1. Why AI in Drug Discovery Is in the Spotlight Now
“AI in drug discovery” has become a buzzword in recent years, but it is not a magic technology that appeared overnight. Traditional in silico approaches – QSAR, docking, and statistical models – have been used for decades. What has changed is that deep learning, generative AI, and foundation models have dramatically increased the scale, speed and quality of what we can do.
Several structural pressures drive this trend:
- Escalating R&D costs per approved drug
- Long development timelines and low success rates
- A shift toward tough disease areas such as oncology, rare diseases, and CNS disorders
- An explosion of data: omics, images, and electronic health records (EHR)
In this environment, AI is emerging as a search engine that tells us where to look across the R&D value chain. It does not autonomously “invent drugs” on its own; rather, it accelerates human hypothesis generation and experimental decision-making.
2. Where Does AI Sit in the Drug Discovery Value Chain?
Let us briefly revisit a typical drug discovery and development value chain and see where AI can plug in.
- (1) Disease understanding and target identification/validation
AI analyzes omics, network data, and CRISPR screen results to identify new targets and patient subtypes. - (2) Hit finding
In small molecules and biologics, AI assists virtual screening and de novo design to generate hit candidates. - (3) Lead optimization
AI supports multi-parameter optimization of potency, selectivity, ADMET, and synthetic accessibility, and helps prioritize “what to make next.” - (4) Preclinical and safety assessment
AI models can predict toxicity risks using in vitro, in vivo, and historical safety data. - (5) Clinical trial design and execution
AI contributes to trial design, patient recruitment, and biomarker discovery, helping identify the right patients and endpoints. - (6) Post-marketing and real-world evidence
AI mines EHR and registry data for real-world effectiveness, safety signals, and appropriate use.
Later parts of this series will dive into specific modalities and stages. In Part 1, the primary goal is to build an intuitive map of AI’s role across this entire chain.
3. What AI Can Do Well in Drug Discovery
Here we summarize cross-cutting strengths of AI that apply across many modalities.
3-1. Narrowing Down the Search Space and Prioritizing Candidates
The chemical space for small molecules, for example, is astronomically large (often quoted as 1060 or more). We cannot experimentally test more than a tiny fraction. By learning from historical data and public databases, AI can:
- Predict structures likely to show activity against a target
- Flag chemotypes with lower toxicity risk
- Suggest properties that favor distribution to specific tissues
This allows us to focus experimental resources on a smaller, more promising subset and potentially improve hit rates.
3-2. Supporting Multi-Parameter Optimization (MPO)
In real drug discovery projects, high potency alone is never enough. Selectivity, safety, solubility, metabolic stability, and ease of synthesis all matter. AI combines prediction models for these properties and provides a composite score, making trade-offs more explicit and quantifiable.
3-3. Integrating Heterogeneous Data and Finding Patterns
Modern projects routinely generate multi-modal data: genomics, transcriptomics, proteomics, pathology images, and clinical records for the same patient. AI is particularly powerful at:
- Discovering molecular subtypes of disease
- Extracting prognostic and treatment-response patterns
- Identifying novel targets and biomarkers
Deep learning is especially suited to high-dimensional, non-linear relationships in images and time-series data.
3-4. Automating Retrosynthesis and Experimental Design
In small-molecule projects, AI-based retrosynthesis can propose synthetic routes, and robotics can execute them. Together, they enable:
- AI proposing candidate molecules
- Robots synthesizing and testing them
- Data feeding back into AI to propose the next set
This kind of closed-loop optimization is beginning to extend beyond chemistry to antibody sequence optimization and cell culture condition optimization.
4. What AI Cannot (Yet) Do – Key Limitations
To avoid unrealistic expectations, we must be equally clear about what AI cannot do today.
4-1. AI Alone Cannot Fully Explain Biology
Most AI models learn statistical correlations, not genuine causality. They cannot fully answer questions such as “Why does this molecule work?” or “Why does toxicity emerge only in this patient subgroup?” Many crucial factors – experimental nuances, tissue–tissue interactions, and rare mechanisms – remain invisible to the data.
Scientific reasoning – generating hypotheses and designing experiments to test them – remains a core human responsibility.
4-2. Performance Collapses Outside the Training Distribution
AI models can perform poorly in data regimes they have never seen, including:
- Emerging modalities (e.g., PROTACs, molecular glues, new BNCT carriers)
- Rare diseases and small patient subgroups
- Under-represented populations and regions
In such cases, blind trust in model predictions is risky. We need frameworks to quantify uncertainty and robust human review.
4-3. Black-Box Models Create Regulatory and Accountability Risks
As we get closer to patients and regulators, explainability matters. Black-box models face skepticism in:
- Internal decision boards and safety committees
- Interactions with regulatory agencies
- Communication with physicians and patients
Explainable AI is an active research area, but we are still far from universally convincing, “courtroom-ready” explanations.
5. Stakeholder Perspectives: What AI in Drug Discovery Really Means
Different stakeholders see very different “pictures” when they hear the term “AI in drug discovery.” This series targets several key audiences.
5-1. Scientists and Developers (R&D, Translational, DMPK, Toxicology)
Goal: Use AI as an extra pair of hands and eyes, not as a threat.
The series helps connect your domain-specific data and expertise with practical AI use cases.
5-2. Headquarters Functions (Corporate Strategy, DX, IT, Corporate Planning)
Goal: Avoid launching isolated, one-off AI pilots and instead build a coherent company-wide AI strategy.
We aim to provide a common language for deciding where – and in what order – to deploy AI for maximum impact.
5-3. Investors, Consultants, and Strategy Professionals
Goal: Learn what to ask and where to look when evaluating AI drug discovery strategies and startups.
We move beyond “we use AI” and focus on data assets, modality focus, and business models.
6. Early Signs of Real-World Impact
When asked “Has AI already delivered results in drug discovery?”, the honest answer is nuanced. So far we see:
- Multiple AI-designed small-molecule candidates entering clinical trials
- Reports of shorter timelines to reach candidate selection compared with traditional approaches
- Marked improvements in protein structure prediction supporting target understanding and antibody design
- Virtual screening and retrosynthesis contributing to efficiency gains in existing pipelines
At the same time:
- Fully AI-originated blockbusters on the market are still rare
- Truly “AI-only” conceptual breakthroughs are still emerging
It is important to distinguish between “efficiency gains in existing workflows” and “new molecules that would have been unreachable without AI”. This series will highlight both types of impact as we move through specific modalities and stages.
7. Setting the Right Expectations for Learning AI in Drug Discovery
Finally, let us summarize a healthy mindset for learning and applying AI in drug discovery.
- AI is an accelerator, not a replacement, for drug discovery.
- No data, no AI. Data strategy is AI strategy.
- Knowing what AI cannot do is a competitive advantage.
- Human roles evolve but do not disappear – hypothesis generation, experimental design, interpretation, and accountability remain human tasks.
From Part 2 onward, we will delve into data and algorithms, modality-specific applications, clinical and business impacts, and long-term risks. After Part 1, it is enough if you have:
- A mental map of where AI sits along the R&D-to-market value chain
- A clear picture of AI’s general strengths and limitations
- An understanding of what “AI in drug discovery” means from your own professional vantage point
My Thoughts and Future Outlook
Public discourse around AI in drug discovery tends to swing between extremes: enthusiastic claims that “AI will revolutionize pharma” and skeptical views that it will eventually disappoint. Reality sits somewhere in between. AI does not guarantee the “quality” of scientific hypotheses, but it does allow us to iterate hypotheses and experiments much faster. By shrinking the practical search space to something manageable, AI frees human intuition and experience to focus on deeper, more meaningful questions – where to look, which mechanisms to prioritize, and how to design the next decisive experiment.
Whether organizations can truly capture the value of AI will depend less on any single algorithm and more on data design and operating models. Which modalities and stages should come first? What data do we collect, how do we share it, and where do we draw the line? Players who patiently build robust data foundations and cross-functional workflows – rather than chasing the latest buzzwords – are likely to win over the next decade. In this series, I will not only cover technical topics but also connect the dots between lab and headquarters, and between business, investment, and policy, to draw a realistic picture of AI in drug discovery and explore where the field might be heading.
This article has been edited by the Morningglorysciences team.
Related Articles
Comments