As AI Reshapes Drug Discovery the Lab Remains the Bottleneck

By Allison Proffitt 

June 9, 2026 | At a plenary session at the Bio-IT World Conference & Expo last month, Jeremy Jenkins, a computational drug discovery leader at Novartis, delivered a frank and wide-ranging assessment of where artificial intelligence is genuinely accelerating pharmaceutical R&D — and where it keeps running into the same old wall: biology. 

“We solve one thing and the bottleneck shifts elsewhere,” Jenkins told the audience, “and almost always it shifts back to the lab.” 

A Pipeline in Need of a Shortcut 

Jenkins opened by framing the challenge. Drug discovery, he noted, runs roughly half research and half development, with an average timeline of 12 to 13 years from target identification to approved therapy. His talk zeroed in on three early-stage areas he sees as ripe for AI-driven acceleration: target identification, small molecule design, and AI-enabled assays. 

The ambition, he argued, is twofold: do the same work faster (efficiency), but also reach hypotheses that human researchers simply wouldn’t have arrived at on their own (creativity). “I think we are getting more creative and we’re coming up with target hypotheses that we wouldn’t have gotten otherwise,” he said, “molecules we wouldn’t have gotten from screening existing molecules.” 

Finding Targets with Foundation Models and LLMs 

On the target identification front, Jenkins described how Novartis is combining several AI approaches. Single-cell RNA sequencing foundation models — transformers pre-trained on vast human cell atlas datasets — are being fine-tuned on proprietary disease-versus-healthy tissue data to identify which genes behave differently in sick cells. These models can then simulate what happens when a specific gene is knocked out, a technique Jenkins called “in silico perturbation.” 

“We’ve been doing this on many, many diseases,” he said, noting that the models predict directional changes in gene expression reasonably well, even if they sometimes miss on magnitude. He was candid about the limitations: the models are largely trained on healthy cells, which haven’t “seen” what it looks like when genes are suppressed by disease. And as with many computational methods, the output is a gene list — useful, but only a starting point. 

To move from gene lists to target hypotheses more efficiently, Novartis has leaned on large language model-powered tools, including a platform called Causaly. Jenkins described querying the system for novel drug targets against the malaria parasite and receiving, in one minute and 26 seconds, a 10-page report citing 32 papers and proposing 29 target classes — a task he estimated would have taken a human researcher one to two weeks. 

“That’s the power of putting gene lists you may be getting from other AI methods into an LLM to quickly annotate and provide more evidence towards a target hypothesis,” he said. 

Designing Molecules for the Pocket 

Jenkins then turned to small molecule design, where he sees generative AI making real but imperfect inroads. The core problem: chemical space is effectively infinite, and a drug candidate must simultaneously hit a target with high affinity, bind in the right orientation, avoid toxicity, and survive in the human body long enough to work. 

He described Novartis’s adoption of a 3D generative chemistry method called Pocket2Mol, an open-source academic algorithm that designs molecules conditioned directly on the three-dimensional shape of a protein binding pocket — growing the molecule atom by atom to fit the target’s geometry. The team industrialized it into an internal pipeline and applied it to a challenging oncology target, the protein-protein interaction between MYC and WDR5. 

Compared to screening 1.1 billion compounds computationally, the generative approach proposed far fewer candidates but achieved a 13-fold better hit rate. Several resulting molecules showed low-micromolar activity and, crucially, appeared to bind at residues that conventional screening had missed. 

“That’s the power of de novo design: to get what you want up front,” Jenkins said. 

Still, designing molecules is only half the problem. Ranking them remains hard. Jenkins surveyed the landscape of co-folding methods (including AlphaFold 3) and drug-target interaction models that try to predict binding affinity from protein and compound features. The best AI affinity models achieve Pearson correlations around 0.5–0.6, he noted — better than classical docking scores, but well short of physics-based free energy perturbation methods, which reach 0.8. The challenge, in his view, is building AI approaches that combine the speed of machine learning with the accuracy of physics. 

Teaching Microscopes to See What Isn’t There 

Jenkins described AI-enabled imaging assays as “transforming the way we do imaging.” 

High-content screening, in which cells are stained with fluorescent dyes and antibodies and then imaged to measure biological effects, is a workhorse of pharma R&D. But staining kills cells, fixing them in time and preventing any further observation. Jenkins described an approach his team calls in silico labeling: training neural networks on paired brightfield (unstained) and fluorescent (stained) microscopy images, then using those models to predict what the fluorescent image would look like without ever staining the cell. 

In a myelination assay—measuring the protective sheath around neurons, relevant to diseases like multiple sclerosis—the predicted images actually looked cleaner than the real stained images to the human eye. Pearson correlation between predicted and actual myelin area reached 0.83. The ROI was tangible: reagent use fell by 80%, imaging speed increased tenfold, and researchers could take repeated measurements over time without destroying their samples. 

Jenkins also described a more speculative extension of the approach, which his team calls pDIFF (Profile-conditioned stable DIFFusion model for high-content imaging). Rather than predicting a fluorescent image from a brightfield image, pDIFF attempts to generate a predicted cell image from a compound’s chemical structure alone — before any experiment is run. The key was representing compounds not by their 2D chemical fingerprint, but by an in silico bio-activity profile built from over 10,000 historical dose-response assays run at Novartis over two decades across cell types, protein types, and disease areas.  

“It’s probably a billion dollars worth of data,” Jenkins said. “We call that profile QSAR. It’s this biological fingerprint. I think the power of it is that it really doesn’t treat all atoms and bonds the same in a molecule. It actually emphasizes pharmacophoric type of motifs that are biased for pharmacologic activity. I think it has more power than a typical chemical descriptor.”  

Jenkins asked the pDIFF model to, “Create me a high content image of U2-OS cells treated with a Chk1 inhibitor?” The results surprised the team. Models trained on standard chemical fingerprints produced what he called “total nonsense.” Models trained on the biological activity profile generated images that closely resembled real cell-painting experiments. Applied to the problem of hit expansion — finding more active compounds similar to a known hit — pDIFF achieved roughly 50% overlap with experimentally validated actives, compared to around 10–12% for random chance and around 40% for the biological fingerprint alone. 

Where AI Stands — and Where It Doesn’t 

In his closing remarks, Jenkins resisted both hype and dismissal. AI in drug discovery, he argued, is genuinely accessing new hypothesis spaces — novel targets and novel scaffolds that conventional approaches would have missed. But efficiency gains remain elusive, partly because each time an AI method solves one bottleneck, the constraint simply shifts back to the laboratory: to experimental validation, to synthesis, to cell-based testing. 

“I don’t think it’s all going to come from AI,” he said. “I think we’re really going to need to hit this AI mission as a combination of lab automation and AI working together.”