Insilico Medicine releases AI models that beat dedicated drug-discovery software

Overview

Updated Sep 4

Insilico Medicine released AI models Sept. 2 that it says outperform the dedicated software drug companies have used for years. The language-model specialists cover chemical synthesis, drug-safety prediction, and target-binding strength.

The Hong Kong-listed company reports state-of-the-art or better scores on more than 50 benchmark tasks, including 28 drug-safety endpoints covering absorption, distribution, metabolism, excretion, and toxicity. Insilico licenses its platform to 13 of the world's top 20 pharmaceutical firms, so the models face real-world testing quickly.

The open question is whether benchmark wins survive real drug programs. Insilico's lead drug, the AI-discovered rentosertib, is in Phase III trials for a lung disease. The company nominated nine development candidates in the first nine months of 2026.

Why it matters

If language-model AI reliably predicts drug safety and potency before lab work, early discovery gets faster and cheaper — and more candidates reach human trials sooner.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

50+

Benchmark tasks with state-of-the-art or better scores

Across chemistry and biology, verified on the company's DDD Bench framework.

13 of 20

Top pharmaceutical firms licensing Insilico's platform

Licensing gives the new models a fast path to commercial use.

Development candidates nominated in nine months of 2026

A company record for annual pipeline productivity, driven by the Pharma.AI platform.

$600M

Potential total value of Takeda collaboration

Includes initiation fees, near-term payments, milestones, and tiered royalties.

Voices

Curated perspectives — historical figures and your fellow readers.

Ever wondered what historical figures would say about today's headlines?

People Involved

Alex Zhavoronkov

Founder and CEO, Insilico Medicine

Leading the company through rentosertib Phase III and the MMAI Gym rollout

Ramin Hasani

Co-founder and CEO, Liquid AI

Partner supplying the underlying compact model architecture

Organizations Involved

Insilico Medicine

Clinical-stage biotechnology company

Publicly listed on the Hong Kong exchange; releasing frontier specialist AI models

A clinical-stage generative AI drug-discovery company listed on the Hong Kong exchange (HKEX:3696).

Liquid AI

Foundation-model company

Technology partner supplying model architecture

An MIT spinoff building lightweight foundation models for enterprise applications.

Timeline

June 2014 September 2026

3 events Latest: September 2nd, 2026 · 1 week ago

September 2026
Frontier specialist AI models released
Latest Product Release

Insilico releases chemistry and biology specialists trained via MMAI Gym, claiming state-of-the-art performance on more than 50 benchmark tasks.
September 2nd, 2026
March 2026
Insilico and Liquid AI announce partnership
Partnership

The companies unveil LFM2-2.6B-MMAI, a lightweight scientific foundation model for pharmaceutical research.
March 3rd, 2026
June 2014
Insilico Medicine founded
Founding

Alex Zhavoronkov founds the company in Baltimore to apply AI to aging research and drug discovery.
June 1st, 2014

Scenarios

MMAI Gym models reshape pharma pipelines

Possible Resolves by End of 2027

Discussed by: Insilico's own benchmarking and industry analysts tracking AI-drug-discovery adoption

Partner firms fold the language-model specialists into discovery workflows. Lower-cost predictions let teams test more chemistry earlier, and a wave of MMAI-Gym-based candidates reaches the clinic.

Benchmark wins don't survive real chemistry

Possible Resolves by End of 2027

Discussed by: Computational-chemistry researchers who note language models historically lagged specialized molecular models

The models overfit benchmark datasets and degrade on novel chemical structures, just as earlier AI claims did. Adoption stalls as pharma teams stick with established computational tools.

Rentosertib approval validates generative AI discovery

Uncertain Resolves by End of 2028

Discussed by: Industry observers and Insilico, which calls it the world's first generative-AI-discovered drug

The Phase III trial for idiopathic pulmonary fibrosis succeeds, and regulators approve the drug. Approval would give Insilico's entire approach, including its new specialist models, a market-proven track record.

Historical Context

2 moments from history that rhyme with this story — and how they unfolded.

1980-2000

Computer-aided drug design, 1980s-1990s

Molecular-modeling and docking software met deep skepticism from bench chemists who trusted wet-lab instincts, and early failures slowed acceptance.

Then

One tool at a time, computational methods earned niches in lead optimization and scoring.

Now

CADD became a standard part of discovery, not a replacement for experiments, shaping what the industry expects from new methods.

Why this matters now

The adoption curve for today's AI mirrors this: breakthrough claims get tested against real programs, and survivors become standard tools rather than wholesale replacements for lab work.

November 2020

AlphaFold (2020)

DeepMind's AlphaFold burst onto protein-structure prediction with results that stunned structural biologists, matching or beating experimental methods after decades of slow progress.

Then

AlphaFold triggered a wave of adoption, with millions of structures predicted and later folded into biology toolkits.

Now

It proved a general AI method could outperform purpose-built scientific software and reset expectations for the field.

Why this matters now

Like AlphaFold, Insilico's models claim to beat tools built specifically for a scientific task. The parallel is also a caution: AlphaFold took years of independent validation before its real limits and uses were understood.