How generative AI improves hit identification in drug discovery

In the pharmaceutical industry, artificial intelligence is currently transforming the landscape of drug discovery. As medicinal chemists face the challenge of investigating vast chemical spaces while balancing time, price, and success rates, generative AI is a valuable tool that enhances rather than replaces human expertise. This transformation is especially apparent in three key stages of drug discovery: hit identification, hit-to-lead optimization, and lead optimization.

Image Credit: Stock-Asso/Shutterstock.com

The promise and reality of generative AI in therapeutic discovery

Generative AI represents a fundamental change in the approach to molecular design. At its foundation, it is an algorithm that produces novel content, in this case, new molecules, based on patterns recognized in training data. For medicinal chemists, this technology is a robust tool that can generate novel molecules based on desired characteristics, investigate larger chemical spaces that would be humanly impossible to navigate, and reveal insights from intricate datasets.

The numbers highlight the untapped opportunities: while current technologies have enabled just 10⁴ to 10⁹ compounds to be explored in chemical space, there remains a massive unexplored territory that could be critical to drugging previously "undruggable" targets. This is where generative AI's computational power becomes invaluable, prompting increased investigation of chemical space that would otherwise require decades to explore.

However, the reality is more nuanced than the hype suggests. Generative AI supports rather than replaces medicinal chemists. This technology faces considerable hurdles, especially in ensuring that generated molecules are both synthetically accessible and possess the right characteristics for pharmaceutical development. Without accurate guidance and context, AI can produce essentially "garbage" molecules, which are chemically valid but biologically irrelevant.

The critical role of high-quality training data

The success of any generative AI application in pharmaceutical discovery depends on the quality of its training data. This principle is essential: a model's quality corresponds to the data it learns from. The pharmaceutical sector has learned this lesson through experience, as early AI models frequently produced compounds that failed when tested experimentally, resulting in wasted time and resources in the crucial design-make-test-analyze cycle.

The solution lies in using robust, experimental data that has been meticulously curated and validated. This entails training data points that have been assayed consistently, using identical experimental protocols, and representing diverse chemical space. The challenge is that much of the published literature contains positive outcomes, with negative data rarely shared – creating a considerable gap in model training.

Organizations that have successfully adopted generative AI have invested heavily in cleaning and curating their internal datasets, often covering decades of experimental work. This includes not just activity data, but extensive ADMET (Absorption, Distribution, Metabolism, Excretion, and Toxicity) profiles that are crucial for forecasting drug-like characteristics.

Improving hit identification workflows

The first stage of drug discovery is hit identification, where the objective is to identify initial compounds that show activity against a target of interest. While this traditionally involves screening large compound libraries, generative AI is changing this process in several ways.

Expanding chemical space exploration: Generative AI allows for more efficient investigation of both known and unknown chemical spaces. While conventional techniques can be used to explore known chemical space, unknown chemical space – where truly novel compounds reside – requires the pattern recognition capabilities of AI. By training models on diverse chemical datasets, scientists can produce novel scaffolds and chemotypes that might not be present in current compound libraries.

Virtual screening at scale: Contemporary generative AI platforms can screen billions of virtual compounds in minutes, a process that would be impossible through conventional high-throughput screening alone. This ability enables scientists to cast a wider net in their initial hit identification efforts, potentially discovering new starting points for pharmaceutical development.

Property-guided generation: In contrast to random compound generation, modern AI systems can be guided by desired characteristics from the outset. This means that hit identification can concurrently consider variables such as synthetic accessibility, drug-likeness, and basic ADMET properties, decreasing the risk of pursuing compounds that will ultimately fail in subsequent stages.

Revolutionizing hit-to-lead optimization

The hit-to-lead stage is characterized by limited data availability and a focus on investigation rather than exploitation. This phase typically begins with one or multiple verified hits and aims to broaden the chemical series while enhancing potency and drug-like characteristics.

Scaffold hopping and analog design: Generative AI is particularly effective at scaffold hopping – the process of finding structurally distinct compounds that maintain or enhance biological activity. AI models can identify bioisosteric replacements and recommend new scaffolds that scientists might not immediately consider. This ability is especially important when dealing with intellectual property limitations or when aiming to enhance certain characteristics.

Fragment-based therapeutic design: AI supports fragment linking, growing, and merging approaches by forecasting how different molecular fragments might combine to produce stronger compounds. The technology can quickly enumerate possible combinations and predict their attributes, helping scientists prioritize which synthetic targets to pursue.

Multi-parameter optimization: One of the most challenging aspects of hit-to-lead optimization is balancing several competing goals – enhancing potency while preserving selectivity, optimizing ADMET characteristics, and ensuring synthetic viability. Generative AI helps navigate this intricate optimization landscape by producing compounds that concurrently satisfy multiple requirements.

Advancing lead optimization approaches

The most focused stage of medicinal chemistry is lead optimization, where researchers have identified promising chemical series and are working to optimize them for clinical development. This phase is characterized by comprehensive experimental data, requiring advanced modeling strategies.

Structure-Activity Relationship (SAR) analysis: AI can speed up SAR evaluation by identifying trends in large datasets that may not be immediately obvious to human chemists. This includes identifying activity cliffs, recognizing essential pharmacophores, and predicting the effect of certain structural modifications.

Predictive ADMET modeling: Sophisticated machine learning models trained on proprietary pharmaceutical data can provide more precise predictions of ADMET properties than conventional strategies. These models can guide optimization efforts by predicting which modifications are likely to enhance certain characteristics such as metabolic stability, permeability, or safety profiles.

Free energy perturbation integration: Leading-edge platforms incorporate generative AI into more robust computational techniques such as free energy perturbation (FEP) computations. This combination allows for both quick idea generation and in-depth thermodynamic evaluation of protein-ligand interactions.

The value of confidence scoring and model validation

One of the most crucial developments in AI-driven pharmaceutical discovery is the development of confidence scoring systems. These systems provide researchers with an assessment of the reliability of each prediction, helping them to make informed decisions about which compounds to prioritize for synthesis and evaluation.

Confidence scores are produced by evaluating the correlation between model predictions and experimental outcomes across multiple validation sets. These confidence scores are continuously updated as more experimental datasets become available, creating a feedback loop that enhances model dependability over time.

This strategy addresses a critical concern medicinal chemists have about AI predictions: understanding when to trust the model and when to depend on human intuition and experience. High-confidence predictions can guide synthetic priorities, while low-confidence predictions indicate areas where further experimental verification is necessary.

Integration with retrosynthesis planning

A crucial aspect frequently underestimated in AI-driven pharmaceutical design is synthetic accessibility. The most promising compound is effectively worthless if it cannot be synthesized efficiently. Contemporary AI platforms overcome this challenge by integrating generative design with retrosynthesis planning tools.

This integration serves two purposes: first, it provides synthetic accessibility scores during the design phase, helping to filter out compounds that would be challenging or impossible to produce. Second, it generates comprehensive retrosynthetic routes for prioritized compounds, allowing scientists to move quickly from virtual design to bench synthesis.

Combining generative AI with retrosynthesis planning represents a major development in enabling medicinal chemists to access AI-designed compounds.

Democratizing access to AI-powered drug discovery

One of the most substantial advancements in AI-driven pharmaceutical discovery is the democratization of access to advanced AI tools. In the past, only large pharmaceutical companies with comprehensive internal datasets and computational resources could engineer and implement sophisticated AI models. Today, commercial platforms are extending these capabilities to smaller biotechnology companies and academic researchers.

These platforms offer access to models trained using decades of pharmaceutical data, enabling smaller companies to benefit from the same AI capabilities as large pharma. In addition, numerous platforms allow users to incorporate proprietary data, making it possible to tailor models according to specific therapeutic areas or compound classes.

Challenges and limitations

Despite the substantial potential of generative AI in medicinal chemistry, several obstacles remain:

Data quality and bias: AI models can perpetuate biases present in training data. If training sets are dominated by specific chemical classes or therapeutic areas, the models may be less effective for new targets or less-explored chemical space.

Interpretability: Although AI can produce new compounds, understanding why specific molecules are predicted to have certain characteristics is difficult. This "black box" nature can make it challenging for medicinal chemists to build intuition and learn from AI predictions.

Experimental validation: Ultimately, every AI prediction must be experimentally confirmed. The time and expense required for synthesis and assessment continue to be major constraints in the pharmaceutical discovery workflow.

Integration with existing workflows: Successful AI tool implementation requires integration with existing medicinal chemistry processes, data management systems, and decision-making workflows.

The future of AI-enhanced medicinal chemistry

The future of generative AI in medicinal chemistry lies in creating powerful human-AI partnerships rather than replacing human expertise. The most effective applications combine AI’s pattern recognition and computational power with the intuition, creativity, and scientific judgment of skilled medicinal chemists.

Emerging trends include expanding AI abilities beyond small molecules to biologics, PROTACs, and other innovative modalities. In addition, integrating AI with automated synthesis and testing platforms is expected to create closed-loop systems capable of quickly iterating through design-make-test cycles with low manual intervention.

As the technology continues to advance, future AI models will likely be more advanced and better understand the nuances of pharmaceutical discovery, enhanced integration with experimental workflows, and improved interpretability that helps researchers learn from AI predictions.

Conclusion

By improving hit identification, hit-to-lead optimization, and lead optimization processes, generative AI is reshaping the medicinal chemistry landscape. Its ability to investigate massive chemical spaces, predict molecular characteristics, and recommend new synthetic targets is already speeding up pharmaceutical discovery timelines and enhancing success rates.

Successful adoption, however, depends on understanding both the strengths and limitations of AI. High-quality training data, robust validation techniques, and confidence scoring systems are crucial for building trust and ensuring dependable predictions. Most importantly, AI should be regarded as a valuable complement to human expertise rather than a substitute for it.

As the field continues to progress, drug discovery organizations that effectively integrate AI capabilities with human creativity and scientific judgment will be the most successful, creating synergistic collaborations that drive discovery and innovation in medicinal chemistry. The combination of both human intelligence and artificial intelligence will ultimately shape the future of drug discovery, allowing researchers to overcome the most complex problems in human health.

About Merck KGaA

Merck Millipore’s purpose is to solve the toughest problems in life science by collaborating with the global scientific community – and through that, we aim to accelerate access to better health for people everywhere.

We provide scientists and engineers with best-in-class lab materials, technologies and services. With the 2015 combination of Merck Millipore and Sigma-Aldrich, we now have a broad portfolio of 300,000 products, an expanded global footprint and an industry-leading eCommerce platform - SigmaAldrich.com.

We are dedicated to making research and biotech production simpler, faster and safer.

Our three life science business areas are designed to best address our customers’ needs:

Research Solutions: Serve customers focused on identifying and developing medicines.

Process Solutions: Deliver end-to-end products and expertise to customers who take what is developed in labs and manufacture it.

Applied Solutions: Support customers in ensuring that drugs, food and beverages are safe for consumption.

Sponsored Content Policy: News-Medical.net publishes articles and related content that may be derived from sources where we have existing commercial relationships, provided such content adds value to the core editorial ethos of News-Medical.net, which is to educate and inform site visitors interested in medical research, science, medical devices and treatments.