My Perspective As an AI/ML student, I am particularly interested in how Generative AI can move beyond content generation and contribute to scientific discovery. Drug discovery is an interesting application because it combines large-scale data, machine learning, optimization, domain expertise and responsible AI. In my view, the most important question is not whether AI can replace scientists, but how AI can help scientists explore complex problems faster while keeping humans responsible for validation and decision-making. Introduction Artificial Intelligence has evolved from systems designed mainly for prediction and classification into systems capable of generating new content, designs and hypotheses. Generative AI is now being explored in healthcare, biotechnology, scientific research, software development and many other areas. One particularly promising application is drug discovery and biomolecular design. Traditional drug discovery is a long and complex process. Researchers need to identify biological targets, search for promising molecules, evaluate their properties, optimize candidate compounds, conduct laboratory experiments and eventually perform preclinical and clinical testing. Each stage requires significant scientific expertise, computational resources, time and experimentation. A major challenge is the enormous size of chemical and biological search spaces. There are an extremely large number of theoretically possible molecular structures, making it impossible to experimentally test every possibility. Generative AI provides a different approach. Instead of only predicting the properties of existing molecules, generative models can learn patterns from chemical and biological data and propose new candidate molecules or biomolecules according to desired objectives. This creates an important shift: Traditional machine learning often asks: "What properties does this molecule have?" Generative AI can help ask: "What molecules could potentially have the properties we are looking for?" This concept is closely related to inverse design and can become a powerful tool for scientific discovery. How Generative AI Can Support Drug Discovery A simplified Generative AI workflow for drug discovery can be viewed as: Data → Representation → Model → Generation → Filtering → Validation → Experimental Feedback The first stage is data. AI systems can learn from molecular structures, biological sequences, protein information, experimental measurements, scientific literature and other relevant datasets. The next challenge is representation. Molecules can be represented using SMILES sequences, molecular graphs, fingerprints or learned embeddings. Proteins can be represented using amino-acid sequences, structural information or learned representations. The representation selected affects what the model can learn. For example, a graph-based representation naturally captures relationships between atoms and chemical bonds, while a sequence representation can allow transformer-based models to learn patterns in molecular strings. After representation, an appropriate machine learning or deep learning model can be trained. The model can then generate or prioritize potential candidates. However, generation is only one stage. Generated candidates must be filtered according to chemical, biological and practical constraints before they can be considered useful. Generative Models for Molecules Different AI architectures can contribute to molecular generation and optimization. Transformers Transformers have become one of the most influential architectures in AI because they can learn relationships between elements in sequences. Similar approaches can be applied to molecular sequences such as SMILES representations and biological sequences. A transformer-based model can learn patterns from existing molecular data and generate new sequences that follow learned patterns. However, a generated sequence being syntactically valid does not automatically mean that it represents a useful or biologically active molecule. Additional validation is required. Graph Neural Networks Molecules naturally have a graph structure in which atoms can be represented as nodes and chemical bonds as edges. Graph Neural Networks can therefore learn information directly from molecular graphs. They can be useful for molecular property prediction, representation learning and molecular generation. This demonstrates an important principle in AI engineering: the representation should reflect the structure of the underlying problem. Variational Autoencoders Variational Autoencoders can learn latent representations of data. In molecular applications, a VAE can learn a latent space representing molecular structures and generate new candidates by sampling from that space. This can allow researchers to explore relationships between different molecular structures and optimize candidate properties. Diffusion Models Diffusion models are another important class of generative models. They learn to generate structured data by progressively transforming noisy representations into meaningful samples. In scientific applications, diffusion-based approaches can be explored for molecular and protein structure generation and design. The important engineering lesson is that the newest architecture is not automatically the best solution. Model selection should depend on the problem, data, constraints, computational resources and evaluation requirements. From Prediction to Inverse Design One of the most interesting aspects of Generative AI in drug discovery is inverse design. Suppose researchers want a molecule with characteristics such as: - strong interaction with a target - acceptable molecular properties - suitable solubility - low predicted toxicity - chemical stability - appropriate drug-like characteristics A conventional predictive model may estimate these properties for an existing molecule. A generative system can instead attempt to produce new molecular structures that satisfy multiple objectives. A simplified workflow could be: Desired properties → Generate candidates → Predict properties → Rank candidates → Filter candidates → Validate → Repeat This creates a feedback loop in which generation and prediction work together. However, optimizing one property can negatively affect another. For example, improving predicted binding affinity might result in poor solubility or unacceptable toxicity. Therefore, molecular design is a multi-objective optimization problem rather than simply finding the "best" molecule. AI for Protein and Biomolecule Design Generative AI is not limited to small-molecule drugs. Proteins, peptides, antibodies and other biomolecules can also be represented computationally and explored using AI. Proteins are particularly interesting because their amino-acid sequences are related to complex three-dimensional structures and biological functions. AI models can learn relationships between protein sequences, structures and functions. Generative approaches can then explore new sequences or structures that may have desired properties. Potential applications include: - Protein engineering - Enzyme design - Therapeutic protein development - Peptide design - Antibody engineering - Biomolecular optimization This creates an important intersection between AI, deep learning and computational biology. For an AI engineer, this also demonstrates why domain knowledge matters. A model may generate a technically valid sequence, but determining whether that sequence has meaningful biological function requires scientific validation. The Role of Large Language Models Large Language Models can also contribute to drug discovery, but their role should be understood carefully. Scientific research produces enormous amounts of information through research papers, patents, clinical information, databases and experimental reports. LLMs can help researchers interact with this information using natural language. For example, a researcher could ask: "What research has been conducted on this protein target?" An AI system could retrieve relevant scientific documents, summarize them and organize the findings. However, relying only on an LLM's internal knowledge can introduce hallucination. This is where Retrieval-Augmented Generation can become useful. RAG for Scientific Knowledge A Retrieval-Augmented Generation system can follow a workflow such as: User Question → Retrieve Relevant Documents → Rank Context → Provide Context to LLM → Generate Answer Instead of asking an LLM to generate an answer entirely from its internal parameters, relevant information can first be retrieved from a trusted knowledge base. For scientific applications, the knowledge base could contain research papers, validated datasets, technical documentation or other approved sources. Embeddings can represent documents and queries in a semantic vector space. A vector database can then be used to identify relevant information. This is particularly interesting from an AI engineering perspective because it connects: - Document processing - Embeddings - Semantic search - Vector databases - Retrieval - LLMs - Prompt engineering - Response validation However, RAG is not a guarantee of correctness. If the retrieved information is incorrect, outdated or irrelevant, the generated answer may still be unreliable. Therefore, retrieval quality and source quality must be evaluated. Major Challenges 1. Data Quality AI models depend heavily on training data. Scientific datasets can contain missing values, inconsistent measurements, experimental noise, duplicate information and biases. A powerful model trained on poor-quality data can still produce unreliable results. Therefore: Poor data → unreliable learning → unreliable predictions or generation. Data preprocessing, validation and governance are as important as model architecture. 2. Hallucination Generative AI models can produce outputs that appear convincing but are incorrect. In scientific applications, this is particularly serious. An AI-generated statement about a molecule, protein or biological mechanism should not automatically be considered scientifically valid. Systems therefore need mechanisms for verification and human review. 3. Chemical and Biological Validity A model may generate a molecule that looks plausible computationally but may not be stable, synthesizable or biologically useful. Similarly, a generated protein sequence may satisfy a model's learned objectives without having the desired real-world function. AI generation must therefore be combined with domain-specific computational and experimental validation. 4. Generalization A model can perform well on a benchmark dataset but fail when applied to a new biological target or unfamiliar chemical space. This is an example of an out-of-distribution problem. Good evaluation should therefore go beyond a single benchmark score. 5. Interpretability Scientists often need to understand why a model recommends a particular candidate. If a model produces a prediction without meaningful evidence or explanation, researchers may have difficulty trusting or validating the result. Interpretability and uncertainty estimation can therefore become important components of scientific AI systems. 6. Bias If training data underrepresents particular biological populations, diseases or experimental conditions, the resulting model can inherit those limitations. This demonstrates why responsible AI should be considered throughout the development lifecycle rather than added at the end. Responsible AI and Safety Healthcare and drug discovery require a high level of responsibility because incorrect decisions can potentially affect human health. An AI system used in this domain should consider: - Data governance - Privacy - Security - Model validation - Reproducibility - Uncertainty - Traceability - Human oversight - Documentation - Regulatory requirements I believe the principle should be: AI should assist scientific decision-making, not replace scientific responsibility. A generated molecule should be treated as a candidate. A predicted biological property should be treated as a hypothesis. An LLM-generated scientific explanation should be treated as information requiring appropriate verification. This mindset can help prevent overconfidence in AI systems. MLOps and Continuous Model Governance Another important consideration is the engineering side of deploying AI systems. A model that performs well during development may behave differently after deployment because data distributions can change. Changes in experimental protocols, datasets or biological populations can affect model performance. Therefore, AI systems should have lifecycle management processes: Data Validation → Model Validation → Deployment → Monitoring → Evaluation → Updating Monitoring can include data drift, model performance, prediction distributions, error rates and system reliability. This is where MLOps becomes important. For high-impact scientific applications, model governance should continue throughout the model lifecycle rather than stopping after deployment. Human-AI Collaboration I believe the most realistic future is not AI replacing scientists but AI augmenting scientists. AI is particularly strong at processing large datasets, identifying patterns, generating candidate solutions and exploring large search spaces. Scientists contribute domain knowledge, experimental judgment, interpretation and responsibility. A future drug discovery workflow could look like: Human defines objective ↓ AI generates candidate solutions ↓ Computational models evaluate candidates ↓ Scientists review promising candidates ↓ Laboratory experiments validate hypotheses ↓ Experimental results become new data ↓ AI systems improve This creates a human-AI feedback loop. My Engineering Perspective From an AI/ML engineering perspective, this application demonstrates that building a useful AI system requires much more than selecting a model. A complete system needs: 1. A clearly defined problem 2. Appropriate and reliable data 3. Meaningful representations 4. A suitable model 5. Evaluation metrics 6. Validation mechanisms 7. Monitoring 8. Human oversight 9. Responsible deployment For example, achieving high accuracy on a benchmark does not necessarily mean that an AI system is ready for real-world use. The more important question is: "Does the system solve the intended problem reliably under realistic conditions?" This mindset is important not only in healthcare AI but in every engineering application. Future Opportunities I believe the future of AI in drug discovery will involve increasingly integrated systems. Instead of using separate tools for literature search, molecular generation, property prediction and analysis, future platforms may combine multiple capabilities into intelligent scientific workflows. Potential developments include: - Multimodal AI combining text, molecular structures and biological data - AI agents assisting researchers with scientific workflows - Improved molecular and protein generation - Automated hypothesis generation - Stronger scientific RAG systems - Better uncertainty estimation - Integration with computational simulations - AI-assisted laboratory experimentation - Learning from experimental feedback However, progress should be measured not simply by how impressive a model appears, but by whether it produces scientifically useful, reliable and reproducible outcomes. Conclusion Generative AI has the potential to transform drug discovery and biomolecular engineering by helping researchers explore enormous chemical and biological search spaces. Its applications can extend from molecular generation and property optimization to protein design, scientific knowledge retrieval and AI-assisted research workflows. At the same time, AI-generated outputs should never automatically be treated as scientific truth. Data quality, hallucination, chemical validity, biological validity, generalization, interpretability, safety and regulatory considerations remain important challenges. The strongest future approach will therefore combine: AI + Scientific Data + Domain Expertise + Computational Validation + Experimental Evidence + Responsible Governance I believe this combination represents the real opportunity of Generative AI in healthcare. For me, the most exciting part is not simply that AI can generate new molecules or biomolecules. It is that AI can become a tool for helping researchers ask better questions, explore more possibilities and accelerate the process of turning scientific hypotheses into validated discoveries. The future of AI in drug discovery should therefore be viewed as human intelligence amplified by machine intelligence, rather than machines replacing scientific expertise. As an aspiring AI/ML engineer, I want to develop the ability not only to build models but also to understand the problems they are solving, evaluate their limitations, and design systems that are useful, reliable and responsible in real-world environments.