Insilico launches specialist AI models for drug discovery chemistry

Clinical-stage drug discovery company Insilico Medicine has released a series of specialist artificial intelligence models designed for chemistry and biology, reporting state-of-the-art performance across more than 70 benchmark tasks.

The models were developed through Insilico’s MMAI Gym for Science framework, which trains compact language models on groups of related scientific problems rather than relying on general-purpose capabilities.

“For years, language models have been seen as versatile conversational tools, but when it came to complex, data-scarce problems in drug discovery, specialised computational methods held the upper hand,” said Dr Alex Zhavoronkov, Founder and CEO of Insilico Medicine. “With MMAI Gym, we are proving that language-model architectures, when trained with domain-specific multi-task precision, can move far beyond high-level general knowledge. They are now competing with and outperforming dedicated scientific methods on real-world chemistry and geroscience benchmarks, marking a fundamental shift toward truly predictive AI for human health.”

Focus on difficult drug discovery tasks

Insilico said the approach is intended to address areas such as ADMET and target potency prediction, where specialised computational models have traditionally outperformed general language models.

The chemistry portfolio includes models for chemical synthesis, ADMET prediction and target activity prediction across GPCR and kinase panels. The ADMET specialist was tested across 28 tasks and delivered strong results in areas including drug-drug interaction risk, cytotoxicity and pharmacokinetic properties.

Two protein family-focused models were also developed for target activity prediction, covering 44 GPCRs and 67 kinases respectively. Insilico said both achieved state-of-the-art IC50 prediction performance on selected targets, with potential applications ranging from virtual screening and lead prioritisation to selectivity profiling.

The chemical synthesis models focus on single-step retrosynthesis and use Liquid AI’s compact 2.6B-parameter architecture. According to Insilico, they outperform leading dedicated methods on both standard benchmarks and more challenging out-of-distribution tests.

Broader biological applications

MMAI Gym has also been extended to biology, with specialist models trained on clinical, omics and molecular biology data. Insilico said selected models have matched or exceeded substantially larger general-purpose frontier models on relevant benchmarks.

The results are documented through DDD Bench, Insilico’s standardised framework for assessing AI systems across drug discovery and development tasks.

The company said the significance lies not simply in language models’ ability to process scientific information, but in their emerging ability to compete with computational methods specifically designed for prediction.

MMAI Gym was first unveiled at NeurIPS 2025. Subsequent work presented at ICLR 2026 and ICML 2026 expanded the framework, including new approaches to chemistry modelling and retrosynthesis evaluation. A further study involving 2.6B-MMAI and 24B-A2B-MMAI has been accepted to EMNLP 2026.

AI-driven pipeline expansion

The announcement comes as Insilico reports rapid commercial and research growth. The company recently reported approximately $106 million in revenue for the first half of 2026, up 287 percent year-on-year, alongside an adjusted net profit exceeding $51 million.

As of late August, Insilico had nominated nine development candidates within nine months of 2026 and achieved eight clinical milestones across proprietary and co-developed programmes.

Its lead programme, Rentosertib (ISM001-055), which Insilico describes as the world’s first drug candidate discovered and developed using generative AI, has advanced to a Phase III trial for idiopathic pulmonary fibrosis.