Evogene Ltd. has announced beginners in the class Generated AI foundation model For small molecule designs, mark breakthroughs on how new compounds are discovered. The model, released in collaboration with Google Cloud on June 10, 2025, expands Evogene’s Chempass AI platform to tackle long-standing challenges in both pharmaceutical and agriculture: finding new molecules that meet multiple complex criteria simultaneously. This development is poised to accelerate R&D in drug discovery and crop protection by enabling simultaneous optimization of properties such as efficacy, toxicity, and stability in a single design cycle.
From sequential screening to simultaneous design
In traditional drug and agrochemical studies, scientists usually test one factor at a time. First, check if the compound works and then test its safety and stability. This stepwise method is slow, expensive, and often ends with obstacles, with many promising compounds lacking at later stages. It also focuses on familiar chemical structures, limiting innovation and making it difficult for researchers to create new, highly patentable products. This outdated approach contributes to high costs, long timelines, and low success rates. 90% of drug candidates fail before reaching the market.
Generation AI changes this paradigm. Instead of filtering one by one, AI models can juggle multiple requirements at once to design molecules that are strong, safe and stable from the start. The new foundational model of Evogene was explicitly built to enable this simultaneous multiparameter design. This approach aims to risk risk during subsequent stages of development due to front load considerations such as ADME and toxicity to initial design.
In reality, it could mean less late failure. For example, fewer drug candidates show lab results that show failure only in clinical trials due to side effects. In short, generator AI allows researchers to optimize many aspects of successful molecules faster, smarter and simultaneously, rather than addressing each one.
Internal Chempass AI: How generative models design molecules
At the heart of Evogene’s Chempass AI platform is a powerful new foundational model trained with a huge chemical dataset. To teach AI the “language” of molecules, the company has assembled a curated database of approximately 40 billion molecular structures across known drug-like compounds and diverse chemical scaffolds. Using Google Cloud’s Vertex AI infrastructure with GPU supercomputing, this model learns patterns from this vast chemical library, giving you an unprecedented breadth of knowledge about what drug-like molecules look like. This large-scale training regimen is similar to training large-scale language models, but instead of human language, AI has learned to express chemicals.
Evogene’s generative model is built on a transformer neural network architecture, similar to the GPT model that revolutionizes natural language processing. In fact, the system is called Chempass-GPT, a unique AI model trained with Smiles strings (text encoding of molecular structures). Simply put, Chempass-GPT processes molecules like sentences. The smile string of each molecule is a series of letters representing atoms and bonds. Transmodels learn the grammar of this chemical language and “write” new molecules by predicting one letter at a time, just as GPT can write sentences in letters. Trained with billions of examples, this model can generate new smiles that correspond to chemically effective drug-like structures.
this Sequence-based Generation Approach Use the strength of the transformer when capturing complex patterns. By training such a huge and chemically diverse dataset, Chempass AI overcomes the problems faced by previous AI models, such as biases and redundant or invalid molecules from small datasets. 90% accuracy Traditional GPT-based models provide 29% accuracy and 29% accuracy when generating new molecules that meet all design criteria.evogene.com. In reality, this means that almost every molecule, Chempass AI, is not only new, but also hits the target profile. This is an impressive improvement over the baseline generation technique.
Although Evogene’s main generation engine uses transformers with linear smiles, it is worth noting that the wider AI toolkit includes other architectures such as Graph Neural Networks (GNNS). Molecules are naturally graphical, atoms are bonds as nodes and edges, and GNNs can be inferred directly in these structures. In modern drug design, GNNs are often used to predict properties and generate molecules by constructing them atom-by-atom. This graph-based approach complements the sequence model. For example, Evogene’s platform also includes tools such as Deepdock for 3D virtual screening. This could use deep learning to evaluate molecular binding in a structure-based context by combining sequence models (optimal for creativity and novelty) and graph-based models (optimal for structural accuracy and characterization prediction). The AI design loop could generate candidate structures and evaluate them via predictive models (probably GNN-based) for criteria such as toxicity and synthesis feasibility, creating a feedback cycle that refines each proposal.
Multi-purpose optimization: potency, toxicity, stability At once
A standout feature of Chempass AI is its built-in capabilities for multi-purpose optimization. While classical drug discoveries often optimize one property at a time, Chempass was designed to handle many goals simultaneously. This is achieved through advanced machine learning techniques that guide generative models to meet multiple constraints. In training, Evogene can impose characterization requirements – molecules must strongly activate certain targets, avoid specific toxic motifs, and improve bioavailability. The model learns to navigate the chemical space under these rules. The ChemPass-GPT system enables “Constraint-based Generation.” This means that from the start, we can only instruct them to propose molecules that satisfy a particular desired property.
How does AI achieve this multiparameter balance? One approach is multitasking learning in which the model not only generates molecules, but also uses learning predictors to predict properties and tailor the generation accordingly. Another powerful approach is reinforcement learning (RL). In RL-enhanced workflows, the generative model acts like a “play game” agent in molecular design. Propose a molecule and obtain a reward score based on the extent to which the molecule meets its objective (e.g., lack of potency, toxicity, etc.). Over many iterations, the model fine-tunes the generation strategy to maximize this reward. This method has been successfully used in other AI-driven drug design systems. Researchers have shown that reinforcement learning algorithms can guide generative models to generate molecules with the desired properties. Essentially, AI can train with reward functions that encapsulate multiple goals. For example, you can give points of predicted effectiveness and subtract the points of predicted toxicity. The model then optimizes “movement” (add or removed atoms, altering functional groups) to effectively learn the trade-offs needed to meet all criteria.
Evogene has not revealed the exact unique source behind Chempass AI’s multipurpose engine, but it is clear that such a strategy is working. The fact that each compound produced “meets essential parameters at the same time” such as efficacy, synthesisability, and safety. Future Chempass AI version 2.0 will push this even further. It is developed to allow for even more flexible multi-parameter tuning, including user-defined criteria tailored to the specific therapeutic area or crop requirements. This suggests that next-generation models can dial up or down the importance of certain factors (for example, prioritize the neurological drugs of pesticides or environmental biodegradable brain penetration), and AI adjusts design strategies accordingly. By integrating these multi-purpose features, Chempass AI can design molecules that hit sweet spots at once with a large number of performance metrics. This is actually not possible using traditional methods.
A leap beyond traditional R&D methods
The emergence of Chempass AI generative models highlights a wide range of changes in research and development in life sciences: movements from workflows of tedious trials and errors Ai-Augmented’s creativity and accuracy. Unlike human chemists who tend to stick to known chemistry series and slowly repeat, AI can speculate on billions of possibilities and challenge 99.9% of the chemical space. This opens the door to finding effective compounds that don’t resemble those we’ve seen before. This is important for tackling new chemicals, pests and pathogens that have evolved resistance to existing molecules. Furthermore, by considering Patentability From Get-go, the generated AI helps you avoid busy intellectual property areas. Evogene explicitly aims to produce molecules that open up a key competitive advantage: fresh IP.
The advantages over the traditional approach can be summarized as follows:
-
Parallel Multi-Trait Optimization: AI evaluates many parallel parameters that design molecules that meet efficacy, safety, and other criteria. In contrast, traditional pipelines often only discover toxicity issues after years of work on otherwise promising drugs. By filtering out these questions preemptively, AI-designed candidates would be better able to achieve success in costly later exams.
-
Expanding chemical diversity: Generation models are not limited to existing compound libraries. Chempass AI can recall structures that have never been created before, but is predicted to be effective. this Novelty-driven generation It avoids reinventing wheels (or molecules) and helps in creating differentiated products with new modes of action. Traditional methods often lead to “like me” compounds that offer almost nothing new.
-
Speed and Scale: What a team of chemists can achieve through synthesis and testing in a year is that AI can simulate in a few days. Chempass AI’s deep learning platform can quickly and quickly screen hundreds of billions of complexes in a single run and generate hundreds of new ideas. This dramatically compresses the Discovery timeline and concentrates wet love experiments only on the most promising candidates identified in Silico.
-
Integrated Knowledge: AI models like Chempass incorporate vast amounts of chemical and biological knowledge into their training (known structure activity relationships, toxicity alerts, drug-like characterization rules). Traditional designs rely on the experiences of pharmaceutical chemists (worthy but limited to human memory and bias, whereas AI can capture patterns across millions of experiments and diverse chemical families.
Practically speaking, in the case of pharma, this can lead to higher success rates in clinical trials and lower development costs. In agriculture, it means faster creation of safer and sustainable crop protection solutions. For example, it is a herbicide that is fatal to weeds, but benign to targeted non-target organisms, and decomposes harmlessly in the environment. AI helps to provide “effective, sustainable, unique” ag chemicals by optimizing efficacy and environmental safety, addressing the challenges of regulation and resistance at once.
Part of Evogene’s wider AI toolbox
Chempass AI steals spotlights for small molecule designs, but it’s part of Evogene’s trio of AI-powered “technology engines” tailored to different domains. The company has microboost AI focused on microorganisms, Chempass AI for chemicals, and generator AI for genetic elements. Each engine applies big data analysis and machine learning to each field.
This integrated ecosystem of AI engines highlights Evogene’s strategy as an “AI-First” life sciences company. They aim to revolutionize product discovery across the board, whether it develops drugs, biostimulants, or drought-resistant crops — Exploit calculations to navigate biological complexity. Engines share a common philosophy. Use cutting-edge machine learning to increase the odds of success in R&D, reducing time and costs.
Outlook: AI-driven discoveries are age
Generated AI transforms molecular discoveries and shifts the role of AI from assistants to creative collaborators. Instead of testing one idea at a time, scientists can now use AI to design entirely new compounds that meet multiple goals (safety, stability, etc.) in one step.
This future is already unfolding. Pharmaceutical teams may target specific proteins, avoid the brain, and require molecules that are verbally available. In agriculture, researchers can produce eco-friendly pest controls tailored to regulatory and environmental constraints.
The recent basic model of Evogene, developed in Google Cloud, is an example of this shift. It allows for multiparameter design and opens new areas of chemical space. Future versions will allow for even more customization, making these models an essential tool across life sciences.
Importantly, the impact depends on actual verification. The model improves when molecules produced in AI are tested and refined. Create a powerful feedback loop between calculations and experiments.
This generative approach is not limited to drugs and pesticides. It can quickly drive breakthroughs in ingredients, food and sustainability. This provides faster and smarter discovery across an industry that is constrained by trial and error.
