Synthetic intelligence (AI) could also be remodeling drug discovery, however one of many business’s greatest issues stays unchanged: a lot of the biology used to develop new medicines nonetheless doesn’t absolutely mirror human illness.
Animal fashions usually fall quick in predicting scientific outcomes, and on the similar time, many organic datasets stay fragmented and closely based mostly on correlations. The problem is much more seen in liver illness, the place scientific trials are advanced and failure charges are excessive.
“We actually simply don’t have sufficient causal human biology data across the liver,” stated Quin Wills, chief government officer (CEO) of Ochre Bio. “The fashions to validate concepts round targets should not actually nice. The gold normal mouse fashions nonetheless don’t actually predict fibrosis regression.”
With AI now built-in in drug discovery, some biotech firms are beginning to argue that the bottleneck is not solely computational energy or algorithms, however the high quality of the organic knowledge itself. That is driving curiosity in large-scale human datasets constructed utilizing transcriptomics, purposeful genomics, and perturbation-based approaches designed to raised seize how human tissues behave in illness.
One instance is a current collaboration between Ochre Bio and RNA sequencing firm Lexogen, centered on producing a big human liver purposeful genomics dataset for AI-driven discovery constructed round human organic knowledge. The collaboration additionally illustrates some extent that’s usually underestimated in AI drug discovery: the sequencing companion isn’t a commodity supplier, however a determinant of whether or not advanced organic knowledge can turn into dependable, interpretable, and usable for machine studying.
Why AI drug discovery has a biology downside
A lot of the thrill round AI in drug discovery has centered on algorithms. From massive language fashions to generative chemistry platforms, advances in machine studying have created the impression that higher fashions will naturally result in higher medicines. However in biology, the mannequin is just as robust as the info it learns from, and that makes experimental design, pattern high quality, reproducibility, and organic context central to AI efficiency.
That is the place specialist transcriptomics capabilities turn into strategic. For AI-driven drug discovery, knowledge era should be deliberate from the start across the organic query, the downstream computational use, and the necessity to distinguish true organic sign from technical noise.
“An AI mannequin is 2 issues. It’s when a very good algorithm is matched with good knowledge, and also you want each of these for that to occur. Sadly, what’s taking place round AI hype in the intervening time is that lots of the main focus is on algorithms.”
Quin Wills, chief government officer (CEO) of Ochre Bio
Wills argues that biology presents a problem that differs basically from the environments by which massive language fashions emerged. “The big language fashions have finished extraordinarily effectively as a result of they’ve had a ton of knowledge to coach on. We don’t have that benefit in biology. We have to generate good knowledge.”
The issue isn’t merely one among amount. In response to Wills, many organic datasets stay largely correlational, making it obscure the underlying mechanisms driving illness. “AI is only a very intelligent approach, for probably the most half, to detect correlations that the human mind can’t see,” he stated. “It’s simply correlation, nothing else. It doesn’t tackle the large organic questions you must reply once you’re doing drug discovery.”
Drug discovery is about answering a sequence of causal questions: if a gene is modified, if a protein is inhibited, or if a pathway is activated and the organic penalties observe? Understanding these cause-and-effect relationships stays significantly tougher than figuring out statistical associations.
For Ochre Bio, the answer is to generate datasets based mostly on causality as an alternative of counting on present public assets. The corporate’s work combines human tissue fashions, genomics and perturbation experiments to reply organic questions that different datasets can’t.
The rise of transcriptomics in AI-driven drug discovery
If the problem for AI-driven biology is producing higher human knowledge, transcriptomics has turn into central to that effort.
Transcriptomics measures RNA expression and gives a snapshot of what cells are actively doing at a given second. That makes it helpful for finding out illness states, therapy responses, and the way completely different cell populations behave beneath altering organic circumstances. Over the past decade, advances comparable to single-cell sequencing, spatial transcriptomics, and multiomics approaches have expanded the quantity of organic element researchers can seize from human tissues.
“We are able to see the biology in rather more element than earlier than,” stated Stéphane Barges, CEO of Lexogen. “You’ll be able to mix differing types of knowledge to get a extra full image. These days, everyone is speaking about spatial biology and multiomics.”
On the similar time, RNA sequencing has turn into considerably extra accessible than it was a decade in the past, each when it comes to value and scalability. In response to Barges, that’s serving to transfer transcriptomics past massive tutorial facilities and into broader drug discovery workflows.
“Right this moment, not solely the big firms or tutorial labs which can be effectively funded can entry it, however smaller labs can even use it,” he stated. “That is actually good for innovation.”
That broader adoption has additionally modified how transcriptomics is being utilized in pharmaceutical analysis. Relatively than serving primarily as a downstream validation software, RNA sequencing is changing into a part of the invention course of itself, notably in areas comparable to purposeful genomics and perturbation biology the place researchers try to know how cells reply when genes or pathways are modified.
Nonetheless, each firms argue that producing helpful transcriptomic knowledge for AI functions requires greater than merely operating sequencing experiments at scale. Experimental design, reproducibility, pattern high quality, organic context, and workflow consistency all stay essential. Lexogen’s position is due to this fact not solely to provide RNA-Seq knowledge, however to assist be sure that the ensuing dataset is clear, comparable throughout samples, and match for computational interpretation.
“If you wish to use AI in a significant approach, you want knowledge that’s designed very rigorously for a selected query,” stated Barges.
“RNA knowledge is very wealthy. It offers you a a lot deeper view than conventional assessments. However this additionally implies that you need to generate this knowledge in a really managed and cautious approach. In any other case, it loses worth. The info should be clear and constant; the info is the gas, and the AI mannequin is the engine.”
Stéphane Barges, CEO of Lexogen
Constructing human-first datasets
For firms making an attempt to construct AI methods round biology, producing higher knowledge means shifting past static datasets and towards experiments that seize how human cells and tissues reply. This new want has seen the emergence of perturbation atlases, the place genes are systematically modified to measure the organic results.
For Ochre Bio, this grew to become the premise of its collaboration with Lexogen. The challenge centered on producing what the businesses describe as one of many world’s largest human liver purposeful genomics datasets, constructed round large-scale gene perturbation experiments in main human hepatocytes throughout a number of donors and illness states.
For Lexogen, the worth proposition was the power to industrialize a posh transcriptomics workflow with out shedding organic decision. Giant-scale perturbation datasets solely turn into helpful for AI if each pattern is processed with constant high quality, if technical variation is minimized, and if the info construction stays suitable with downstream modeling. That mixture of RNA experience, scalable operations, and high quality management is why a specialist companion might be important reasonably than elective.
In response to Wills, the corporate intentionally selected to prioritize human organic methods reasonably than relying totally on simplified mobile fashions or animal datasets.
“We tradition diseased liver tissue out right here in Asia. We now have these organ perfusion methods in New York the place we’re holding entire human livers alive,” he stated. “Along with Lexogen, we now have generated lots of causal knowledge the place we’ve gone and knocked out all of the genes in a number of human liver donors, illness states, RNA sequence to construct up this basic wiring diagram.”
The concentrate on main human tissues can be a part of a broader development throughout drug discovery, notably in areas the place animal fashions have struggled to foretell outcomes. In liver fibrosis, for instance, preclinical translation stays very difficult.
One other essential side of the challenge was donor variability. Relatively than treating illness biology as uniform, the dataset included a number of donors and illness circumstances to seize how organic responses differ throughout human populations. The significance of any such variability is now clearer for AI-driven biology as a result of fashions skilled on homogeneous datasets could fail to generalize throughout actual affected person populations.
For Wills, the goal was not merely to generate a big dataset however to create a organic basis that might help future AI predictions.
“For those who go forward and knock out each single gene in several donors and completely different illness states and RNA sequences, it offers you a wiring diagram, a basic gene regulatory community that you may go away and take into consideration. And you’re then capable of ask organic questions, comparable to which genes in hepatocytes may upregulate FGF21? This can be a crucial protein in the intervening time within the liver fibrosis area,” he defined.
“We’re obsessive about complexity. We don’t consider that you may appropriately mannequin very advanced processes like fibrosis in quite simple cell tradition fashions or inappropriate animal fashions.”
Quin Wills, chief government officer (CEO) of Ochre Bio
Can this really change drug discovery?
The tougher take a look at is whether or not human-first datasets can enhance translation, and that stays unproven at an business stage. AI has already modified early discovery, particularly speculation era or goal prioritization. However there’s nonetheless a spot between prediction and scientific success. Ochre Bio thinks that AI must be paired with experimental methods which can be nearer to human illness.
Wills described this as a “lab within the loop” strategy, whereas additionally acknowledging that the phrase has turn into overused within the discipline. For Ochre Bio, the thought is to check a manageable variety of genes experimentally in advanced human fashions, then use AI skilled on causal perturbation knowledge to foretell the habits of genes that haven’t but been examined.
If algorithms turn into extra accessible, proprietary organic knowledge and the power to generate it reproducibly could turn into a extra essential aggressive benefit. Which will additionally favor firms and partnerships that can tightly join experimental biology, RNA sequencing, high quality management, and computational evaluation reasonably than treating these steps as separate silos.
The regulatory setting can be shifting, cautiously, in a path that makes these approaches extra related. In April 2025, the U.S. Meals and Drug Administration introduced a plan to cut back and doubtlessly substitute some animal testing necessities with new methodologies, together with AI-based computational fashions, cell strains, and organoids.
Wills sees these items starting to return collectively in liver illness, though he stays cautious about how far the sphere has progressed. “The intent is there, however we’re firstly of the start,” he stated. He pointed to rising regulatory curiosity in human fashions, biomarkers, and extra versatile scientific trial designs as indicators that drug growth may turn into extra refined in the way it makes use of human knowledge.
In that sense, the Lexogen-Ochre Bio collaboration factors to a broader shift in drug discovery. Claims about refined AI fashions are not enough on their very own. The sector additionally wants companions that may ask the suitable organic questions, design experiments round them, and generate high-quality human knowledge at a scale and consistency that enables causal predictions to be examined. That is the place Lexogen’s contribution turns into central: enabling the form of AI-ready transcriptomic knowledge basis that drug discovery more and more requires.
For a deeper dive into these subjects, Ochre Bio’s work on liver illness and Lexogen’s perspective on transcriptomics and AI-ready organic knowledge, hearken to the Past Biotech podcast episode that includes Quin Wills, CEO of Ochre Bio, and Stéphane Barges, CEO of Lexogen.
