As AI reshapes the pharmaceutical industry, it is opening a new chapter of opportunities in drug discovery, design, and development. To help accelerate the advent of Pharmaceutical Superintelligence, Insilico Medicine (03696.HK), a clinical-stage generative AI-driven drug discovery company, welcomed over 600 scientific professionals worldwide, part of the Pharma.AI 2026 Quarterly Webinar Series
Following the earlier deployment of MMAI Gym for Science for large language model (LLM) post-training, Insilico Medicine has expanded its generative AI capabilities through a strategic collaboration with Liquid AI, yielding a 2.6B parameter model optimized for retrosynthesis and a high-capacity 24B parameter model, both available for on-premise deployment and via the Microsoft Marketplace. To enable enterprise-grade automation, Insilico has unified its entire Pharma AI suite via the Model Context Protocol (MCP), allowing researchers to execute complex workflows directly within developer environments like Claude Code and Cursor. Concurrently, the company launched PandaClaw, an autonomous AI agent for multi-omics target discovery, alongside critical upgrades to its core suite: dynamic multibody simulation scaling in MDFlow, an expanded kinase profiling panel, advanced ionization states in generative chemistry, and enhanced MD simulations for metalloproteins.
Generative Biologics: From Trial-and-Error to Rational Engineering: New Paradigms in Parallel Simulation and Precision Sifting
This quarter’s updates to the Generative Biology platform transition biologics design from empirical trial-and-error to a predictable, rational engineering ecosystem through three key enhancements:
This paradigm shift is driven by three architectural enhancements: Batch MDFlow replaces sequential molecular dynamics workflows with a high-throughput parallel execution pipeline, enabling researchers to bulk-upload and validate hundreds of candidates simultaneously to save roughly 4 hours of setup time per 100-file campaign. Simultaneously, the Interactive Optimization Workspace shifts antibody affinity maturation into a browser-native visual environment,automatically annotating CDR regions and mapping binding interfaces so teams can select variants, rebuild 3D structures, and recompute contact metrics live before entering the wet lab. Finally, our novel Cofolding Scoring Function neutralizes ranking ambiguity in structural prediction models. Paired with metrics like ipTM and trained on nearly 700,000 examples, this graph-based scorer elevated the Top-1 hit rate by 3.2x (from 12% to 38%) on highly challenging, low-homology targets. By unifying parallel compute, real-time structural manipulation, and deep learning scoring, the platform empowers teams to surface high-affinity, structurally stable therapeutics with unprecedented velocity.
Chemistry 42:Shaping the Future of Small-Molecule Discovery via High-Precision Physics and Conversational Orchestration
The Chemistry 42 Summer update underscores an impactful evolution in AI-driven, small-molecule drug discovery. By fusing robust physics-based simulation with flexible computational chemistry toolkits, the platform neutralizes deep technical complexities to deliver immediate, actionable predictive insights to preclinical teams. This shift toward precision and usability is defined by architectural updates across its core components. In Generative Chemistry, the engine moves past single-pose structural assumptions by introducing intensive multi-conformational docking sampling, calculating mutliple distinct interaction profiles per molecule to guide reward optimization alongside a fully dynamic, SMARTS-driven ionization enumerator and multi-conformation SDF advanced export toggles. Data-wrangling friction is eliminated across structure-based Apps through native .mmcif file conversions and direct PDB cloud mirroring. Alchemistry post-processing workflows gain three new transparent free-energy (ΔG) estimation modes (Zero-based, Experimental-based, and Absolute-based). Furthermore, MDFlow expands simulation robustness for biologically critical metalloproteins by deploying a 12-6-4 Lennard-Jones interaction approach coupled with a dummy atom model to maintain highly realistic zinc coordination geometries over extended 500-snapshot trajectories, while the Profiling module adds 67 new predictive models, with calibrated confidence and uncertainty indicators now implemented for all models.
The defining announcement of this release is Chemistry 42+, an AI-driven scientific orchestrator built on the Model Context Protocol (MCP). Operating natively within the platform or via secure self-hosted servers connected to third-party tools (Cursor, VS Code, Claude Code), Chemistry 42+ transforms plain natural language requests into automated, multi-step computational chemistry workflows. Live demonstrations highlighted how researchers can seamlessly execute dataset standardization, SAR analysis, activity cliff flagging, and complex protein-ligand protonation within a single conversational exchange. Ultimately, these integrated upgrades elevate Chemistry42 into an interactive scientific teammate. By pairing frontier AIDD techniques with natural-language automation, Chemistry42+ is designed to open sophisticated computational chemistry to a far broader audience — putting advanced workflows within reach of scientists across the discovery team, not just CADD specialists, and radically compressing the path from raw queries to validated therapeutic leads.
PandaOmics: Transforming Multi-Omics Discovery via Advanced Data Integration and the PandaClaw Autonomous Agent
The expansion of PandaOmics introduces PandaClaw , a conversational AI agent equipped with nine specialized bioinformatics skills designed to replace manual data wrangling with immediate, context-rich biological insights. This quarter's release advances early-stage target discovery across four core methodologies. For advanced target characterization, the Target Evaluation engine instantly surveys the entire evidence landscape—synthesizing genetic associations, structural data, patent filings, safety signals, and competitive landscapes—to generate comprehensive Go/No-Go reference reports tailored for portfolio strategy or partnering discussions. To streamline target selection, Target Claw deploys a machine learning model trained specifically on clinically validated targets to screen thousands of genes into three highly stratified tiers ( Hot Clinical-Stage Targets, High-Confidence Targets, and Novel Targets ), delivering up to 25 prioritized opportunities complete with pathway context and proposed mechanisms of action. Furthermore, the Longevity Lobster workflow uncovers unique, dual-purpose therapeutic targets that address specific disease indications while simultaneously targeting the hallmarks of aging, offering a highly differentiated edge for healthspan-focused pipelines. Finally, the platform’s analytical depth is expanded through the Signatures Module , which integrates seven new gene set libraries—bringing the total to ten—to map transcriptomic signatures directly onto human genetic evidence, upstream transcriptional regulators, and drug-gene interactions via instant, pre-calculated switching. By pairing conversational automation with machine learning scoring and massive multi-omics datasets, the platform effectively de-risks early-stage target prioritization with unprecedented speed and precision.
MMAI Gym: Revolutionizing Unified Scientific Computing with MMAI Gym and Pharma LLMs
Driven by breakthroughs in the MMAI Gym post-training space, Insilico Medicine has introduced its InsilicoMMAI-Pharma-LLMs portfolio. At its core is a unified, LFM-based multitask model (LFM2-24B-InsilicoMMAI-Chem-MT) that improved performance across 78 distinct drug design tasks, consolidating what would otherwise require many separate specialist models into one. Trained in the MMAI Gym, it matches or exceeds state-of-the-art (SOTA) domain specialists — predicting complex ADMET safety profiles, executing 3D pocket reasoning for protein–ligand binding, and mapping single-step retrosynthetic routes — while vastly outperforming generic LLMs. The portfolio also includes a highly compact 2.6-billion-parameter retrosynthesis specialist (LFM2-2.6B-InsilicoMMAI-Chem-SSRS) that, despite its small footprint, demonstrated SOTA performance and exceeds both dedicated domain models and generalist LLMs on single-step retrosynthesis, and was presented as a featured poster at ICML 2026. Availability rolls out in stages: the 2.6B retrosynthesis specialist is live now as a managed compute service on the Microsoft Marketplace, with the LFM2-24B multitask model to be onboarded later, while both models are already available for on-premise deployment today.
To explore further, please visit Pharma.AI to schedule a personalized session with our scientific experts or email Pharma.AI@insilicomedicine.com.