Pharma.AI Webinar 2026 | Deciphering the Spring Kickoff Updates from Insilico Medicine

As artificial intelligence and foundation models evolve rapidly, the biopharmaceutical sector is quickly advancing into a new phase of AI-driven R&D. Recently, Insilico Medicine (03696.HK), a generative AI–driven, clinical-stage drug discovery company, hosted the Pharma.AI 2026 Spring Kickoff webinar.

The webinar focused on key progress for the Pharma.AI platform in Q1 of 2026, including: the official launch of MMAI Gym for Science; the rollout of the new benchmarking portal and leaderboard; and a live demonstration of how MMAI Gym transforms general-purpose foundation models into domain-specific models for drug discovery.

Meanwhile, the webinar also announced a series of functional updates to AI tools, including Generative Biologics, Chemistry42, and PandaOmics. These updates aim to build on the validated capabilities of the Pharma.AI platform by introducing more powerful AI agents and lightweight foundation models, further improving the efficiency and precision of drug discovery.

To address the bottleneck of general-purpose models underperforming on specialized drug discovery tasks, Insilico launched MMAI Gym for Science in January 2026, a benchmarking and training platform for scientific foundation models. MMAI Gym aggregates diverse benchmarking tasks covering drug discovery, biology, chemistry, and broader scientific fields.

The platform highlights three core benchmarking categories:

• ScienceAI Bench (Science Bench): Evaluates broad scientific reasoning capabilities across biology, chemistry, longevity, material science, and agricultural science.
• Drug Discovery Benchmark (Drug Discovery Bench): Evaluates end-to-end drug discovery tasks, including target identification, molecular design, and optimization in real-world R&D scenarios.
• Insilico Bench: Proprietary benchmarks developed by Insilico Medicine for drug discovery and other complex scientific challenges.

Insilico has launched public leaderboards across these categories, showcasing model performance across over 200 benchmarking tasks.

Using a two-stage approach that combines supervised fine-tuning (SFT) and reinforcement fine-tuning (RFT), Insilico has systematically strengthened several models. For example, Qwen 3-14B achieved up to a 10-fold improvement in some drug development tasks post-training. The smaller Qwen 3-4B significantly improved performance on high-complexity tasks such as clinical trial outcome prediction, even outperforming massive general models like GPT-5 in specific scenarios. Furthermore, the LFM 2.6B lightweight model from Liquid AI, trained via the MMAI Gym, surpassed larger specialized or general models in property prediction, functional reasoning, and retro-synthesis.

Generative Biologics is an AI-driven biologic engineering platform focusing on designing and optimizing antibodies, peptides, and other biologics. It leverages high-precision physics-based computational tools to predict developability metrics such as hydrophobicity and self-association, helping researchers screen candidates more efficiently.

This update introduces a new Cyclic Peptide design workflow that supports head-to-tail cyclization and disulphide-bonded architectures. Core to this is a multi-step generative engine combining All-atom Diffusion, Inverse Folding, re-scoring, and co-folding. This allows for the generation of hundreds of potential binding peptides in just a few hours.

In a GLP-1R peptide optimization case study, the team used the platform to start from lead peptide P3, identify binding modes via co-folding and Molecular Dynamics (MD), and fine-tune the Reward Model for new design cycles. The entire cycle took only 4 weeks, with all 30 candidates successfully expressed and 27 showing measurable activity. The top variant, P3-1, achieved a 6-fold increase in activity, reaching a sub-nanomolar level of 828 pM.

Chemistry42 has achieved a leap in its core engine by merging generative AI with physics-based methods. The next-generation 3D generative engine Nach01 has iterated into Nach01 MMAI through intensive training in the MMAI Gym. This model natively supports multi-objective 3D molecular optimization, accepting 3D point clouds, ligands, 3D anchors, and pharmacophores, and optimizing ADMET, PLI, and pocket occupancy via prompts. Following training on 14 million molecular pairs and 1 million ligand-protein environments, Nach01 MMAI performs 5 times better than previous models in optimization tasks.

Additionally, Generative Chemistry now supports multi-target generation, allowing users to upload additional PDB files to penalize or optimize binding against auxiliary targets. The Alchemist module has added Absolute Binding Free Energy (ABFE) calculations using a Double System Solvent Box (DSSB) setup. This allows direct estimation of △G from first principles without relying on pairwise transformations between ligands, providing a robust baseline for challenging scenarios such as scaffold changes or net charge differences.

PandaOmics further improved the precision and automation of target identification. In the field of omics research, it has transitioned from tissue-bulk to single-cell resolution, supporting over 1,400 single-cell datasets across more than 100 indications. This enables researchers to pinpoint differentially expressed genes at the cell-type level.

Simultaneously, the newly released AI agent PandaClaw transforms bioinformatic analysis through a conversational interface. Users can perform complex analyses—such as querying EGFR expression across top cancers—via natural language, with the agent automatically generating plans, executing calculations, and producing publication-ready charts and textual interpretations.

The series of updates presented at the 2026 Spring Kickoff demonstrates that Insilico is advancing AI drug discovery to the next stage, centred on AI-driven decision-making. Whether through injecting expert-level scientific reasoning via MMAI Gym or lowering the barrier to complex R&D workflows with PandaClaw, the Pharma.AI platform continues to evolve toward being smarter, more efficient, and easier to use.

Please visit pharma.AI for more information, or scan the QR code to set up a demo with our team of experts today.