AI in Drug Discovery: Surveying the Breadth of the Challenges

The biotech industry is in the midst of an artificial intelligence (AI) investment surge, with companies pouring billions into the promise of faster, smarter drug discovery. As leaders in R&D, we have a responsibility to ask a hard question: are we investing in real improvements in effectiveness, or are we chasing a narrative? The honest answer, for now, is that it depends on what you’re asking AI to do, how you’re measuring success, and whether your expectations are grounded in what the science can deliver today.

This is not a pessimistic view. AI and machine learning are genuinely transforming parts of our work, and the algorithmic advances of the past decade, running in parallel with the explosion in compute power, represent a real step forward. But realizing that value requires strategic clarity, not hyperbole. The companies that will benefit most from AI are those that deploy it thoughtfully: identifying specific problems it can solve well, building the data infrastructure to support it, and maintaining the scientific rigor that remains essential even in the emerging era of AI drug discovery.

What AI is good at and where it struggles

AI excels at pattern recognition within familiar territory. When a model has seen thousands of variations of a problem, it can identify structure and make reliable predictions. This is why AI has transformed fields like image recognition, transcription, and language. The training data is abundant and the task is essentially one of recall and interpolation.

Drug discovery, in the areas of chemistry and biology, is a fundamentally different challenge. The relevant chemical space is almost incomprehensibly vast, with estimates suggesting more than 10×10⁶⁰ potential drug-like molecules. The goal is to invent something new: a molecule that doesn’t yet exist, that will behave safely and effectively in a biological system. Advances in basic biology and our biological mechanisms underpinning disease remain inadequate in many cases. If our best understanding of challenging diseases (like Alzheimer’s and pancreatic cancer) is rudimentary today, a highly sophisticated LLM model won’t be able to find answers in scientific publications where they don’t exist. AI models trained on historical data are, by definition, extrapolating beyond what they have seen. How well they do that, what researchers call performance at distance from training data or within the domain of applicability, is one of the central open questions in the field.

Even celebrated breakthroughs like AlphaFold demonstrate this dynamic. These models perform well when structurally similar examples exist in training data, but performance degrades significantly for truly novel cases. 

One of the fundamental intellectual challenges in drug discovery is target selection, a multi-parametric decision with many aspects of scientific judgment. An effective team of scientists will not only assess a target across many of its aspects, even when data are incomplete, but they must also decide how to weigh those various aspects to yield good decision-making. AI currently cannot perform this higher-level task. Looking ahead, it may be possible for AI to review a curated dataset of target selection decisions and the rationale for those decisions. Target selection is just one example, but it illustrates a broader point – understanding where the science stands is essential to making optimal investment decisions and to ensuring that AI is guided by, rather than substituting for, experienced drug developers.

The data problem is underappreciated

Behind every AI model is a training dataset, and in drug discovery, that dataset is expensive to generate, difficult to curate, and imperfect by nature. Unlike consumer applications where training data can be harvested at scale, pharmaceutical data comes from carefully controlled experiments that take months and cost significant resources. In addition, a meaningful portion of it is not directly comparable across experiments. Different assay conditions, cell types, and timepoints can produce data that look similar but are not.

This is distinct from the broader reproducibility challenges in biomedical research, though those are real too. The more immediate issue for AI applications is data curation: ensuring that models are trained on internally consistent and scientifically comparable data. This kind of curation requires deep domain expertise and sustained investment, and it is not a problem that AI can solve on its own currently.

Where to focus: Targeted efficiency over broad promises

Given these constraints, the most productive R&D strategy is to identify the specific, rate-limiting steps in drug discovery where AI can deliver reliable efficiency gains and focus investment there, rather than asking AI to autonomously explore chemical space or replace scientific judgment.

One well-suited application is helping teams navigate the sheer volume of candidate data generated during drug design. AI can sift through large datasets and evaluate proposed molecules against known parameters, e.g., potency, selectivity, and ADME properties, to help scientists prioritize which candidates are worth pursuing. Rather than replacing scientific judgment, AI can augment it by surfacing the most relevant information faster and reducing the time spent on manual triage.

AI also has genuine utility in compressing time-consuming but lower-judgment tasks. For example, AI can be used to read text from patents and publications, which does not require a training set. In those cases, AI can read huge amounts of text and consolidate conclusions and answers. Other lower-judgement tasks include literature monitoring, synthesis planning support, and administrative documentation. Turning a two-day process into a two-hour one, at scale, compounds meaningfully over the course of a drug program. These gains are less dramatic than the headlines, but they are achievable now. As we learn more about the utility of AI on lower-judgment tasks, we can assess strengths and weaknesses along the way and apply the technology to more complex tasks. 

A strategic framework for AI investment

For R&D leaders evaluating AI investments, a few principles are worth considering.

First, define the problem before selecting the tool. The question should never start with “how do we use AI?” It should start with “what is slowing us down, and is this the right tool to address it?” AI is one instrument in a broader scientific toolkit, not a strategy in itself.

Second, invest in data infrastructure as seriously as in models. The quality of your training data will determine the ceiling of what any model can do. Companies that treat data curation as a foundational capability will have a lasting advantage.

Third, measure performance objectively. Benchmark your models not just on familiar test sets, but on genuinely novel scenarios that reflect real drug discovery challenges. If a model only performs well on problems similar to its training data, you need to know that before you rely on it in a live program.

Finally, be skeptical of dramatic timeline compression claims. The setbacks that delay drug programs, including unexpected toxicity findings, novel safety signals, or target biology that doesn’t translate, occur sporadically in historical data and tend to manifest differently each time – too inconsistent for AI to learn from reliably. The human judgment, creativity, and domain expertise of experienced scientists remain irreplaceable at the hardest inflection points in drug development.

The real opportunity 

When it comes to AI, the algorithmic progress is genuine, the compute resources are unprecedented, and the industry’s willingness to invest in it creates real opportunity. AI can speed up drug discovery and decrease attrition rates in the clinic, but it is important to recognize that both are tall orders.

The companies that will benefit most are those that stay grounded, set realistic expectations, and keep experienced scientists at the center of decisions that require genuine creativity and judgment. That discipline is what turns AI from a promising technology into one that can deliver in many aspects of drug discovery with the ultimate goal of delivering for patients.

Photo: metamorworks, Getty Images

This post appears through the MedCity Influencers program. Anyone can publish their perspective on business and innovation in healthcare on MedCity News through MedCity Influencers. Click here to find out how.