#104- Unpacking AI in drug discovery
Unpacking the hype, challenges, and pockets of opportunity
In my last post, I talked about the exciting premise (& promise?) of AI in drug discovery—faster timelines, lower costs, and higher success rates. It sounds like the perfect solution to fix an industry full of inefficiencies and high failure rates. But is AI really living up to the hype? Beyond headlines about big partnerships and billion-dollar valuations, the reality is a bit more complicated.
In this post, I will try to share where AI is delivering results, the challenges it still faces, and the opportunities that can arise in the future. Whether you are a founder or an investor, it is important to understand these aspects as we move towards a future where AI and pharma together to reshape healthcare.
Current Landscape of AI in Pharma
AI’s role in pharma has grown a lot in recent years, with startups and big companies using it to make drug discovery better and faster.
AI startups are more and more partnering with or getting acquired by big pharma companies. This shows a trend where they work together rather than compete with each other.
Challenges of AI in drug discovery
1. Data bottlenecks- AI has access to large datasets like genomics and molecular interactions, but the quality of data is a big challenge. Data gaps, errors, and biases can lead to wrong predictions. Plus, combining different datasets into one framework is still very hard. Sometimes, traditional models like homology still perform better than advanced ones like AlphaFold.
2. Predictive limits- AI simulations are improving but still cannot handle all the complexities of human biology. For example, unexpected interactions, like off-target effects or metabolic issues, are often missed by AI models.
3. Novelty- AI relies too much on the data it is trained with, which means it often focuses on known compounds instead of discovering new, innovative drugs. This bias in training data means unexplored targets or rare conditions are ignored. Predicting new molecular classes or mechanisms is still a tough challenge.
4. Regulation and validation- Regulatory agencies still require AI findings to be checked with traditional clinical and preclinical methods, which adds extra cost and time. Also, there are no clear rules for evaluating AI-driven drug discoveries, making it harder for the industry to move forward.
My understanding of AI in drug discovery
AI still needs more creative breakthroughs to meet the hype it is getting. Too much hype can even slow down progress because people become too cautious when real breakthroughs are possible. The current AI systems need more reliability and better reasoning algorithms, not just relying on fixed weights for predictions.
The issue is not AI itself, but the over-reliance on current models. Maybe in the future, new models will solve the problem better. Comparing today’s AI in drug discovery to computers in the 90s makes sense—it’s not that AI is early or buggy, but the current models are not ready to produce truly groundbreaking discoveries yet.
In fact, AI has actually been helping drug discovery for over 20 years. It’s made biologics therapy and other areas much faster. Tools like AlphaFold and custom AI models help analyze data, making research more efficient. While it hasn’t delivered everything people expected, it has improved predictions, especially when working with complex data sets.
For example, AlphaFold can predict protein folding without crystal structures, but this success is partly due to better data from experiments, not just better AI. Today, AI allows faster protein engineering and better analysis, showing its strength lies in managing and processing large data faster than ever.
Pockets of opportunity
The merging of AI, biotech, and personalized medicine offers huge growth opportunities, using new tech to solve big problems in drug discovery and health care. Also, quantum computing, even though still new, fits in our bigger interest in game-changing technologies that can change industries.
For early stage investors, while drug discovery in itself is a very exciting and profitable proposition, it is also a space which requires an enormous amount of capital. Hence it’s important for me as an early stage investor to find business models that give key tools, platforms, or infrastructure to support the whole ecosystem of AI-based drug discovery.
Below are some models that excite me, considering AI’s increasing role in this space:
Data platforms/infrastructure providers
Companies offering basic data systems and platforms for AI to work better. This can include genomic data storage, clinical trial data hubs, or special molecular databases made for training AI. Companies can source clinical data or mix experimental data (genomic, proteomic, clinical) to make datasets for AI training.
Why it excites me? AI needs lots of clean, good-quality data to work well. Companies that organize and prepare such data for AI use will be very important. These platforms can also sell access via APIs or SaaS, working like "AI data middlemen."
Tools for preclinical and clinical phases
Companies creating AI-powered tools for different steps in drug development like molecular screening, in silico tests, or patient selection for clinical trials. A platform using AI to make better trial designs, improve patient selection, and predict responses to treatments is of key interest.
Why it excites me? Drug discovery has many slow, costly, manual tasks. AI tools to speed up processes like compound testing or trial planning can save time and money.
Drug repurposing platforms
Companies using AI to find new purposes for old drugs, saving time and money in development by using safe drugs for new diseases. AI lab or SaaS tool that suggests new therapeutic areas for known drugs, licensing this data to pharma companies is a perfect example. Read more about this here.
Why it excites me? Repurposing known drugs can skip some clinical trial steps. AI can find new uses from data patterns people might miss.
Drug design tools
Companies which have tools for molecular design, predictive modeling, or finding new disease targets. These can be platforms for compound simulation or predicting drug interactions, helping small biotech firms work faster and cheaper.
Why it excites me? AI tools to help scientists design and improve drug candidates will grow to become more valuable than some small biotechs due to mortality risk in biotech and breadth of opportunities in a platform.
Biomarker Discovery
Tools to find biomarkers for early diagnosis or treatment response. These can be startups using AI to identify and validate biomarkers for unmet needs like cancer or rare or orphan diseases.
Why it excites me? Biomarkers are key for precision medicine, and AI can make finding them faster and easier.
Predictive analytics/decision support
These can be predictive AI tools for R&D teams, helping with drug success forecasting, dose planning, or toxicity checks. AI platforms can potentially predict the success rate of drug candidates based on past trial data.
Why it excites me? Such tools lower trial risks and make drug pipelines efficient, saving time and resources.
Regulatory and compliance tools
AI solutions and platforms to simplify regulatory paperwork and reporting for drug approvals which can automate FDA report prep, speeding up approval steps.
Why it excites me? Regulatory steps take time and are expensive. Tools that automate submissions and compliance checks save cost and reduce errors.
While AI in drug discovery has enormous potential, I believe the real value lies in businesses enabling this tech through tools, data systems, and platforms. These “shovel-sellers” can power AI’s impact in biotech, offering lower-risk but high-return chances especially for early stage investors.
Reality is AI still faces limits like needing better models and higher quality data. Quantum computing may disrupt the field in a few years, but now, focus on AI’s realistic use in precision medicine and biotech. This sector suits our strategy to invest in real-world deeptech, balancing innovation with practical growth paths. By backing companies with strong science and reliable AI applications, we aim to support healthcare’s transformation while generating solid returns.
Updates
I saw mixed responses on my ‘shovels’ approach i.e. selling tools for AI in drug discovery being better than making drugs itself. It strongly believe that companies focus on building systems and platforms are making rapid changes in the industry.
Many AI companies failed because they only solved small problems (relative to opportunity, not that these are insignificant) like virtual screening or docking. But new age companies are building better technology to solve the inherent biology problem from a wider, larger perspective. This simply means they create tools that help others find drugs faster.
Lab automation and biobanks are very important for drug discovery today. Companies like Insilico build wet labs that run on AI, and this saves lot of time and money. This shows that making infrastructure for drug research is more valuable than just finding one drug (if you remember it takes more than a billion dollars and a decade to do this).
Finally, two of the biggest problems in drug discovery are predicting toxicity or finding the right patients for trials. And these can not be solved by simple AI models. We need platforms that work with ton of data and use advanced AI to make good predictions. These tools can use by many researchers to do their work better.
One counterpoint is that some companies are doing both tools and drug discovery at same time. If they succeed in both, then they will capture more value than just selling tools. But even these companies depend on having good infrastructure to do their work, so tools are still very important.
The best opportunities in AI drug discovery are in making platforms and systems that everybody can use. These businesses may give more stable growth and less risk compare to companies that only focus on making drugs. Investors who support these companies can fully underwrite this enormous change happening in this field.
Fin on this topic for now.