How We Choose Drug Target is Changing - Fast
Developing a new drug begins with a critical question: what should we target to treat a disease? For decades, scientists answered this by educated guesswork – pinpointing one suspect protein or gene based on limited clues, then painstakingly testing one drug at a time. This traditional approach has been slow, expensive, and risky. In fact, roughly 90% of drug candidates fail in clinical trials, often because the initial target or model was not predictive of human disease. Lab dishes and young lab animals cannot fully mimic the complex biology of an older human body, so many “promising” targets fall apart when tested in real patients. Today, however, target discovery is undergoing a rapid transformation. New technologies – from CRISPR genome editing and multi-omics data mining to AI-driven analysis and phenotypic screening – are taking the “hunt in the dark” out of target selection. As I write to learn, I want to know how these advances are making target discovery more systematic, high-throughput, and predictive than ever, and what that means for the future of drug development. I’ll write a little about Gordian too - because I find them super cool :D
What we did (traditionally)
For most of modern drug discovery, choosing a target was a hypothesis-driven process based on fragmentary evidence. Researchers would start by identifying a molecule (often a protein) suspected to play a key role in a disease – say an overactive enzyme or signaling receptor. This suspicion might arise from human genetics (e.g. a gene variant linked to the disease), animal models, or even serendipitous clues (like a natural product that affects a pathway). Armed with a candidate target, scientists then screened libraries of compounds to find any that could hit it. A “hit” compound would be optimized in the lab and validated in preclinical tests, moving from cell culture to animal studies. Only after years of work would a drug targeting that molecule enter human trials.
This linear process yielded many successful drugs, but it has major drawbacks. First, it’s extremely slow and low-throughput. Testing one target hypothesis at a time means many years before finding out if you guessed right. Second, it’s often misleading, because disease biology is far more complex than a single target. A drug can look effective in an artificial system yet fail in humans because of off-target effects or compensatory pathways not seen in simpler models. Indeed, most drug candidates that enter trials fail to become approved drugs, frequently due to lack of efficacy in humans. Finally, traditional models lack realistic context. Human diseases – especially chronic and age-related diseases – involve dozens of cell types, dynamic interactions, and long-term processes like inflammation or fibrosis. A cell in a plastic dish or a tumor in a young mouse cannot recapitulate that complexity. This “gap” in translation has been recognized as a key factor in R&D’s high failure rate. As one analysis noted, improving preclinical models’ predictive validity is crucial to improving drug development success.
In short, the old paradigm of target discovery was hypothesis-driven and reductionist. Researchers picked a target based on partial information and hope, then validated it through a slow pipeline of in vitro and animal tests. It’s akin to looking for your keys under a lamppost because that’s where the light is – while knowing they might be lost somewhere in the dark. Today, new approaches are lighting up the entire landscape, allowing us to find drug targets in a more unbiased and comprehensive way.
Living animals as discovery engines
One of the most radical shifts in target discovery is the move to high-throughput in vivo screening – essentially using living disease models as discovery engines rather than as one-at-a-time validators. A striking example of this is the approach taken by Gordian Biotechnology, a company that literally tests hundreds of gene therapies simultaneously in a single living animal. This approach turns the traditional workflow on its head: instead of optimizing one therapy and taking it through multiple animal studies, Gordian delivers a library of candidate gene therapies into one complex, diseased animal to see which ones work in situ.
How is this possible? Gordian’s platform consists of three key components:
Patient Avatars™: Large, naturally diseased animals that closely model human conditions. For example, they use older horses with osteoarthritis and primates with metabolic syndrome as stand-ins for human patients. These animals develop disease through natural aging and genetics, so their physiology (immune responses, tissue complexity, comorbidities) is much more human-like than an artificially induced disease in a young lab mouse.
Mosaic Screening™: A pooled in vivo screening method where hundreds of distinct gene therapy vectors, each carrying a unique DNA barcode, are injected into a single animal. Each cell in the diseased tissue takes up at most one vector, creating a “mosaic” of different therapeutic experiments within one living organism. Because the barcodes mark which therapy each cell got, the outcomes of hundreds of interventions can be read out from that one animal. In effect, one Avatar hosts hundreds of mini-experiments in parallel – compressing months or years of sequential testing into one study.
Pythia™: An AI-powered analytics engine that reads the results of the mosaic screen at single-cell resolution. After the animal is treated, researchers perform single-cell RNA sequencing on its tissues, sequencing each cell along with the therapy barcode it received. Pythia then sifts through this massive dataset to see which gene therapy (which target) produced beneficial changes in the cells. By linking each barcode to molecular and cellular outcomes, the AI can identify which targets had the desired effect (such as prompting cartilage repair in an arthritic joint). This approach cuts through the noise of complex in vivo data to pinpoint promising targets with far greater confidence than traditional guesswork.
The results from this bold strategy are extremely encouraging. In an osteoarthritis study, Gordian screened hundreds of gene therapy candidates in horses suffering natural joint degeneration; the top hits identified by the screen went on to show ~80% accuracy in predicting cartilage regeneration when tested in follow-up lab models. In another experiment, a pool of 50 gene therapies was delivered to a mouse model of fatty liver disease (MASH); remarkably, the single in vivo screen correctly recapitulated 13 out of 16 known positive targets – essentially predicting the outcomes of many separate trials in one go. These proof-of-concept results suggest that high-throughput in vivo screening can dramatically improve the hit rate of finding true therapeutic targets.
Why is this approach so powerful? It comes down to context, scale, and data. By testing therapies inside a living, diseased, aged organism, you automatically account for factors that are impossible to capture in a petri dish – the immune system, the tissue architecture, the systemic environment, and the effects of aging. Promising targets emerge in the proper context of disease. By doing hundreds of tests at once, you massively accelerate the cycle of discovery – compressing what used to be years of work (one target, one animal at a time) into a single experiment. And by applying AI to rich single-cell data, you can detect subtle cellular signals and cause-effect relationships that human eyes or bulk measurements would miss. In short, this strategy starts with biology rather than simplifying it away. It embraces complexity but tames it with smart technology.
Gordian is not alone in this philosophy. A broader trend is emerging to use more physiologically relevant systems earlier in drug discovery. For example, Ochre Bio – a biotechnology startup focused on liver disease – conducts its target and drug screens on actual human organs. They maintain donated human livers (ones not suitable for transplant) on specialized perfusion machines, essentially a “liver ICU” that keeps the organ alive outside the body for days. In this living human tissue, Ochre Bio can test dozens of genetic or RNA-based therapies to see which best improve liver function or reverse disease markers. According to CEO Jack O’Meara, with whom I had some of my greatest chats, using real human organs has “dramatically shortened” their development timeline to days instead of months, because they can observe drug effects in a human physiological context immediately. Indeed, Ochre Bio’s innovative platform helped it secure a $1 billion+ collaboration deal with a major pharmaceutical company, underscoring big pharma’s belief in more predictive preclinical models.
The take-home message is that in vivo and ex vivo screening platforms are turning complex biology from a liability into an asset. Rather than working around the complexity of human disease, these approaches dive straight into it – testing potential interventions in systems that behave like the human body. This is a paradigm shift: living systems are becoming discovery platforms. We are likely to see more use of aged animal models, organoids and organ-on-chip systems, and even patient-derived tissues as “high-fidelity” testing grounds for target discovery. The early successes of these strategies suggest that the best way to find what works in humans is, whenever possible, to start with real biology upfront.
Looking into our genes
In parallel with advances in where we search for targets (i.e. in more realistic models), there has been a revolution in how we search the target space: functional genomics. At the heart of this revolution is the genome-editing tool CRISPR-Cas9, which has made it feasible to systematically interrogate the function of essentially every gene in the genome. So-called CRISPR screens allow researchers to knock out, activate, or otherwise modulate thousands of genes in a single experiment, then read out which genetic perturbations produce a desirable effect (such as killing cancer cells or increasing a cell’s insulin sensitivity). This approach provides an unbiased way to discover which genes really matter for a given disease phenotype, as opposed to the old approach of guessing based on prior knowledge.
CRISPR screening technology has rapidly become a cornerstone of target discovery in both academia and industry. With libraries of guide RNAs, scientists can perform genome-wide loss-of-function screens – e.g., disabling each gene one by one in a pool of cells – and then identify which gene knockouts cause a meaningful change in the cells’ behavior. For example, a CRISPR screen in cancer cells might reveal which genes are essential for the cancer’s survival, highlighting vulnerabilities that a drug could target. Indeed, CRISPR-based screens have identified numerous promising targets across cancer, cardiovascular disease, neurodegeneration, and other areas. Because these screens test every gene impartially, they can uncover novel disease drivers that would have been missed by hypothesis-driven research (including genes that were previously “anonymous” with little literature but turn out to be linchpins in a disease pathway).
Several technical advances are enhancing the power of CRISPR functional genomics. One major innovation is combining CRISPR screens with single-cell sequencing readouts. Instead of simply measuring whether cells live or die after a gene knockout (as early screens did), researchers can now read the full transcriptional profile of each perturbed cell. Techniques like Perturb-seq allow a massively parallel experiment where each cell is barcoded with the guide RNA that hit it, and after the screen, you sequence all the cells to see how each gene knockout shifted the cell’s gene expression program. This yields rich information on how a target gene influences cellular state, not just whether it’s essential. It also means you can screen for complex phenotypes – for instance, find genes whose disruption makes immune cells adopt a more anti-inflammatory profile, or genes that push a fibroblast from a pro-fibrotic state to a regenerative state.
Another advance is conducting CRISPR screens in more physiologically relevant models. Initially, most screens were done in cancer cell lines or immortalized cells grown in a dish. Now, labs are using 3D organoids (miniature organ-like cell clusters), primary cells from patients, and even in vivo CRISPR screens in live animals. For example, researchers can introduce a pooled CRISPR library into a mouse and look for guide RNAs enriched in metastases (to find genes that promote cancer spread) or in surviving cells after drug treatment (to find genes conferring drug resistance). Combining CRISPR perturbations with organoid models of diseases is especially powerful – it marries unbiased gene discovery with a disease model that has architecture and cell-type diversity closer to real tissue. All of this means CRISPR-based functional genomics is yielding target insights with far greater biological validity.
Being data-driven
Another major change in how we choose drug targets is the rise of data-driven, system-level discovery. Diseases are being studied at an unprecedented scale and depth thanks to high-throughput “omics” technologies – genomics, transcriptomics, proteomics, metabolomics, epigenomics, and beyond – often applied at the single-cell level. These technologies are producing a flood of data that, with the help of advanced computational tools (including machine learning and AI), can be mined to uncover new insights about disease mechanisms and potential intervention points.
In the past, target discovery often started from a single observation (e.g., “Protein X is higher in diseased tissue, let’s target X”). Now, researchers can take a far more holistic view: for example, profiling all the gene expression changes, protein modifications, and metabolic shifts that occur in a disease versus healthy state, and using that rich dataset to identify key nodes in the network. This integrative approach is called multi-omics, and it’s quickly becoming the norm. A recent review noted that drug-target identification has “shifted towards integrated multi-omics techniques, gradually replacing traditional single-omics techniques”. The reason is simple – complex diseases rarely have a single molecular cause. By analyzing multiple layers of biology together, we can find patterns and drivers that a one-dimensional view would miss.
Consider an example: in heart failure research, a multi-omics study might combine patient genome sequencing (to find genetic risk variants), transcriptomics of heart tissue (to see which genes are up/down-regulated in failing hearts), proteomics of the blood (to find circulating factors that correlate with disease severity), and single-cell RNA-seq (to see which cell types are changing and how). Such a study could reveal, for instance, that a particular cell type (say, fibroblasts) in the heart is shifting into a pathological state characterized by a certain regulatory network – and within that network, a protein emerges as a central hub controlling many disease-related genes. That protein could be a promising target to modulate the fibroblasts’ behavior. This is not a hypothetical scenario; approaches like this are already re-classifying diseases based on molecular profiles. In cancer, comprehensive atlases have identified new subtypes of tumors defined by specific cell state signatures and vulnerabilities, pointing to targets (enzymes, transcription factors, cytokine receptors, etc.) that drive those states.
Artificial intelligence and machine learning are essential allies in this data-driven approach. The datasets are often too large and complex for humans to interpret unaided. Machine learning models can find subtle correlations and construct predictive models: for example, identifying a combination of gene expression changes that causes a particular disease phenotype rather than is merely associated with it. Techniques like causal inference, network analysis, and knowledge graphs can help distinguish true driver genes from passengers. Causal AI models, in particular, aim to move beyond correlation – they integrate data from perturbation experiments (like CRISPR screens or drug perturbations) with observational data to infer what factors lie upstream of others. One can imagine an AI model ingesting multi-omic data from hundreds of patients and outputting: “The data are most consistent with the hypothesis that Pathway Y is a master regulator of this disease; within Y, enzyme Z appears to be a choke point that, if inhibited, would reverse the disease signature.” This is increasingly how targets are being nominated.
We’ve already seen practical outcomes of data-driven target discovery. A classic example is the identification of PCSK9 as a cholesterol-regulating target – human genome data (rare mutations) flagged PCSK9 as a key player in LDL levels, leading to the development of PCSK9 inhibitor drugs. Now apply that concept on a genome-wide scale with machine learning: companies like BenevolentAI and Insilico Medicine use AI systems trained on biomedical data to predict novel targets for diseases ranging from fibrosis to neurodegeneration. In 2020, for instance, BenevolentAI’s platform identified a drug targeting the IL-6 receptor as a potential treatment for COVID-19 by mining connections in literature and omics data – a hypothesis that was later clinically validated. This kind of rapid target repurposing would have been unlikely via traditional methods.
It’s noteworthy that single-cell atlases of human tissues – such as the Human Cell Atlas project – are providing a new map for target discovery. By cataloguing every cell type and state in healthy and diseased tissues, scientists can pinpoint exactly which cells to target and when. For example, single-cell analysis of lung fibrosis might reveal a specific subpopulation of fibroblasts that drives scarring, characterized by co-expression of certain pathways; a drug could then be designed to selectively hit that subpopulation’s vulnerabilities. Single-cell data often highlight cellular programs (like the senescence program, or an inflammatory activation program in macrophages) that span multiple genes. This shifts our thinking away from single genes to network states – and suggests that intervening in a network (perhaps by targeting a master regulator or a combination therapy) could be more effective. We’ll talk more about targeting cell states in the next section.
Finally, the integration of machine learning into all these analyses has greatly improved our ability to generate hypotheses from big data. Machine learning models can absorb vast amounts of molecular data and output ranked lists of candidate targets. For example, a deep learning model might be trained on known disease-gene relationships and then used to predict new gene-disease links from multi-omic data; or a model might classify patients into subgroups based on molecular profiles and identify what makes one subgroup’s biology distinct (potentially revealing a targetable abnormality). Modern drug discovery increasingly employs such AI tools not only for target identification but also for target prioritization – scoring targets by druggability, safety profiles (e.g. expression in essential tissues), and synergy with existing therapies.
The bottom line is that we have unprecedented data and computational power to guide target selection. We are no longer limited to what we can observe with a microscope or a single assay; we can let the data speak, often revealing unexpected drivers of disease. Multi-omics ensures we aren’t missing the forest for one tree, and AI helps us find meaningful patterns in the forest. This data-driven paradigm complements the experimental advances (like CRISPR and in vivo screening) by providing a broad, unbiased view of disease biology from which new targets emerge. It transforms target discovery from an expert-driven art into a more objective, evidence-driven science.
Let cells lead the way
While molecular data and targeted approaches have boomed, there’s also a renaissance in a more old-school strategy: phenotypic drug screening. Phenotypic screening means testing compounds (or genetic perturbations) on disease models and simply asking: does it make the phenotype better? – without initially worrying about the specific target. It’s a throwback to how many drugs were originally discovered (by observing functional effects in cells or animals), but updated with modern tools. Intriguingly, phenotypic screening has often yielded first-in-class drugs by finding molecules that produce a desired change in a disease model, even when the underlying target was not known upfront.
In the early 2000s, drug discovery shifted heavily toward target-based screening (thanks to genomics and combinatorial chemistry). Yet, analysis of success stories revealed that many breakthrough drugs – especially for complex diseases – came from phenotypic approaches. Today, with advanced imaging and AI, phenotypic screening is more powerful than ever and is helping identify new targets and drugs in tandem.
One driver of this resurgence is high-content screening. We can now measure phenotypes in far more detail than just cell death or a single biomarker. Automated microscopes can capture images of cells under thousands of different treatment conditions, and these images contain a wealth of information about cell morphology, organelle state, protein localization, etc. Machine learning algorithms (especially deep neural networks for image analysis) can detect subtle phenotypic changes that humans might miss – effectively quantifying the “state” of a cell in a high-dimensional way. Companies like Recursion, Insitro, and others have built enormous image-based screening platforms where millions of microscopic images of cells treated with various compounds or genetic perturbations are analyzed by AI. The AI clusters compounds that cause similar changes, hints at their mechanisms, and can flag compounds that reverse a disease phenotype even if we don’t initially know or care how they work.
For example, Recursion’s platform grows human cells engineered to exhibit disease-relevant phenotypes (such as cells with a certain toxic protein aggregation). They then apply hundreds of thousands of chemical compounds or CRISPR-based gene knockouts to these cells and use computer vision to see which interventions restore the cells to a healthy state. In one case, they identified an existing drug that unexpectedly normalized a cellular defect in a model of a rare genetic disease – a finding that suggested a new use for the drug (target repurposing) and also helped identify a previously unknown molecular pathway involved in that disease. Phenotypic hits like this can then be followed up to “deconvolute” the target – i.e. figure out which protein the drug was modulating to achieve the effect. But importantly, the initial discovery didn’t require knowing the target; it only required a measurable phenotype and a library of diverse perturbations.
Modern phenotypic screening is not limited to cells in a dish. It extends to organoid models, zebrafish larvae, and other whole-organism or 3D systems where complex phenotypes (like developmental processes or behavior) can be observed. High-throughput zebrafish screens have been used, for instance, to find compounds that suppress epileptic seizures or improve cardiac function, by literally adding chemicals to tanks of tiny transparent zebrafish and using imaging or electrophysiology readouts. These systems let the biology “tell us” what works, which can lead to drugs acting on novel targets or multi-target mechanisms.
The integration of multi-omics with phenotypic screening further boosts its power. After finding a drug that phenotypically “works”, researchers can apply transcriptomics or proteomics to treated vs untreated cells to identify which pathways were altered – providing clues to the drug’s mechanism and potential targets. This approach was unheard of decades ago but is now quite feasible. For example, if a mysterious compound makes sick cells look healthy, you can perform RNA-seq on those cells and see that it down-regulated a specific set of stress-response genes – pointing you to the pathway, and eventually the target, that the compound affects.
It’s worth noting that phenotypic screening is essentially what Gordian’s in vivo approach and Ochre’s human liver experiments are doing, just at a larger scale. They test many interventions and ask which ones improve the overall phenotype (cartilage regeneration in a horse joint, or liver function in a human organ). Thus, the line between phenotypic and target-based discovery is blurring: new technologies allow us to do unbiased phenotypic discovery and identify the targets responsible once we have a hit.
A real advantage of phenotypic screening is that it can reveal therapeutic strategies that target complex processes or multiple targets at once. Some diseases may best be treated not by a single magic-bullet target, but by nudging a cellular program. A phenotypic readout (say, neurite outgrowth in a neuronal model for neurodegeneration) might be restored by compounds that act on different targets but converge on promoting regeneration. This approach is agnostic to whether one or several proteins are involved – if the phenotype improves, it’s a win. In an era when polypharmacology (drugs hitting multiple targets) is recognized as sometimes beneficial (e.g. in psychiatric meds or cancer combos), phenotypic assays naturally accommodate that.
Of course, phenotypic screening also relies on having a relevant phenotype. The quality of the disease model is crucial – which is why advanced models like 3D co-culture organoids, human iPSC-derived cells, and in vivo disease models are being adapted to screening. We want phenotypes that translate to clinical outcomes, not just any change in a cell. This is where the field has learned from past failures (simple cell assays that didn’t predict anything). Today’s phenotypic screens aim to be disease-relevant and data-rich, often incorporating human cells or tissues and readouts that correlate with disease biology.
In summary, phenotypic screening has made a comeback as a powerful complement to target-based design. It’s a way of listening to the cells: you throw a wide net of perturbations and let the biological system reveal what can correct the disease features. Coupled with modern imaging and AI analysis, this approach can unveil both new drugs and the novel targets or pathways they act on. It rebalances the discovery process by ensuring that efficacy in a biological context is front-and-center, rather than an afterthought. Many in the industry now advocate a hybrid strategy: use phenotypic screens to find efficacious interventions, then decode the targets and mechanisms with omics and AI – effectively harnessing the strengths of both unbiased and hypothesis-driven methods.
As always, we need some lessons:)
As we’ve seen, multiple paradigm shifts are happening in parallel. What general principles can we learn from these successful strategies? A few key themes stand out:
Start with Biology, Not Just Hypotheses: The new approaches embrace unbiased exploration of disease biology. Whether it’s screening in live animals or letting an AI sift multi-omic data, they allow the system itself to highlight what’s important, rather than relying solely on our preconceived theories. This reduces the chance of chasing a target that sounds good but isn’t truly central to the disease.
Scale Up and Speed Up: High-throughput methods – from pooled CRISPR libraries to automated phenomic imaging – massively increase the scale of experiments. This “scale-smart” mentality means we can test hundreds or thousands of possibilities in parallel, dramatically accelerating discovery and avoiding false leads that would cost years if done sequentially. The faster we can cycle through ideas, the faster we learn and converge on the right targets.
Combine Technologies for Synergy: The biggest breakthroughs often come from integrating multiple cutting-edge tools. For instance, in vivo screening marries gene therapy, single-cell omics, and AI. Phenotypic screening platforms combine robotics, imaging, and deep learning. Multi-omic data is intertwined with causal inference algorithms. Combining tech is crucial – each tool compensates for others’ limitations and provides a more complete picture.
Don’t Oversimplify – Embrace Complexity (When Needed): A recurring lesson is that using more human-relevant and complex models early can pay off. Simplified systems are easier and will remain valuable for certain studies, but we’ve learned that sometimes you must tackle complexity head-on (e.g., using aged animals or human organoids) to avoid misleading results. The trick is to choose the right model for the question – and not to shy away from complexity if it’s critical for translational accuracy.
Be Ambitious and Bold: Lastly, these new approaches all required a certain boldness – testing hundreds of therapies at once in an animal, or perturbing every gene in the genome, or trusting an AI to guide biology. The lesson is that to significantly improve success rates, incremental tweaks aren’t enough; we need ambitious leaps that rethink the process. The good news is that the tools to do this are now in hand, and early successes (like Gordian’s animal studies or AI-designed proteins) validate the risk.
These principles are increasingly influencing how biotech and pharma R&D teams structure their discovery programs. We can expect future projects to be designed around them – for example, incorporating humanized models and CRISPR screens from day one, or using AI to nominate targets which are then vetted in high-throughput phenotypic assays. The ultimate goal is to pick better targets (ones that truly impact disease in patients) and to do so faster and more efficiently than before.
With those insights in mind, let’s look ahead at where these trends might take drug discovery in the coming 5–10 years.
Some educated prediction for drug discoveries in the next 5 - 10 years
Data-Driven Target Discovery Becomes the Norm
We will see a continued shift from hypothesis-driven to data-driven target identification. Rather than starting with a pet theory about a single gene, researchers will start with large-scale data – patient multi-omics, population genetics, and unbiased screens – to let the key signals emerge. Technologies like comprehensive single-cell atlases of disease tissue, genome-wide CRISPR perturbation datasets, and deep learning models trained on biomedical data will guide target selection. In practical terms, this means new targets will increasingly be found by mining patterns across thousands of genes and cells, identifying network hubs or cell-state regulators that drive disease, instead of focusing on one pathway at a time. This system-level approach acknowledges that diseases (especially complex ones like Alzheimer’s, diabetes, autoimmune disorders) are not caused by one faulty gene but by dysregulated networks. By harnessing big data, scientists can pinpoint leverage points in those networks. Importantly, these data-driven findings will be filtered through experimental validation – for example, a causal inference algorithm might highlight 10 candidate targets, which are then quickly tested in a CRISPR screen or organoid model to see which actually modulate the disease phenotype. The net effect is a more robust and comprehensive pipeline for target discovery, with fewer blind spots. We can also expect better disease subtyping from multi-omics (leading to more personalized targets). Overall, embracing data-driven discovery will likely increase the success rate of picking targets that translate into effective drugs, because those targets will come pre-vetted by human data and unbiased functional evidence.
Complex In Vivo Models Become Gold-Standard Early Tests
The era of relying on a single cancer cell line or young mouse model to decide on targets is waning. In the next decade, high-throughput in vivo platforms and other complex models will become a gold standard for early target validation. We will routinely see approaches like Gordian’s mosaic screening in aged animals applied to various diseases – imagine pooled in vivo CRISPR screens in mice to find drivers of fibrosis, or barcoded viral vectors in primates to assess gene therapy targets for neurodegeneration. Likewise, advanced human-derived models (such as organoids that contain multiple cell types, or cultured human tissues with microfluidic perfusion) will be used to evaluate target perturbation effects in a realistic context before a project commits to drug development. This trend is driven by the clear evidence that simplistic models often fail to predict human outcomes. As one 2012 study famously highlighted, many costly late-stage failures stem from the translational gap between preclinical models and human biology. To close that gap, the industry is moving testing that used to happen in Phase 1 or 2 (in patients) earlier into the preclinical phase, via humanized models and in vivo surrogates. In 5–10 years, it may be routine for a pharma company to say: “We have 5 candidate fibrosis targets from multi-omics; let’s run a mini in vivo trial in parallel on all 5 by using a pooled CRISPR approach in a mouse model that truly mimics human fibrosis. The data will tell us which target has the best therapeutic index in a realistic setting.” Approaches like this will significantly de-risk projects. High-throughput in vivo data combined with AI analysis will identify winners and weed out losers far earlier, saving time and money. In short, predictive models that capture human complexity will increasingly replace overly reductionist models, leading to targets that have already proven their mettle in lifelike systems.
Therapies Aim at Cellular Programs and States, Not Just Single Proteins
Drug mechanisms will gradually shift from knocking out single proteins to modulating entire cellular programs. As our understanding of diseases deepens, we recognize that conditions like cancer, fibrosis, or neurodegeneration aren’t solely caused by one aberrant protein, but by cells entering pathological states (e.g. a pro-fibrotic state, a senescent state, an inflammatory macrophage activation state). Future therapies will target these cell states or programs. This could mean combinations of drugs hitting multiple nodes of a network, gene therapies reprogramming cell fates, or “smart” biologics that reset a cell’s behavior. Already, startups like Unity Biotechnology are developing drugs to selectively eliminate or alter senescent cells (cells that have entered an aged, pro-inflammatory state implicated in many diseases of aging). These senolytic or senomorphic agents don’t target one enzyme – they target a whole class of cells defined by a state (senescence). Similarly, the burgeoning field of cell reprogramming (pioneered by companies/orgs/centers like Altos Labs and Calico) is essentially aiming to reverse the program of aging or disease in cells, making them behave as if they were younger or healthier. We can expect future drugs that, for example, push immune cells from a tolerogenic state to an active tumor-fighting state (in cancer therapy), or convert scar-forming cells in the heart into a regeneration-permissive state after a heart attack. Achieving this may involve hitting multiple targets at once or upstream master regulators (like epigenetic modulators). The Human Cell Atlas and similar efforts are providing the blueprints by cataloging the molecular features of every cell state. Using that knowledge, researchers can design interventions not just to inhibit Protein X, but to drive a whole cellular phenotype from “disease” toward “health.” In 5–10 years, we will likely see the first FDA-approved drugs explicitly indicated to modify a biological hallmark of aging or a dysfunctional cell program (for instance, a therapy that treats multiple age-related conditions by flushing senescent cells). These will open up a new category of treatment that is broader in mechanism yet precision-guided to the right cells.
Generative AI for Drug and Target Design
Artificial intelligence will move from assisting discovery to directly designing drugs and even suggesting therapeutic biological mechanisms. The success of tools like DeepMind’s AlphaFold in predicting protein structures has been a game-changer for understanding targets. Building on that, companies (including DeepMind’s own spin-off, Isomorphic Labs) are using AI to design novel molecules that can modulate those targets. In the coming years, generative AI models – analogous to language models like GPT, but trained on chemical and genomic data – will be routinely used to invent new drug candidates (small molecules, protein therapeutics, gene therapies) with desired properties. For example, given a protein target structure, an AI can generate novel drug-like compounds that fit into its pocket; given a certain biological function needed, AI can generate a protein that performs that function (as demonstrated by Salesforce’s ProGen model, which created new enzymes that were functional in the lab). We have already seen AI-designed proteins (such as new antimicrobial enzymes) and even early AI-designed small molecules entering clinical trials. Over 5–10 years, these capabilities will expand.
Beyond designing the molecules, AI will help design entire therapeutic strategies. We might have AI propose, for instance, a multi-gene cell therapy: “to make T-cells better at attacking cancer X, modify gene A, B, and C in this way,” based on analyzing massive datasets of cell circuitry. AI could also optimize vectors, dosages, and even clinical trial designs (through digital twin simulations). In short, AI and machine learning will permeate all stages of drug R&D – not just finding targets but creating the solutions. One exciting possibility is AI-driven synthetic biology where we design circuits or microbes as therapies (e.g. a synthetic bacterial strain that senses an inflammatory environment and secretes an anti-inflammatory drug – AI can help design the genetic circuit for that). Foundation models trained on biomedical knowledge might even identify new therapeutic modalities we haven’t thought of, by analogy and creativity drawn from data. The lab of the future could involve scientists working closely with AI systems, where the AI can generate hypotheses (“target this allosteric site to tweak the protein’s activity”) and even draft experimental plans to test them. While human expertise will remain crucial, much of the heavy lifting in design and analysis will be offloaded to powerful computational models. This should drastically cut down the time from target idea to a viable drug candidate and open up novel solution spaces (e.g., protein designs humans would never conceive on their own).
Aging as a Treatable Condition
Aging itself – long considered an inevitable, untreatable process – is on track to become an accepted therapeutic target in its own right. In the next decade, we anticipate the first interventions that are explicitly approved to prevent or delay multiple age-related diseases at once. The rationale here is that aging drives diseases like arthritis, Alzheimer’s, heart disease, and more; if you can slow biological aging or remove its hallmarks, you could broadly improve healthspan. This concept has gained enough traction that the FDA approved a trial design (the TAME trial – Targeting Aging with Metformin) to test an existing drug’s ability to delay the onset of age-related chronic diseases. Results from such studies in the coming years could provide proof that aging rate can be modified in humans. Meanwhile, numerous biotech companies are working on aging-focused therapies: senolytics that clear senescent cells (e.g., Unity’s candidates), NAD+ boosters that improve cellular metabolism, telomerase activators, young-blood factor mimetics, and even partial cellular reprogramming techniques to rejuvenate tissues. We are likely to see some of these approaches reach clinical validation. For example, if a senolytic drug shows it can meaningfully improve an aging biomarker or functional aspect in older adults, regulators may begin to consider aging as an indication (perhaps measured by something like “time to onset of any of several age-related diseases”). Startups like Loyal are even starting in pet dogs to gather data on lifespan extension in mammals. In 5–10 years, it’s plausible that we’ll have the first geroprotective drug – a medication given to at-risk older individuals to postpone frailty and disease, much as statins are given to prevent cardiovascular events. The mindset shift here is enormous: treating aging itself rather than each disease separately. Even before official indications catch up, people may be using such therapies off-label. From a target discovery perspective, this means new targets are being identified in pathways of aging (senescence, proteostasis, mitochondrial function, etc.), and they often have broad-spectrum effects on multiple diseases. Aging biology is yielding drug targets like TERT (telomerase), mTOR, AMPK, BCL-2 family (for senolytics) and others that drug hunters are actively pursuing. If successful, the impact on public health would be revolutionary – extending healthy lifespan rather than just treating disease in siloed fashion.
Faster, Modular, and Distributed Drug Discovery
Finally, the entire process of drug discovery is poised to become much faster and more modular, resembling an engineering project more than an artisanal science experiment. Innovations in automation, cloud computing, and collaboration are driving this. Cloud laboratories – such as Emerald Cloud Lab and Strateos – allow researchers to conduct wet-lab experiments remotely with robots executing the protocols. This means an experiment that might take a team weeks to do by hand can be run overnight by a robot, and multiple experiments can be run in parallel 24/7. As more labs come online, scientists will be able to prototype and test drug ideas as easily as running software code (a concept investors call “Bio as APIs” or programmatic biology). In addition, digital twins and in silico modeling will let researchers simulate aspects of drug behavior before ever going to the lab. We already see AI models predicting drug-target binding, pharmacokinetic properties, and even likely side effects by structure – these will only improve.
Another aspect is the idea of modular discovery platforms. Companies are increasingly built around reusable modules: for example, a platform that can generate antibodies to any target quickly, or a gene therapy vector that can be retargeted to different organs by swapping a tissue-specific promoter. Once you build a core capability (like an AAV delivery method to the liver, or a CAR-T cell framework), you can reuse it for multiple programs, making each new drug faster to develop than the last. This modularity is enhanced by collaboration – we see big pharma partnering with tech firms or AI startups to plug in capabilities. The result is an ecosystem where a small company might design the molecule, a cloud lab tests it, another specialized company does the toxicity profiling, etc., all coordinated in a more distributed R&D pipeline. This breaks the traditional monolithic in-house model and allows multiple innovations to be combined. In the next 5–10 years, such arrangements will shorten the timeline from idea to IND (Investigational New Drug) dramatically – possibly from ~4 years currently to 1–2 years in best cases. For instance, imagine a researcher uses AI to identify a target and design a compound in silico within weeks, orders it synthesized via an automated chemistry platform, screens it on disease organoids via a robotic lab, and iterates a few times – all before any significant manual effort. What used to be sequential steps done by different departments over years could be done in a largely parallel, largely automated fashion in a fraction of the time. Investors like a16z have dubbed this the era of “Bio 2.0” or “TechBio”, emphasizing practices like version control, rapid prototyping, and scaling that mirror software development. The upshot for target discovery is that the feedback loops will tighten – if a target isn’t panning out, you’ll know and pivot quickly, and if it shows promise, you can immediately plug it into a machine-oiled pipeline to make a drug. Drug discovery is on its way to being faster, more predictable, and more reproducible, powered by these modular and automated approaches.
These future trends all feed into a central outcome: better therapies reaching patients faster. By transforming how we choose targets – making it more data-driven, context-aware, and tech-enabled – we increase the chances that when we aim our therapeutic “arrow,” it hits the bullseye of clinical benefit. The next decade of advances, from AI-designed molecules to aging-delaying drugs, promises to not only speed up the lab work, but to redefine what is treatable. Diseases long deemed intractable may succumb to combination strategies and novel targets uncovered by unbiased screens. And we’ll likely tackle entirely new frontiers, like the biology of aging, that were outside the scope of medicine until recently.
In the end, the goal is that future medicines will be more effective and rationally designed because they’re built on a foundation of deeper biological insight. The process of discovering those medicines is becoming more of an engineering challenge and less of a shot in the dark. It’s an exciting time in biomedical research – one where centuries-old problems may finally yield to the fast-evolving toolkit of modern science. By embracing these changes in target discovery, we move closer to a world where finding the right target is no longer the bottleneck to curing disease, but rather a swift and surmountable step on the path to new cures.
See ya next time:)
