Evogene Ltd. has unveiled a first-in-class generative AI foundation model for small-molecule design, marking a breakthrough in how recent compounds are discovered. Announced on June 10, 2025, in collaboration with Google Cloud, the model expands Evogene’s ChemPass AI platform and tackles a long-standing challenge in each pharmaceuticals and agriculture: finding novel molecules that meet multiple complex criteria concurrently. This development is poised to speed up R&D in drug discovery and crop protection by enabling the simultaneous optimization of properties like efficacy, toxicity, and stability in a single design cycle.
From Sequential Screening to Simultaneous Design
In traditional drug and agriculture chemical research, scientists often test one factor at a time—first checking if a compound works, then later testing for safety and stability. This step-by-step method is slow, expensive, and infrequently ends in failure, with many promising compounds falling short in later stages. It also keeps researchers focused on familiar chemical structures, limiting innovation and making it harder to create recent, patentable products. This outdated approach contributes to high costs, long timelines, and a low success rate—around 90% of drug candidates fail before reaching the market.
Generative AI changes this paradigm. As a substitute of one-by-one filtering, AI models can juggle multiple requirements directly, designing molecules to be potent and protected and stable from the beginning. Evogene’s recent foundation model was explicitly built to enable this simultaneous multi-parameter design. This approach goals to de-risk later phases of development by front-loading considerations like ADME and toxicity into the initial design.
In practice, it could mean fewer late-stage failures – as an example, fewer drug candidates that show great lab results only to fail in clinical trials as a result of unintended effects. In brief, generative AI allows researchers to innovate faster and smarter, concurrently optimizing for the numerous facets of a successful molecule reasonably than tackling each in isolation.
Inside ChemPass AI: How Generative Models Design Molecules
At the guts of Evogene’s ChemPass AI platform is a strong recent foundation model trained on an infinite chemical dataset. The corporate assembled a curated database of roughly 40 billion molecular structures– spanning known drug-like compounds and diverse chemical scaffolds – to show the AI the “language” of molecules. Using Google Cloud’s Vertex AI infrastructure with GPU supercomputing, the model learned patterns from this vast chemical library, giving it an unprecedented breadth of information on what drug-like molecules appear like. This massive training regimen is akin to training a big language model, but as a substitute of human language, the AI learned chemical representations.
Evogene’s generative model is built on transformer neural network architecture, much like the GPT models that revolutionized natural language processing. In actual fact, the system is known as ChemPass-GPT, a proprietary AI model trained on SMILES strings (a text encoding of molecular structures). In easy terms, ChemPass-GPT treats molecules like sentences – each molecule’s SMILES string is a sequence of characters describing its atoms and bonds. The transformer model has learned the grammar of this chemical language, enabling it to “write” recent molecules by predicting one character at a time, in the identical way GPT can write sentences letter by letter. Since it was trained on billions of examples, the model can generate novel SMILES that correspond to chemically valid, drug-like structures.
This sequence-based generative approach leverages the strength of transformers in capturing complex patterns. By training on such an enormous and chemically diverse dataset, ChemPass AI overcomes problems that earlier AI models faced, like bias from small datasets or generating redundant or invalid molecules The inspiration model’s performance already far outstrips a generic GPT applied to chemistry: internal tests showed about 90% precision in producing novel molecules that meet all design criteria, versus ~29% precision for a conventional GPT-based modelevogene.com. In practical terms, this implies nearly all molecules ChemPass AI suggests should not only recent but in addition hit their goal profile, a striking improvement over baseline generative techniques.
While Evogene’s primary generative engine uses a transformer on linear SMILES, it’s value noting the broader AI toolkit includes other architectures like graph neural networks (GNNs). Molecules are naturally graphs – with atoms as nodes and bonds as edges – and GNNs can directly reason on these structures. In modern drug design, GNNs are sometimes used to predict properties and even generate molecules by constructing them atom-by-atom. This graph-based approach complements sequence models; for instance, Evogene’s platform also incorporates tools like DeepDock for 3D virtual screening, which likely use deep learning to evaluate molecule binding in a structure-based context By combining sequence models (great for creativity and novelty) with graph-based models (great for structural accuracy and property prediction), ChemPass AI ensures its generated compounds should not just novel on paper, but in addition chemically sound and effective in practice. The AI’s design loop might generate candidate structures after which evaluate them via predictive models – some possibly GNN-based – for criteria like toxicity or synthetic feasibility, making a feedback cycle that refines each suggestion.
Multi-Objective Optimization: Potency, Toxicity, Stability All at Once
A standout feature of ChemPass AI is its built-in ability for multi-objective optimization. Classic drug discovery often optimizes one property at a time, but ChemPass was engineered to handle many objectives concurrently. That is achieved through advanced machine learning techniques that guide the generative model toward satisfying multiple constraints. In training, Evogene can impose property requirements – corresponding to a molecule must activate a certain goal strongly, avoid certain toxic motifs, and have good bioavailability – and the model learns to navigate chemical space under those rules. The ChemPass-GPT system even enables “constraints-based generation,” meaning it could be instructed to only propose molecules that meet specific desired properties from the outset.
How does the AI accomplish this multi-parameter balancing act? One approach is multi-task learning, where the model just isn’t just generating molecules but in addition predicting their properties using learned predictors, adjusting generation accordingly. One other powerful approach is reinforcement learning (RL). In an RL-enhanced workflow, the generative model acts like an agent “playing a game” of molecule design: it proposes a molecule after which gets a reward rating based on how well that molecule meets the objectives (potency, lack of toxicity, etc.). Over many iterations, the model tweaks its generation technique to maximize this reward. This method has been successfully utilized in other AI-driven drug design systems – researchers have shown that reinforcement learning algorithms can guide generative models to provide molecules with desirable properties. In essence, the AI could be trained with a reward function that encapsulates multiple goals, for instance giving points for predicted efficacy and subtracting points for predicted toxicity. The model then optimizes its “moves” (adding or removing atoms, altering functional groups) to net the best rating, effectively learning the trade-offs needed to satisfy all criteria.
Evogene hasn’t disclosed the precise proprietary sauce behind ChemPass AI’s multi-objective engine, nevertheless it’s clear from their results that such strategies are at work. The indisputable fact that each generated compound “concurrently meets essential parameters” like efficacy, synthesizability and safety. The upcoming ChemPass AI version 2.0 will push this further – it’s being developed to permit much more flexible multi-parameter tuning, including user-defined criteria tailored to specific therapeutic areas or crop requirements. This implies the next-gen model could let researchers dial up or down the importance of certain aspects (as an example, prioritizing brain penetrance for a neurology drug or environmental biodegradability for a pesticide) and the AI will adjust its design strategy accordingly. By integrating such multi-objective capabilities, ChemPass AI can design molecules that hit the sweet spot on quite a few performance metrics directly, a feat practically unattainable with traditional methods.
A Leap Beyond Traditional R&D Methods
The arrival of ChemPass AI’s generative model highlights a wider shift in life-science R&D: the move from laborious trial-and-error workflows to AI-augmented creativity and precision. Unlike human chemists, who are likely to keep on with known chemical series and iterate slowly, an AI can fathom billions of possibilities and enterprise into the unexplored 99.9% of chemical space. This opens the door to finding efficacious compounds that don’t resemble anything we’ve seen before – crucial for treating diseases with novel chemistry or tackling pests and pathogens which have evolved resistance to existing molecules. Furthermore, by considering patentability from the get-go, generative AI helps avoid crowded mental property areas. Evogene explicitly goals to provide molecules that carve out fresh IP, a vital competitive advantage.
The advantages over traditional approaches could be summarized as follows:
-
Parallel Multi-Trait Optimization: The AI evaluates many parameters in parallel, designing molecules that satisfy potency, safety, and other criteria. Traditional pipelines, in contrast, often only discover a toxicity issue after years of labor on an otherwise promising drug. By preemptively filtering for such issues, AI-designed candidates have a greater shot at success in costly later trials.
-
Expanding Chemical Diversity: Generative models aren’t limited to existing compound libraries. ChemPass AI can conjure structures which have never been made before, yet are predicted to be effective. This novelty-driven generation avoids reinventing the wheel (or the molecule) and helps create differentiated products with recent modes of motion. Traditional methods often result in “me-too” compounds that supply little novelty.
-
Speed and Scale: What a team of chemists might achieve via synthesis and testing in a yr, an AI can simulate in days. ChemPass AI’s deep learning platform can virtually screen tens of billions of compounds rapidly and generate a whole lot of novel ideas in a single run. This dramatically compresses the invention timeline, focusing wet-lab experiments only on probably the most promising candidates identified in silico.
-
Integrated Knowledge: AI models like ChemPass incorporate vast amounts of chemical and biological knowledge (e.g. known structure-activity relationships, toxicity alerts, drug-like property rules) of their trainingThis means every molecule design advantages from a breadth of prior data no single human expert could hold of their head. Traditional design relies on the experience of medicinal chemists – invaluable but limited to human memory and bias – whereas the AI can capture patterns across hundreds of thousands of experiments and diverse chemical families.
In practical terms, for pharma this could lead on to higher success rates in clinical trials and reduced development costs, since fewer resources are wasted on doomed compounds. In agriculture, it means faster creation of safer, more sustainable crop protection solutions – for instance, an herbicide that’s lethal to weeds but benign to non-target organisms and breaks down harmlessly within the environment. By optimizing across efficacy and environmental safety together, AI may also help deliver “effective, sustainable, and proprietary” ag-chemicals, addressing regulatory and resistance challenges in a single go.
A part of a Broader AI Toolbox at Evogene
While ChemPass AI steals the highlight for small-molecule design, it’s a part of Evogene’s trio of AI-powered “tech-engines” tailored to different domains. The corporate has MicroBoost AI specializing in microbes, ChemPass AI on chemistry, and GeneRator AI on genetic elements. Each engine applies big-data analytics and machine learning to its respective field.
This integrated ecosystem of AI engines underscores Evogene’s strategy as an “AI-first” life science company. They aim to revolutionize product discovery across the board – whether it’s formulating a drug, a bio-stimulant, or a drought-tolerant crop – by harnessing computation to navigate biological complexity. The engines share a standard philosophy: use cutting-edge machine learning to extend the probability of R&D success and reduce time and price.
Outlook: AI-Driven Discovery Comes of Age
Generative AI is transforming molecule discovery, shifting AI’s role from assistant to creative collaborator. As a substitute of testing one idea at a time, scientists can now use AI to design entirely recent compounds that meet multiple goals—potency, safety, stability, and more—in a single step.
This future is already unfolding. A pharmaceutical team might request a molecule that targets a particular protein, avoids the brain, and is orally available—AI can deliver candidates on demand. In agriculture, researchers could generate eco-friendly pest controls tailored to regulatory and environmental constraints.
Evogene’s recent foundation model, developed with Google Cloud, is one example of this shift. It enables multi-parameter design and opens recent areas of chemical space. As future versions allow much more customization, these models will turn into essential tools across life sciences.
Crucially, the impact will depend on real-world validation. As AI-generated molecules are tested and refined, models improve—creating a strong feedback loop between computation and experimentation.
This generative approach isn’t limited to drugs or pesticides. It could soon drive breakthroughs in materials, food, and sustainability—offering faster, smarter discovery across industries once constrained by trial and error.
