Target identification, not the drug itself, is the key challenge in first-in-class therapeutics
The industry has a chemistry execution problem only because it first has a biology selection problem.
BLUF
The pharmaceutical industry loses most of its bets before a molecule is ever synthesized. Roughly 85% of drugs entering clinical trials fail, and the two dominant causes both trace upstream to how the target was chosen. About 45% of failures stem from safety and toxicity concerns and about 40% from poor efficacy or target selection, which means the overwhelming majority of clinical attrition reflects targets that were never adequately validated in humans. Preclinical animal models, still the primary validation tool, carry an estimated 92.6% false discovery rate, while drug mechanisms with human genetic support succeed in the clinic at 2.6 times the rate of those without. Human genetic evidence is decisive here because it predicts both efficacy and safety; targets whose genes are highly constrained or broadly expressed across tissues fail disproportionately for safety reasons. Of roughly 4,500 druggable genes, approved drugs address only 716, leaving the highest-value first-in-class opportunities in precisely the territory where causal evidence is thinnest. Insilico Medicine has demonstrated that AI-driven target identification can reach mid-stage clinical validation, Genomics plc is systematizing the genetic-evidence advantage directly, and BenevolentAI's clinical failures illustrate what happens when correlative target selection lacks causal grounding. The companies that solve target identification with human-genetics rigor, rather than generating more molecules against unvalidated targets, will capture the most durable returns in drug discovery.
Connected to Thesis 1: Lab Automation & Closed-Loop Discovery. Better target ID reduces the synthesis burden on automated labs. Better robotics generates the biological data needed to validate targets. These theses are mutually reinforcing.
Background & historical context
From empirical pharmacology to causal target biology
Drugs were discovered by observation. Willow bark reduced fever. Opium relieved pain. The concept of a molecular target did not exist. Efficacy was the only signal; mechanism was unknown.
Ehrlich articulated the idea that drugs must bind selectively to specific cellular targets; the first conceptual framework for target-based drug discovery. His compound 606 (arsphenamine) for syphilis was the first designed molecular therapy.
Black, Stephenson, and others formalized receptor theory. Drugs could now be rationally designed to agonize or antagonize specific receptor classes. Beta-blockers and H2 antagonists followed directly from this framework.
Köhler and Milstein produced the first monoclonal antibody, opening the door to highly specific biological targeting. Nobel Prize 1984. Biologics would eventually become the dominant modality for truly novel targets.
The HGP changed the target identification frame entirely: from 'find a protein that binds your drug' to 'find the gene causally implicated in this disease, then drug it.' This was the conceptual shift that made genomics-driven target ID possible.
The complete sequence revealed ~20,000 protein-coding genes, of which approximately 4,500 are considered druggable. All drugs approved to date target only 716 distinct proteins, a vast white space of unexplored biology.
Genome-wide association studies began producing statistically rigorous links between genetic variants and disease. The key insight, formalized over the next decade: targets with human genetic support are substantially more likely to succeed in clinical trials than those derived from animal models alone.
Minikel et al. (Nature 2024, building on Nelson 2015) rigorously quantified: drug mechanisms with human genetic support have a 2.6× higher probability of clinical success. This single empirical finding is the foundation for the current wave of genetics-driven target ID investment.
Insilico's PandaOmics platform identified TNIK as a novel target for IPF. Rentosertib, the resulting inhibitor, was nominated as a preclinical candidate 18 months after project initiation at a cost of $150K (ex wet lab). It entered Phase IIa and showed dose-dependent FVC improvement, the first AI-identified, AI-designed drug to reach mid-stage clinical validation.
Co-founded by Nobel laureate David Baker, Xaira raised over $1 billion to rebuild drug discovery end-to-end using AI. Its X-Cell platform combines generative protein design, perturbational multi-omic data generation, and foundation models in a single closed loop. The scale of capital reflects consensus that target biology, not generative chemistry, as the core unsolved problem.
Genomics plc launches Mystra (Oct 2025), the first platform explicitly built on the 2.6× multiplier, ingesting 20,000+ GWAS studies to surface target-indication pairs with the highest human genetic support. Meanwhile, Insilico advances a second AI-discovered asset and the field awaits the first Phase III readout from a genetics-first, AI-designed drug.
Market size & opportunity
The numbers behind the bottleneck
Cause of failure (share of the 85%)
Phase 1 plus Phase 2+ safety-related failures
Inadequate efficacy or target selection
Commercial, PK
Segments are shares of the 85% that fail, not of all drugs. Sources: Dowden & Munro 2019; Kola & Landis 2004. Safety and target-selection failures both trace upstream to inadequate target validation, which human genetic evidence addresses.
20 yr
Median target ID to FDA approval, target-based first-in-class drugs
716
Approved drugs acting on distinct targets (of ~4,500 druggable genes)
Nature Reviews Drug Discovery, 2026: Target identification and assessment in the era of AI
Target identification sits upstream of every dollar spent in drug development, which makes its failure the most expensive error in the industry. Roughly 85% of drugs entering clinical trials fail, and the two largest causes both originate at target selection. About 45% of failures are attributable to safety and toxicity, a figure derived by combining Phase 1 failures, which are predominantly safety and tolerability driven, with Phase 2-and-later safety-related failures, and about 40% to poor efficacy. Because both failure modes trace to whether the target was the right one in human biology, the economic weight of target selection compounds across the entire pipeline. The market is repricing where value is created; capital that once flowed primarily to chemistry and clinical execution is moving upstream toward platforms that can establish causal confidence in a target before a program begins, because the cheapest failure is the one avoided at selection.
The investable wedge is human-genetics-grounded target identification. Drug mechanisms with human genetic support succeed at 2.6 times the rate of those without, yet most target selection still rests on correlative evidence from animal models with a 92.6% false discovery rate. Platforms that systematize causal genetic evidence, such as Genomics plc's Mystra, are being adopted by incumbents including Novo Nordisk that concluded they could not match the capability internally. The durable returns accrue to whoever owns the data and methods that separate causal targets from correlative noise, not to those generating more molecular hypotheses against targets that were never validated.
Competitive landscape
Target identification: from correlation to causation
Hover any company for funding stage and description.
Deep Dives
Insilico Medicine is the strongest evidence that AI-driven target identification can reach the clinic. Its PandaOmics engine integrates multi-omic and text data to prioritize novel targets, and its Chemistry42 platform generates molecules against them. The flagship validation is Rentosertib, an inhibitor of TNIK, a target PandaOmics identified for idiopathic pulmonary fibrosis. The program advanced from project initiation to a preclinical candidate in roughly 18 months at a reported discovery cost near $150,000 excluding wet-lab validation, and Rentosertib subsequently produced positive Phase IIa data showing dose-dependent improvement in lung function. That single readout is the most cited proof that an AI-identified, AI-designed drug can demonstrate human efficacy. The company now reports multiple assets across Phase I and II and major partnerships including a multibillion-dollar collaboration with Sanofi. The principal caveat is that no AI-discovered asset has yet completed Phase III, so the validation is mid-stage rather than complete. For the thesis, Insilico matters because it converts the target-identification argument from theory into a clinical data point; it demonstrates that the upstream selection problem is tractable with the right computational and experimental loop, which is precisely the capability the thesis argues will compound in value.
Xaira Therapeutics represents the largest single capital commitment to the thesis that better target biology, not better chemistry, is the binding constraint. Co-founded by Nobel laureate David Baker and launched in 2024 with over $1 billion, Xaira combines generative protein design built on RFdiffusion with large-scale perturbational biology and multi-omic data generation. The strategic premise is that owning the data which establishes how perturbing a target changes biology is more defensible than owning any individual molecule. In 2025 the company released one of the largest Perturb-Seq datasets to date, signaling that its moat is intended to be proprietary functional-genomics data rather than a single pipeline asset. The scale of the raise is itself a market signal; sophisticated capital concentrated a billion dollars on upstream target biology at a moment when generative chemistry was commoditizing. The risk is the mirror image of the ambition. Xaira is entirely pre-clinical, its output to date is platform and data rather than validated assets, and its time horizon is long. For the thesis, Xaira is the clearest demonstration that the field's largest conviction is being placed on the target-identification layer, which corroborates the central claim that this is where durable value is migrating.
Genomics plc is the most direct embodiment of the 2.6x finding that anchors the thesis. The company spent a decade building what it describes as the world's largest harmonized human genotype-phenotype database, now spanning more than 45,000 genome-wide association studies and trillions of data points, and in 2025 launched Mystra, an AI platform that lets drug developers query that evidence base to identify and validate targets in minutes rather than months. The core claim is precise and verifiable: targets with human genetic support are 2.6 times more likely to succeed in clinical trials, and Genomics has used the dataset to identify over 100 drug targets across cancer, heart disease, and diabetes. The strongest validation is adoption; Novo Nordisk partnered with Genomics after concluding it could not match the capability through internal software engineering, and BridgeBio and Relation Therapeutics are also early users. The company raised a $30 million round from Foresite and F-Prime. The limitation is that human genetics is most decisive for traits with clear causal variants; for multifactorial common diseases, mapping a GWAS signal to the causal gene remains imprecise. For the thesis, Genomics is the systematized version of the central insight.
BenevolentAI is the instructive failure that proves the thesis by counterexample. The company built a sophisticated knowledge-graph approach to target identification, mapping relationships across biological literature and data to surface novel target hypotheses. The platform was scientifically advanced and well-funded, yet its clinical programs disappointed, including a Phase IIa failure in atopic dermatitis, and the company subsequently cut a large fraction of its staff and saw its public valuation collapse. The lesson is the heart of the thesis. Correlative network biology can generate plausible target hypotheses, but it cannot reliably distinguish a causal driver of disease from a confounded association, which is exactly the distinction that human genetic evidence provides. BenevolentAI selected targets that looked compelling in a knowledge graph and failed in humans, the predictable outcome when target selection lacks causal grounding. For the thesis, BenevolentAI is not a reason for skepticism about AI in target identification; it is the strongest available evidence that the specific approach matters. Genetics-anchored platforms are structurally different from correlative ones, and the difference shows up where it is most expensive, in the clinic.
The core challenges
Preclinical models used to validate drug targets have an estimated false discovery rate of 92.6%. Around 1 in every 200 protein-disease pairings in preclinical research is estimated to be truly causal. The rest are correlative noise that looks real in a mouse model and fails in humans.
Every dollar spent on chemistry against a poorly-validated target is almost certainly wasted. Roughly 85% of drugs entering clinical trials fail, and the pattern is not random noise; it is the predictable output of relying on preclinical models with a 92.6% false discovery rate to stand in for human biology.
Many proteins change during disease without causing it. A biomarker of disease progression is not necessarily a driver. Most target ID methods — proteomics, transcriptomics, network biology — identify correlates. Only human genetic studies (Mendelian randomization, rare variant analysis) can demonstrate causal roles of genes in human disease.
The only real validation is if a drug turns out to be safe and efficacious in a patient — but by then, you have spent a decade and hundreds of millions of dollars. Upstream causal evidence is the only tractable solution.
Of ~4,500 druggable genes in the human genome, all approved drugs today target only 716 distinct proteins. That leaves over 3,700 druggable targets never addressed by any approved medicine. Most first-in-class opportunities sit in this unexplored space — where the causality evidence is thinnest and the target ID problem is hardest.
The highest-value targets are precisely the ones hardest to validate. First-in-class drug development requires solving target ID from scratch, not borrowing validation from existing drug classes.
The median interval from target identification to FDA approval for target-based first-in-class drugs approved between 1999 and 2013 was 20 years (25 years for non-target-based drugs). This is not primarily a clinical trial problem; Phase I through approval averages 6-8 years. The lag is in the pre-clinical target validation phase, where targets are tested in wrong models for years before the error is discovered.
Compressing target ID from decades to months, as Insilico demonstrated, is not just a speed advantage. It enables a fundamentally different R&D portfolio strategy.
Phenotypic screening (observe disease cellular behavior and find drugs that change it) vs. target-based screening (pick a target, design a drug for it) have been in tension for 30 years. Both have produced approved drugs. Phenotypic approaches avoid the target selection problem but create a target deconvolution problem: you don't know what your drug is hitting.
Recursion's phenomics-first approach is a real alternative to genetics-first target ID. The field has not yet resolved which produces better clinical success rates at scale.
GWAS identifies genomic regions associated with disease, not individual genes. A single GWAS locus may contain dozens of candidate genes. Mapping from association signal to causal gene requires fine-mapping, colocalization, functional genomics, and eQTL analysis, a complex pipeline that is still semi-manual and error-prone.
The 2.6× clinical success multiplier is strongest when the causal gene is clear (Mendelian traits, coding variants). For common complex diseases, GWAS-to-target mapping remains imprecise.
Steelmanning the bears
“Target validation, not target identification, is the harder problem; you are conflating the two.”
strengthensThis is the most important distinction to get right. Target ID (finding a candidate) and target validation (proving it's causal in human disease) are both broken, but for the same underlying reason: preclinical models have a 92.6% false discovery rate. The fix is the same for both: human genetic evidence that demonstrates causality before any animal experiment is run. The thesis encompasses both, because both fail at the same step.
“AI generative chemistry is the real value driver — target ID is solved by genomics already.”
strengthensHuman genomics has identified thousands of target-indication pairs, but most lack sufficient causal confidence for drug development. The 716 targets currently addressed by approved drugs are the easy ones. The next 3,700 are harder. Generative chemistry tools are commoditizing rapidly — GPT-4-level molecular design is available to every lab. The bottleneck is deciding what to design against.
“BenevolentAI's Phase IIa failure proves AI target ID doesn't work.”
strengthensBenevolentAI's failure is actually the strongest evidence for the thesis. Their platform used correlative knowledge graph approaches without strong human genetic grounding. The failure demonstrated precisely what the thesis predicts: correlative AI without causal validation selects targets that look compelling and fail in humans. Genomics-anchored approaches (Insilico, Xaira, Genomics plc) are structurally different.
“Phenotypic screening bypasses the target ID problem entirely and has produced approved drugs.”
complementaryPhenotypic approaches produce real drugs; aspirin was 'phenotypic' before its target was known. But phenotypic screening creates a target deconvolution problem: you don't know what your drug hits, which complicates dose selection, toxicology prediction, and combination strategies. And for first-in-class biologics (antibodies, gene therapies), phenotypic approaches are essentially inapplicable -- you need a target.
“The 20-year timeline is historical — AI has already compressed this.”
monitorInsilico's 18-month target-to-candidate is real and impressive. But it has produced one Phase IIa dataset. The field needs 5-10 more examples across diverse disease areas before the compression claim is validated at scale. The 20-year median remains the baseline against which all AI platform claims should be measured.
Quotes & signal
Lack of efficacy in the intended disease indication is the major cause of clinical phase drug development failure. Around one in every 200 protein-disease pairings was estimated to be causal, giving a false discovery rate in preclinical research of 92.6%.
The probability of success for drug mechanisms with genetic support is 2.6 times greater than those without. We are far from reaching peak genetic insights to aid the discovery of targets for more effective drugs.
A target is only fully validated once a drug based on it receives approval by regulatory agencies.
The only real validation is if a drug turns out to be safe and efficacious in a patient. A major challenge is that many targets fail late in the clinic, several years after initiation of the project.
References
- 1.Dowden & Munro, 2019 · of drugs fail in clinical trials
- 2.Hingorani et al., Sci Rep 2019 · Preclinical model false discovery rate
- 3.Minikel et al., Nature 2024 · Clinical success uplift with human genetic support
- 4.Eder et al., Nat Rev Drug Discov, 2014 · Median target ID to FDA approval, target-based first-in-class drugs
- 5.Nature Reviews Drug Discovery, 2026: Target identification and assessment in the era of AI · Approved drugs acting on distinct targets (of ~4,500 druggable genes)
- 6.Hingorani et al.: PMC6906499 — foundational paper on human genomics improving drug target ID success rates
- 7.Minikel et al.: Nature 2024 — refining the impact of genetic evidence on clinical success
- 8.Nature Reviews Drug Discovery: Target identification and assessment in the era of AI, Nature Reviews Drug Discovery, 2026
- 9.Technology Networks: Target Identification and Validation in Drug Discovery review
- 10.Dowden, H. & Munro, J., 2019: Trends in clinical success rates and therapeutic focus, Nature Reviews Drug Discovery 18, 495-496
- 11.Kola, I. & Landis, J., 2004: Can the pharmaceutical industry reduce attrition rates? Nature Reviews Drug Discovery 3, 711-715
Confidential · 9/3/2026