AI tools are becoming increasingly common in early drug discovery, allowing scientists to analyse data and navigate large volumes of research. However, according to Dr Raminderpal Singh, turning that potential into consistent scientific workflows is far from straightforward.

Artificial intelligence (AI) is becoming an increasingly prominent part of drug discovery, particularly in early-stage research where computational tools promise faster data analysis and more informed decision making. Yet translating these advances into reliable scientific workflows remains a significant challenge.

At SLAS Boston 2026, Drug Target Review spoke with Dr Raminderpal Singh, Global Head of AI and GenAI Practice at 20/15 Visioneers, about where AI is delivering value in early discovery and why expectations for the technology often outpace practical implementation.

The origins of AI in modern drug discovery

According to Singh, the current conversation around AI in drug discovery began more than a decade ago with major advances in biological data generation and computing infrastructure.

“Ten to fifteen years ago machine learning became very popular because a lot of data became readily available, quite cheaply,” Singh explained.

The expansion of genomics and other omics technologies produced vast datasets, while cloud computing made it far more affordable to store and process them.

“Genomics became a huge mega topic because of the availability of sequencing data,” he said, adding that cloud computing was also a significant contributor and that it helped drive AI interest.

Generative AI changes the conversation

Interest in AI surged again around 2022 with the arrival of generative AI (GenAI) tools. Large language models (LLMs), such as ChatGPT, demonstrated the ability to analyse and generate natural language, enabling researchers to extract knowledge from scientific literature.

“Large language models can do natural language. They can do knowledge extraction,” Singh said.

shutterstock_2415395573

Source: Krot_Studio / Shutterstock

Large language models (LLMs) are machine learning models trained on large text datasets to predict and generate language. They can perform tasks including question answering, text summarisation and information extraction.

This capability allows researchers to work not only with numerical datasets but also with written information, including research papers, diagrams and experimental descriptions. These developments have renewed enthusiasm across the industry. However, Singh believes the excitement is again moving faster than the practical realities of implementation.

Rising expectations for generative AI

According to Singh, generative AI has triggered another surge of interest across technology and pharmaceutical communities.

“In 2025 the word ‘agent’ comes along and everybody’s building agents,” he said. “Everybody’s building an agent for this and an agent for that.”

Part of the excitement stems from the experience of interacting with conversational AI systems such as ChatGPT, which can provide immediate answers to questions.

“We’re now in a TikTok world,” Singh said. “Because of ChatGPT we ask a question and we expect a good answer straight away.”

However, scientific research requires far more structured systems than simply asking questions.

Because of ChatGPT we ask a question and we expect a good answer straight away.

“These tools are processing engines just like a machine learning model. They work as part of workflows. Somebody must design those workflows.”

According to Singh, without careful system design the potential of AI technologies can easily be overstated. In drug discovery these systems need to operate within structured workflows with curated data, validation steps and clear guardrails. Without that framework even powerful models may produce outputs that appear convincing but are difficult to reproduce or translate into real experimental decisions.

The hidden cost of generative AI

Another challenge emerging with generative AI is the cost of using large language models at scale.

“When I’m using large language models I’m paying for access,” Singh said. “Every time the model is reasoning through something for me or doing a task, I’m paying tokens.”

This token-based pricing model means costs can increase quickly for researchers who rely heavily on the tools.

“You hear people saying, ‘This thing’s really great but I’m spending two or three thousand dollars a month on it,’” Singh explained. At the scale of large pharmaceutical companies employing thousands of scientists, this can create new budgeting challenges.

“If you’ve got thousands of scientists running around, each of them can’t be spending two or three thousand dollars a month on tokens,” Singh said.

Workflows remain the central challenge

Despite rapid advances in AI technology, Singh believes the biggest barrier to adoption lies in how organisations design their research workflows.

“The barriers come back to discipline in creating workflows,” he said.

Building effective systems requires careful planning around data acquisition, data engineering, processing and interpretation.

“Building any complex workflow takes time,” Singh explained.

Large language models are also probabilistic systems, meaning they do not always produce identical responses to the same question.

“You ask the same question twice and you don’t get the same words,” Singh said.

For that reason, organisations must invest time in designing systems that guide how the models are used.

Looking beyond large language models

While large language models are becoming increasingly integrated into research workflows, Singh believes the next major step in AI development may come from systems known as world models.

“The next generation beyond large language models is world models,” he said.

World models aim to simulate complex systems by integrating different computational approaches to represent biological processes.

The next generation beyond large language models is world models.

“World models are the use of large language models and other types of models to create massive simulations of systems,” Singh explained.

Such simulations could eventually allow researchers to test hypotheses computationally before conducting laboratory experiments.

“When world models become real, early drug discovery will become a completely different experience,” Singh said.

Advice for scientists navigating AI

For researchers who feel overwhelmed by the rapid pace of AI development, Singh recommends a simple starting point: begin using the tools that already exist.

“The first step for any scientist is to become very engaged using large language model apps,” he advised.

These tools can help with tasks such as literature analysis, knowledge extraction and report generation.

“For around twenty dollars a month you can do a huge amount. It’s like having a colleague in the room, an intelligent colleague.”

Different models offer different strengths, but Singh encourages scientists to experiment and find the tools that work best for them.

For around twenty dollars a month you can do a huge amount. It’s like having a colleague in the room, an intelligent colleague.

“For me, it’s Claude,” he said. “But I use Perplexity when I’m searching the web a lot because it’s very good at scraping and collating information.”

Ultimately Singh believes the best way for researchers to understand AI’s potential is simply to start using it.

“Pick a tool, find a tool you like and use it,” he concluded.

As AI technologies continue to evolve, many organisations are exploring how best to integrate them into existing scientific processes. While tools such as large language models are already helping researchers navigate complex datasets and scientific literature, their long-term impact on drug discovery will depend on how effectively they are incorporated into research workflows and experimental decision making.

However, Singh argued that the technology being adopted during that period was not always what people understood as AI.

“It wasn’t really an AI wave. Nobody even knows what AI means. It was more like we can do a lot more sophisticated maths with a lot of data,” he said.

These developments created strong expectations that AI would rapidly transform pharmaceutical research. Many in the industry believed that data-driven approaches would significantly accelerate target identification and drug development.

Ten to fifteen years ago machine learning became very popular because a lot of data became readily available, quite cheaply.

“Drug discovery is going to be this big winner,” Singh recalled people saying at the time.

Yet several years later, according to Singh, the reality appeared less impressive.

“You dial forward five or seven years into the late 2010s and people have got their heads in their hands going ‘this AI thing is rubbish’,” he said.

Much of that frustration arose as the early optimism around AI met the practical challenges of drug discovery. While machine learning could analyse large datasets and generate predictions, translating those insights into validated targets or clinically viable drugs proved far slower and more complex than many had anticipated.

Read The Article

News – Curated by Amanda Scott, Alias Group Creative
Follow her on Bluesky

GHCE Concept

From the preface of the book Global Health Care Equivalency in the Age of Nanotechnology, Nanomedicine and Artificial Intelligence, Edited by Frank Boehm: Since the publication of my first book (Nanomedical Device and Systems [...]

Healthcare Headlines: Challenges and Advances in 2026

Health-related updates reveal financial adjustments by Universal Health Services due to Medicaid reimbursement uncertainties, significant pollution-linked health concerns from French-British oil firm Perenco in Congo, drug trial setbacks, potential restructuring at major medical firms, [...]