AI in drug discovery has shifted from an interesting experiment to a real strategic question for pharmaceutical companies and biotech teams. Rising development costs, long timelines, and stubborn failure rates in the clinic mean that any credible way to make better decisions earlier has outsized impact right now. At the same time, the volume of biomedical data across genomics, proteomics, imaging, and clinical records has exploded beyond what traditional informatics can digest, which is precisely the situation modern AI systems are built to handle. This urgency is further underscored by the recent funding surge, exemplified by Chai Discovery’s $400 million round, reflecting strong venture capital interest in AI drug discovery.
AI turns biomedical data overload into a strategic advantage in modern drug discovery
How we got here: from scoring functions to foundation models
AI in drug discovery did not start with today’s foundation models and generative chemistry. Early systems focused on narrow tasks such as quantitative structure activity relationships and docking score prediction, using hand-crafted features and relatively small datasets. These tools helped rank compounds but did not fundamentally change how targets were chosen or how chemists designed molecules.
The deep learning wave brought neural networks that could operate directly on molecular graphs and protein sequences, improving property prediction and binding affinity estimation by learning richer representations from data. Structural breakthroughs such as accurate protein structure prediction further expanded what could be modeled, turning previously undruggable biology into something that could at least be reasoned about computationally.
The current phase is defined by biological foundation models. These are large representation learning systems trained in self-supervised fashion on diverse chemical and biological corpora, then adapted to many downstream tasks rather than a single prediction objective. They include protein language models trained on massive sequence databases, small molecule models that ingest millions of structures, and multimodal architectures that fuse omics data with literature and clinical information. Recent surveys count well over two hundred such models published for drug discovery applications since 2022, an unusually fast growth curve for a domain as conservative as pharmaceutical R&D. This surge in foundation models for drug discovery illustrates how quickly general-purpose AI systems are being integrated into pharmaceutical R&D workflows.
Industrial and academic groups are building families of these models focused on different slices of the biomedical space, from antibody-antigen interactions to small molecule protein binding and single-cell transcriptomics. Platforms that offer ready-to-use foundation models for genomics, proteomics, and molecular design make it possible for smaller teams to stand on the shoulders of these large systems and fine-tune them with their own proprietary data instead of training from scratch.
What AI actually does across the discovery pipeline
At the earliest stages, AI helps tame the complexity of target selection. Multimodal foundation models can ingest genomic variants, expression profiles, pathway databases, and phenotypic data to highlight mechanisms that appear causally linked to disease rather than merely correlated. For oncology and immunology in particular, these models reduce some of the uncertainty around which nodes in a network are most promising to modulate and which patient subgroups are likely to benefit.
In virtual screening, machine learning predictors now sit beside or inside traditional docking workflows. Neural networks trained on known ligand-receptor interactions can rapidly estimate binding affinity, physicochemical properties, and early activity indicators for millions of candidate structures, narrowing the search space to a focused set of hits that are more consistent with project constraints. Some platforms use foundation model embeddings of molecules and proteins to prioritize compounds even before expensive three-dimensional docking is run, effectively filtering the library with learned biological intuition.
Once hits emerge, AI-supported lead optimization tackles the classic balancing act of potency, selectivity, and developability. Property prediction models estimate absorption, distribution, metabolism, excretion, and toxicity profiles, flagging liabilities such as hERG risk or metabolic instability and guiding chemists toward more robust structures. Generative models and multiparameter optimization frameworks help teams explore chemical transformations that may improve selectivity or reduce off-target engagement without blowing up synthetic complexity. In practice, this shortens iteration cycles and makes it easier to maintain consistency between medicinal chemistry decisions and the eventual target product profile.
Preclinical and translational stages are also being reshaped. Biological foundation models trained on large-scale omics and imaging datasets support the detection of subtle safety signals, mechanism-based biomarkers, and early indicators of efficacy that might otherwise be missed in noisy data. In oncology, for example, integrated models link cellular phenotypes to clinical outcomes and can suggest which combinations or dosing strategies are more likely to succeed when compounds move into first-in-human testing.
Foundation and multimodal models as the new substrate
The unifying thread across these use cases is representation learning. A foundation model compresses vast chemical and biological datasets into embeddings that capture useful structure and context, which can then be reused in property prediction, interaction modeling, and safety assessment tasks with relatively modest labeled datasets. This makes it possible to build robust models in areas where experimental data is scarce or expensive, by standing on the broad statistical knowledge embedded in the foundation layer.
Multimodal variants go further by learning jointly over different data types such as protein sequences, molecular graphs, three-dimensional structures, omics readouts, and text from the literature. These models do not just predict whether a molecule binds a given target. They can learn associations between variants, pathways, and phenotypes, and reason about off-target risks by integrating signals across modalities. When combined with tools that predict conformational changes in proteins upon ligand binding and map these changes onto cellular functions, they offer a more unified view of mechanism and risk than isolated task-specific models ever could.
There is also a strong strategic element. Because foundation models are expensive to train but relatively cheap to adapt, they encourage a layered approach where industry players share or license general-purpose models and keep the fine-tuned layers and proprietary data private. This shifts competition toward data quality, integration, and domain knowledge rather than raw compute alone, and it gives smaller biotechs a way to plug into cutting-edge AI without building everything themselves.
Generative chemistry moves from novelty to workflow
Generative models mark a conceptual turning point. Instead of ranking existing molecules, the system proposes entirely new structures that satisfy project constraints such as potency, selectivity, and drug-like properties. Early work relied on recurrent neural networks and autoencoders that learned to generate valid SMILES strings or graph structures by modeling the distribution of known molecules. Subsequent generations introduced generative adversarial networks, transformer-based architectures, and reinforcement learning loops to better control the properties of the sampled compounds.
Recent progress focuses on two broad families. Ligand-based generation learns from large corpora of drug-like molecules and samples new candidates, optionally steered toward specific property objectives. Structure-based generation uses diffusion and related models to grow molecules directly within a protein binding pocket or around a known binding site, allowing the design process to respect geometry and interaction patterns from the outset. This structural conditioning makes scaffold hopping and exploration of underrepresented regions of chemical space more systematic and less dependent on human intuition alone.
Industry and academic perspectives emphasize that successful generative drug design depends on more than clever architectures. Practical frameworks treat chemical synthesizability, realistic development timelines, and hard ADMET constraints as first-class objectives alongside binding affinity. They also incorporate human feedback from experienced drug hunters, ensuring that generative suggestions are evaluated against the messy realities of medicinal chemistry and project strategy rather than only against artificial reward functions. In other words, generative chemistry is gradually being absorbed into standard workflows, not left as a curiosity on the side.
Toward closed loop discovery platforms
As foundation, multimodal, and generative models mature, they are increasingly coupled with automated synthesis and high-throughput experimentation. Robotic platforms that can assemble, purify, and test compounds feed results back into AI models, enabling design-make-test cycles where each iteration is informed by the latest experimental data instead of static training sets. This closed-loop approach turns discovery into an adaptive process, with models proposing compounds, experiments evaluating them, and the models updating their understanding in response.
In oncology-focused discovery, integrated AI agents orchestrate multiple steps at once. Biological foundation models reduce uncertainty in target and pathway selection, generative methods create candidates that match those mechanisms, and autonomous platforms manage assay scheduling and readout analysis to keep the loop moving. The vision is a system that can explore vast regions of chemical and biological space while continually refining its beliefs, compressing timelines and reducing the number of dead ends pursued before reaching the clinic.
These platforms are still emerging and are far from fully autonomous. Most remain heavily supervised, with scientists validating designs, inspecting failure modes, and deciding when to trust or override model recommendations. Regulatory constraints, safety considerations, and the high cost of mistakes ensure that human judgment will stay central even as more of the mechanical work is delegated to AI-driven systems.
Opportunities, risks, and what this means for the industry
For large pharmaceutical companies, AI in drug discovery offers a way to improve pipeline quality and make better allocation decisions under uncertainty. Foundation models can help prioritize which targets and indications to pursue, while generative chemistry and closed-loop platforms promise shorter cycles and more focused experimentation. The upside is not just speed but the possibility of rescuing promising biology that would previously have been dismissed as too complex or too risky to explore.
Smaller biotechs gain leverage by building on shared foundation models and focusing their scarce resources on proprietary data and niche expertise. They can quickly prototype AI-powered programs in areas such as rare diseases, where data is sparse but mechanistic insight matters, by fine-tuning existing models and integrating them with carefully curated experimental datasets. This shifts some innovation toward the edges of the ecosystem instead of concentrating it only in the largest players.
There are real risks and limitations. Foundation models inherit biases and gaps from the data they are trained on, which can lead to overconfident predictions in regions of biology or chemistry that are underrepresented in those corpora. Generative systems can propose molecules that look attractive in silico but are impractical to synthesize or have hidden liabilities that were not captured in the training data. Methodological reviews stress the need for rigorous multiparameter evaluation, robust uncertainty estimation, and prospective validation rather than relying solely on retrospective benchmarks.
Regulatory science is still catching up. Agencies must decide how to evaluate AI-assisted decisions, what documentation is needed for models that influence candidate selection, and how to ensure transparency without exposing proprietary details. There is also a talent gap. Effective use of these tools requires teams that understand machine learning, chemistry, biology, and clinical development and can translate between those domains without falling for hype or dismissing genuine advances.
From a societal perspective, the hope is that AI-supported discovery will eventually widen access to effective therapies, especially for diseases that have historically been neglected due to poor commercial incentives or complex biology. The concern is that if only a few organizations control the most powerful models and the richest datasets, benefits could remain concentrated, reinforcing existing inequities in global health and pricing. How data is shared, how models are governed, and how results are validated will shape which of these futures becomes real.
The next chapter: embedding AI in everyday decision making
The central question is already shifting away from whether AI can help drug discovery. Across target selection, virtual screening, lead optimization, and translational analysis, there is enough evidence that well-engineered models add value when used thoughtfully and evaluated rigorously. The more pressing issue is how organizations redesign their decision processes, data infrastructure, and culture to embed these tools responsibly.
Teams that treat foundation and generative models as partners in reasoning rather than black box oracles are likely to make better use of them. That means investing in data quality, interpretability, and experimental feedback loops, and being explicit about where model predictions are trusted and where they remain exploratory. It also means accepting that high failure rates are intrinsic to innovation in drug discovery, and using AI to fail more intelligently, not to chase the illusion of certainty.
Over the next few years, expect a move from demonstrations of isolated successes to more systematic evaluations of portfolio impact. The most convincing stories will not be single molecules but pipelines where AI has consistently improved decision quality across multiple programs and modalities. As those case studies accumulate, AI in drug discovery will look less like a separate discipline and more like a standard part of how modern medicines are discovered and developed.
Conclusion
Artificial intelligence is moving from promise to practice in drug discovery, at a moment when the world is still reckoning with the cost, speed and resilience of its therapeutic pipeline. The excitement is real, but so are the constraints, and understanding both is essential for anyone trying to separate durable progress from temporary hype.
Viewed from a distance, AI now looks ready to turn drug discovery into a more data driven and hypothesis rich discipline, surfacing patterns that were previously buried in massive collections of molecular, cellular and clinical data. Yet the prospect of faster and more precise medicines still depends on careful validation, reliable datasets and strong governance. As automation spreads through laboratory workflows, AI is emerging as a powerful but partial instrument that extends human judgment rather than replacing it, pushing therapeutics toward more targeted and resilient innovation whose ultimate clinical impact is only beginning to appear in outcomes data.
Why AI drug discovery matters right now
Drug development has always been a high stakes, high attrition business, with many candidates failing in late stage testing due to toxicity, lack of efficacy or unforeseen interactions. That reality collides with rising expectations for precision therapies and the urgent need to respond rapidly to emerging threats, as the COVID 19 pandemic made painfully clear.
Over the past decade, the volume of biomedical data has exploded, from omics datasets and high throughput screening results to real world evidence from electronic health records and registries. AI provides tools to navigate this complexity, and activity has accelerated in recent years. An analysis of Scopus publication data for 2024 estimated that AI drug discovery accounted for about 1147 research papers, roughly 11 percent of projected AI clinical and discovery publications, with AI drug discovery output rising about 421 percent since 2019. In the same period, AI methods for protein structure prediction and clinical trials design also expanded rapidly, highlighting how AI is becoming embedded across the therapeutic lifecycle rather than treated as a niche experiment.
This matters for technology leaders and clinicians not because AI will magically generate perfect drugs, but because it changes the economics and tempo of the search process. When models can prioritize targets, suggest repurposing options and highlight safety signals earlier, decision makers can shift resources more quickly and test more hypotheses with fewer redundant experiments.
From early computational tools to modern AI
AI in drug discovery did not start with deep neural networks. The field has a long history of computational chemistry, molecular docking and quantitative structure activity relationship modeling, which used statistical techniques to relate chemical features to biological activity. Over time, researchers layered in more sophisticated machine learning methods, including support vector machines, random forests and k nearest neighbour models, to predict drug target interactions and classify compounds for further study.
Network medicine then added another important dimension. Instead of studying single targets in isolation, scientists built heterogeneous biological networks connecting drugs, protein targets and diseases, and used machine learning to infer new interactions. These approaches set the stage for the modern era, where deep learning, graph convolutional networks and representation learning can operate on large biomedical graphs and omics datasets to discover potential therapeutics.
A recent systematic review of AI in drug discovery illustrates how this evolution has played out in practice. Across published studies, roughly 40 point 9 percent used machine learning methods, about 20 point 7 percent relied on molecular modeling and simulation, and 10 point 3 percent employed deep learning, with AI being used most heavily in the earliest stages of discovery such as target identification and virtual screening. The same review found a gradual extension of AI into preclinical and early clinical development, but far less penetration into late phase trials and post approval monitoring, which remain more tightly regulated and data constrained.
What AI is actually doing in labs today
At a practical level, AI is already reshaping several core tasks in drug discovery and development.
Target identification and validation
Models trained on genomics, transcriptomics and proteomics data can highlight pathways and molecular targets associated with disease phenotypes, helping researchers focus on mechanisms with stronger causal support. In COVID 19, for example, BenevolentAI applied machine learning to disease maps and signaling circuits to identify members of the numb associated kinase family as potential drug targets, offering a rational basis for repurposing an existing therapy.
Virtual screening and candidate generation
Machine learning and deep learning methods are widely used to score large libraries of molecules against predicted binding sites, narrowing the list of candidates that need to be tested experimentally. This includes deep architectures such as graph convolutional networks that operate directly on molecular graphs, as well as models that predict how compounds will perturb gene expression profiles in human cells. In one example, the DeepCE framework combined graph convolutional networks with feedforward neural networks to predict differential gene expression triggered by new chemicals, and applied this to screen compounds for COVID 19 relevance.
Drug repurposing
COVID 19 became a real world stress test for AI driven repurposing strategies. Researchers used mechanistic models of signal transduction combined with multi task learning algorithms to infer causal links between known drug targets and signaling circuits involved in SARS CoV 2 infection, generating ranked lists of repurposable candidates for further evaluation. Network based deep learning approaches such as CoV KGE integrated open data across transcriptomics, proteomics and clinical trials to identify 41 repurposable drugs, including dexamethasone, indomethacin, niclosamide and toremifene, whose associations with COVID 19 were supported by molecular and clinical evidence.
Other teams built machine learning pipelines using Naive Bayes classification, achieving around 73 percent accuracy in predicting which approved drugs could be repurposed for COVID 19 and highlighting about ten candidates for further testing. Broader efforts used ensembles of support vector machines, random forests, artificial neural networks and deep learning to predict repurposed drug candidates across SARS CoV 2, SARS and MERS, demonstrating that multiple algorithmic families can contribute meaningfully to the search space.
Despite this burst of activity, a careful postmortem on AI driven COVID 19 repurposing has noted that none of these efforts have yet produced a new repurposed drug that progressed fully through clinical trials and into standard care. That does not mean the methods failed, but it underscores the gap between promising signal detection and the arduous path to proven therapies.
Safety, toxicity and pharmacokinetics
AI models are also being used to predict absorption, distribution, metabolism, excretion and toxicity profiles, helping to flag compounds that are likely to fail due to off target effects or poor pharmacokinetic behaviour. When integrated early, these predictors can reduce the number of candidates that progress into expensive animal studies or human trials only to be withdrawn later.
Decision support across clinical phases
Beyond discovery, AI tools contribute to trial design, patient stratification and endpoint selection, particularly in oncology where molecular subtypes and biomarkers are crucial for matching patients to treatments. The same data driven logic applies to adaptive trials and real world evidence analysis, though regulatory frameworks are still catching up to these possibilities.
Lessons from COVID 19 as a proving ground
COVID 19 provides one of the most instructive case studies for AI in drug discovery, because the urgency forced the community to push methods into production at unprecedented speed.
Several groups built medical knowledge graphs for COVID 19 that linked drugs, protein targets, pathways and disease phenotypes, then applied deep learning and graph algorithms to prioritize existing drugs for repurposing. DeepDTnet, for example, embedded heterogeneous biological networks using autoencoders and inferred new drug target associations with higher accuracy than previous approaches, helping identify interactions that might be clinically relevant for COVID 19.
Another line of work emphasised fit for purpose selection of AI strategies, arguing that different tasks such as target discovery, mechanism elucidation and candidate ranking require distinct modeling approaches and evaluation metrics. Reviews of these efforts highlight that AI can rapidly narrow the field of possibilities, but that outcomes depend heavily on data quality, mechanistic grounding and the ability to validate predictions in vitro, in vivo and ultimately in patients.
In parallel, docking combined with machine learning was used to target specific viral proteins such as the main protease 3CL, with decision tree regression and quantitative structure activity relationship modeling identifying drugs with high predicted binding affinities. These approaches showed that even relatively classical machine learning models can add value when applied thoughtfully to well curated datasets.
The mixed results from COVID 19 repurposing are an important check on expectations. AI succeeded in generating ranked lists of plausible candidates and suggested non obvious mechanisms, but clinical translation remained limited, largely because biology is complex and randomized trials are still the final arbiter of value. That experience is shaping how organisations now integrate AI into pipelines for other diseases, emphasising rigorous validation and the avoidance of overconfident claims.
Implications for technology, business and society
For technology teams, AI in drug discovery is a test of whether data intensive methods can integrate smoothly with experimental science rather than sitting apart from it. The most successful projects tend to combine computational predictions with iterative wet lab validation, allowing models to be retrained as new evidence arrives and to focus on realistic questions such as which chemical series deserves more investment.
Pharmaceutical and biotech companies see AI as a way to de risk portfolios and compress timelines. By using AI to screen vast chemical spaces, predict properties and prioritise candidates, they aim to reduce the number of late stage failures, which carry enormous financial and reputational costs. The shift is also changing partnerships and funding models, with AI focused startups collaborating closely with larger pharma firms and contract research organisations to embed their platforms into existing workflows.
For society, the stakes are broader. If AI can help identify repurposing opportunities or accelerate the design of new therapies, patients may gain earlier access to effective treatments, particularly in areas like oncology and rare diseases where options have historically been limited. At the same time, introducing complex predictive systems into medicine raises questions about transparency, accountability and bias. Data used to train models may reflect historical inequities, and if that is not corrected, AI assisted decisions could perpetuate uneven access or skewed risk assessments.
Regulators are beginning to grapple with these issues, examining how AI is used in trial design, evidence generation and post market surveillance. Guidance increasingly emphasises explainability, robust performance across diverse populations and clear documentation of how models are validated. The path forward will likely involve iterative regulation that adapts as real world evidence accumulates, rather than a single static rulebook.
Risks, limitations and how to build trust
Trust in AI driven drug discovery depends on more than technical accuracy. It requires a culture of scientific humility and rigorous testing.
Several common limitations are already apparent. Models are often trained on narrow datasets that do not capture the full variability of human biology, leading to brittleness when applied to new populations or indications. Many studies are retrospective, evaluating performance on held out data rather than prospectively guiding decisions, which can inflate expectations about real world impact. There is also a temptation to treat complex architectures as black boxes, even though mechanistic understanding remains essential for safety and regulatory acceptance.
Responsible teams counter these risks by emphasising interpretability, using tools such as feature attribution and network visualisation to connect model outputs back to biological rationales. They design experiments that challenge models with edge cases and adverse scenarios, and they publish both successes and failures to allow the field to learn collectively. Importantly, they treat AI as an augmentation of expert judgment, not a replacement, keeping clinicians, pharmacologists and statisticians in the loop.
From a governance perspective, robust data stewardship, reproducible pipelines and clear documentation of assumptions are as important as algorithmic sophistication. Without these, even a technically impressive model can become a liability.
The road ahead
Looking forward, AI is likely to become a routine part of drug discovery rather than a headline novelty. Integration with advances in protein structure prediction and large scale generative modeling will allow teams to explore chemical and biological space more systematically, but the fundamental constraints of biology and clinical testing will still apply.
The most transformative impact may come not from single breakthrough models, but from the cumulative effect of many modest improvements across target selection, compound optimisation, safety prediction and trial design. Each step that becomes more evidence driven and transparent can shorten the path from idea to therapy, especially when AI systems are continuously updated with real world outcome data.
AI has already shown that it can find promising signals hidden in massive scientific datasets and guide more focused experimentation. The next chapter will be written by how well the community aligns technical ingenuity with rigorous validation, ethical oversight and patient centred priorities. Whether this promise turns into sustained clinical benefit will depend on embracing AI as a powerful yet partial tool, one that earns trust through consistent performance and openness rather than sweeping claims reddit
1 comment
Comments are closed.