more robust synthesis planning than either model family alone 🤝
4/n
📄 Read the paper: arxiv.org/abs/2608.18940
🤗 Join the discussion on Hugging Face: huggingface.co/papers/2608.1…
📊 Recent #ChemCensor benchmark results on #DDDBench: dddbench.insilico.com
Kudos to team @sumrexromanus @chem_logos @chemcensor @mathieu_reymond @VladAladin @AlexAliper @biogerontology