Patients Want Pills
Why oral small molecules remain brutally difficult to develop, and how AI could finally change that.
TL;DR
• Pills are the only modality that scales to billions of patients. Given the choice, that’s what they pick.
• Antibody discovery has been steadily commoditizing. A competent team can get from validated target to IND in 12 to 18 months for $5 to $15 million. Oral small molecules haven’t had that collapse: 4 to 6 years and $20 to $50M+ is still the norm, and the hard targets push into nine figures.
• The gap is structural. Antibodies have to solve one problem well (target binding) and a handful of developability filters around it. Orals have to solve six at once: binding, selectivity, permeability, metabolic stability, safety, and synthesis. Joint probability is what makes small molecules as a modality so brutal.
• The starting scaffold matters more than any other single variable. Good chemical matter at hit-to-lead is the difference between a two-year program and a six-year one.
• Derek Lowe has been arguing for years on In the Pipeline, the two things that kill most clinical programs (target selection and human toxicity) remain largely outside its reach for AI. AI is however making real progress all along the preclinical value chain: AI powered target selection (not discussed here, maybe a subsequent piece?), generative chemistry, selectivity prediction within protein families and early ADMET.
• Human clinical trials are nevertheless the tests that matter and the early signal has started to come in. Rentosertib (Insilico) cleared Phase 2a with positive data in Nature Medicine in 2025. Zasocitinib (Schrödinger/Nimbus/Takeda) hit both Phase 3 psoriasis endpoints later that year.
• Zasocitinib is arguably the closest the field has to an AI-enabled approval. If Takeda's filing succeeds in FY2026, it would be the first commercial validation of physics-plus-ML drug discovery at the approval stage, distinct from end-to-end generative AI but a real milestone.
• Per industry trackers, roughly half a dozen AI-enabled small molecules are in or approaching pivotal trials in 2026: zasocitinib, zovegalisib, lirafugratinib, REC-1245, and a small handful of others. The AI-discovered biologics pipeline is earlier, with most programs still in Phase 1-2.
• We might be closer than expected by industry experts: “The future has already arrived. It's just not evenly distributed yet.”
Ask a patient with diabetes or obesity to choose between a weekly injection and a daily pill at comparable efficacy, and a lot of them will pick the pill. The market has been telling us this for years. Rybelsus, the first oral GLP-1, has been a multi-billion-dollar product since 2022, with roughly $3.2 billion in 2025 sales from type 2 diabetes alone. Oral Wegovy, the first oral peptide GLP-1 for obesity, launched in December 2025 and hit roughly half a million U.S. prescriptions in its first three months. Lilly’s just-approved Foundayo (orforglipron) is the first true small-molecule oral GLP-1. Analyst peak-sales estimates run from $10 billion bearish to north of $30 billion bullish.
These are not niche numbers. Even with long-acting injectable biologics now the default for chronic disease, patients still want tablets when they can get them. No cold chain, no needle anxiety, no pharmacy routine, no sharps disposal. And if a side effect becomes intolerable, you can stop the drug for a few days. You can’t take back a long-acting injection.
The deeper reason oral drugs matter is access at scale. Cold-chain logistics, infusion centers, and specialty pharmacies are the defining bottlenecks of modern medicine’s access problem, and a pill sidesteps every one of them. It ships at room temperature, costs pennies to manufacture, and doesn’t need a healthcare worker to administer. Nothing else on the menu scales to eight billion patients the same way.
Why, then, don’t we have more of them? The flagship orals of the last few years (TYK2 inhibitors, oral GLP-1s, KRAS G12C binders) are still the exception. Most of cancer immunotherapy, cytokine immunology, and targeted protein degradation remains stubbornly injectable.
The commoditization gap
Antibody discovery is moving toward commoditization. Twenty years ago, getting from a validated target to a clinical antibody took three to five years and a world-class immunology platform. Today, between display technologies, transgenic animals, and AI-native antibody design, a competent team can deliver a developable lead in twelve to eighteen months for a few million dollars. The risk has shifted downstream. You’ll probably get an antibody. Whether it does anything useful in humans is the open question.
Oral small molecules have had no equivalent collapse. Timelines and costs from validated target to clinical candidate still look like they did in 2005: four to six years, thirty to fifty million dollars, with the hard targets pushing into nine figures. Most of the spend lands in the back half of discovery and early safety. For allosteric sites, shallow pockets, and intracellular protein-protein interactions, the numbers get worse, not better.
That is the gap. Biologics are converging toward a cost and timeline floor. Small molecules are not.
Selectivity, inverted
Start with what I’ll call the selectivity inversion. It is the cleanest way to see why orals are structurally harder than antibodies.
Antibodies get selectivity almost for free. The immune system spent a few hundred million years evolving combinatorial machinery for telling one molecular surface from another, and we’ve industrialized that machinery. Hand a modern antibody platform a target, and off-target binding is usually the easy part. What you pay for that selectivity is distribution. Antibodies are enormous. They don’t cross membranes, don’t enter cells, don’t cross the blood-brain barrier, and don’t survive the gut.
Small molecules are the mirror image. A well-designed one can reach most tissues, including the brain, and engage intracellular targets that are invisible to biologics. What you pay for that distribution is selectivity. The druggable proteome runs to several thousand proteins, most with pockets that look broadly similar to the one you’re trying to hit. A drug that can go anywhere can bind anywhere, and every off-target interaction is a potential toxicity.
That trade-off has dictated drug modality choice for forty years whenever the goal is to inhibit or antagonize. If selectivity is the binding constraint, use an antibody. If distribution is, use a small molecule. Most of the targets considered undruggable by orals sit in a regime where neither modality wins cleanly. They are typically intracellular, putting them out of reach of antibodies, while their pockets are shallow, transient, or featureless, defeating standard small-molecule design. You need the distribution of a small molecule and the selectivity of an antibody. No modality has yet delivered both reliably.
Agonism flips parts of this picture. Where the natural ligand is a peptide or protein, small-molecule agonism is the harder problem. A small molecule has a fraction of the contact surface a peptide ligand can deploy, and reproducing the same activating geometry through a much smaller footprint is a hard structural problem. That’s why the recent oral GLP-1 small molecules are such a notable achievement.
Six problems stacked in series
There’s a second way to count the same problem. An oral small molecule has to clear six filters at once, and joint probability is what makes the chemistry brutal.
It has to bindthe target with enough potency to be dosable. It has to be selectiveenough not to hit the rest of the proteome. It has to permeate the gut wall, survive first-pass metabolism in the liver, and reach the target tissue at adequate concentration. It has to be metabolically stable enough for a reasonable half-life without being so stable that it accumulates. It has to be safe: no hERG liability, no reactive metabolites, no idiosyncratic signals. And it has to be synthesizable at a cost and scale that makes the whole enterprise viable.
Each is its own sub-field. Each has its own failure modes, its own assays, its own predictive models. They aren’t independent. Push potency up and solubility often collapses. Improve metabolic stability and you can create tox liabilities. It’s a six-dimensional optimization where every axis is non-convex, and most of the gradients are invisible until after you’ve made the molecule.
An antibody program faces analogues of several of these filters: cross-reactivity, thermal stability, immunogenicity, manufacturability. The difference is dimensional. Most antibody developability liabilities can be addressed by engineering the framework, surface, or constant region without touching the binding paratope, so binding and developability are largely decoupled. Small-molecule chemistry has no such separation. Every atom does multiple jobs at once: a change made to fix metabolic stability often destroys potency or selectivity, and improving solubility can wreck permeability. That coupling, more than the filter count itself, is what makes the optimization brutal.
Organic synthesis is still hard, slow, and expensive
ADMET and safety testing eat most of an oral program’s budget. They are also the filters that retire compounds. But synthesis is what gates iteration speed.
A catalog compound from Enamine runs about $100. A fully custom synthesis compound costs $1,000 to $2,000 per molecule. A single chemist makes only one to five fully custom molecules per week. A round of serious optimization means hundreds of compounds. Best case, that round takes a month. Realistically, two to three. Then you look at the data, redesign, and do it again. Programs routinely run ten or more such rounds before declaring a candidate. Library platforms like Enamine REAL can deliver hundreds of catalog-friendly compounds in parallel within days, and for a hit-finding screen this is now the default. The bottleneck reappears at hit-to-lead and beyond, where novel scaffolds, complex stereochemistry, and bespoke bioisosteres pull you back to one to five custom molecules per chemist per week.
¹ The unit of comparison is asymmetric. For mAbs, this counts lead variants characterized through developability assessment. For small molecules, it counts distinct chemical structures synthesized and assayed. The best-case AI-native column reflects published numbers from Insilico’s lead programs (rentosertib and garutadustat); other AI-native programs report higher counts after virtual screening triage.
Look at the variance in the small-molecule columns. Ask any experienced medicinal chemist what separates a two-year program from a six-year one and they will tell you: the quality of the starting scaffold. Good chemical matter at hit-to-lead means fast optimization, few surprises, clean ADMET. Mediocre matter means years of suffering. The field has never been good at predicting which scaffolds will be which.
The clinical coda
Even if you clear preclinical, the small-molecule modality has one filter that often comes back in the clinic. Drug-induced liver injury (DILI), hepatotoxicity caused by the drug itself rather than by underlying liver disease, is a leading cause of post-approval withdrawals and a major contributor to late-stage attrition. Some estimates put it at 20 to 25% of clinical failures.
Two flavors matter. Intrinsic DILI is dose-dependent and reproducible in animal models. Acetaminophen overdose is the textbook case. Mostly a solved problem; it gets caught in Phase 1. Idiosyncratic DILI is the dangerous form. Immune-mediated rather than dose-dependent, it shows up in roughly one in ten thousand patients, is routinely missed by rodent and primate studies, and usually surfaces in Phase 3 or post-marketing. Both forms have killed promising programs in the last decade.
Fasiglifam (TAK-875) is the canonical case. Takeda's oral GPR40 agonist for type 2 diabetes was killed mid-Phase 3 in December 2013 after liver enzyme elevations emerged across nine global trials. 2.1% of treated patients crossed 3× the upper limit of normal versus 0.5% on placebo. The drug worked. HbA1c reductions of 0.6 to 0.8% were significant and reproducible. Takeda walked away anyway, because the signal couldn't be reconciled with a 5,000-patient cardiovascular outcomes trial already enrolling. The mechanism, bile acid transporter inhibition plus reactive metabolite formation, would have been flagged by the right in vitro panel run early enough. Which is why fasiglifam is now the validation compound for almost every newer in silico DILI tool.
Of course, biologics have their own hepatotoxicity stories, checkpoint inhibitors most prominently. The mechanisms are different, and they’re not the topic of this piece. For oral small molecules, DILI is a structural risk of the modality itself: every CYP-metabolized scaffold is one reactive metabolite away from trouble.
What AI actually changes
So does AI fix this?
Some of it. Not all of it. And so far, mostly the parts that were already getting better. Derek Lowe has been making this point for years on In the Pipeline: AI is best at problems in almost inverse proportion to how much they cost the industry. Target selection and human toxicity, the two things that kill most clinical programs, have remained largely outside its reach.
What’s working today
The real wins are in generative chemistry and selectivity prediction. The last three years have produced a generation of models (diffusion-based, flow-based, equivariant) that propose drug-like molecules conditioned on a pocket. The molecules they generate often carry binding profiles and physicochemical properties that would have taken a medicinal chemist months to reach by hand. Selectivity prediction within protein families is improving where the structural and bioactivity data is dense, kinases being the clearest case. Early ADMET endpoints (solubility, permeability, microsomal stability, baseline CYP inhibition) are improving fast. In silico triage is now useful rather than misleading. Physical automation is catching up. Chemify and similar automated synthesis platforms are closing the loop between design and synthesis, compressing cycle times from months to weeks for full custom synthesis of complex scaffolds.
First clinical signal
3 programs are among the most prominent examples in the public conversation about whether AI can ship a real drug. They’re worth looking at in detail.
Rentosertib (Insilico Medicine). Insilico’s PandaOmics platform pulled TNIK out of multi-omics data as a candidate target for idiopathic pulmonary fibrosis, with no human hypothesis driving the choice. Their generative chemistry platform, Chemistry42, generated tens of thousands of virtual TNIK candidates and ultimately synthesized 78 molecules to nominate the preclinical candidate. The drug went from project start to preclinical candidate in 18 months for a disclosed budget of approximately $2.6 million, and from target to Phase 1 in 30 months. The Phase 2a topline, published in Nature Medicine in June 2025, showed +98.4 mL FVC at 60 mg QD versus a placebo decline. This is the first clinical proof-of-concept for a drug where both target and molecule were AI-originated. Insilico is preparing for pivotal trials.
Garutadustat (Insilico Medicine). A gut-restricted PHD inhibitor for inflammatory bowel disease. Twelve months from project start to PCC, on roughly 115 synthesized molecules. Phase 1 cleared, Phase 2a just started. Two fast preclinical-to-clinic walks from the same platform are hard to dismiss as luck. Reproducibility is the right question to ask.
Zasocitinib (Schrödinger / Nimbus / Takeda). Currently the most clinically advanced AI-enabled small molecule by phase. TAK-279 is a selective TYK2 pseudokinase domain inhibitor discovered at Nimbus on Schrödinger’s physics-plus-ML platform. Free energy perturbation calculations drove selectivity over JAK1/2/3. Takeda acquired the program from Nimbus in February 2023 for $4 billion upfront plus up to $2 billion in sales-based milestones, one of the largest M&A bets ever placed on a computationally-discovered small molecule. Takeda ran the LATITUDE-PsO-3001 and LATITUDE-PsO-3002 Phase 3 trials in moderate-to-severe plaque psoriasis. Both hit primary and secondary endpoints in late 2025. Roughly half of treated patients achieved PASI 90 (90% reduction in Psoriasis Area and Severity Index, the standard measure of skin clearance) at week 16 and ~30% achieved PASI 100. Takeda is filing for FDA approval in fiscal year 2026. This is physics-plus-ML chemistry rather than end-to-end generative AI, and, on current evidence, the more clinically validated of the two approaches.
The category-level numbers point the same way. Boston Consulting Group’s 2024 analysis of AI-native pipelines (Jayatunga et al., Drug Discovery Today, 2024) reported Phase 1 success rates of 80 to 90% (21 of 24 completed trials) versus a historic 40 to 65% industry average. Nature Biotechnology cited the same range in a 2025 editorial. Phase 2 success falls back to roughly 40%, in line with historic norms. AI helps with drug-like properties and ADMET.
The next test is Phase 3. Industry trackers count roughly 15 AI-discovered programs entering pivotal trials in 2026. That is where the early-stage advantage either converts to an approval or stops mattering. Benchmark wins don’t pay for Phase 3 trials.
Pooling ADMET data
An ecosystem around ADMET and safety is finally taking shape. Axiom Bio is making the most serious attempt yet at in silico DILI prediction from primary human hepatocyte data, with reported sensitivity and specificity that beat traditional assays at a fraction of the cost. Inductive Bio is building ADMET foundation models on a pre-competitive industry consortium and recently won the Polaris blind ADMET competition outright.
Lilly’s TuneLab is unprecedented for a top-five pharma. It opens roughly a billion dollars’ worth of proprietary preclinical ADMET data to biotech partners through federated learning. Insitro is training in vivo ADMET models against that dataset. Apheris has launched a parallel federated network anchored by Lundbeck, Orion, Servier, and Recursion. No single one of these will solve the problem. Taken together, they are the first real attempt to pool what has been the industry’s most jealously guarded data type.
The regulatory tailwind, with caveats
The regulatory environment is shifting, but slower and more narrowly than the headlines suggest. The FDA’s April 2025 roadmap to reduce animal testing in preclinical safety studies lists AI-based computational models among acceptable alternatives. Phase one is monoclonal antibodies. New chemical entities come later.
Since then: draft FDA guidance on reducing nonhuman primate testing for monoclonal antibodies (December 2025); draft guidance on validating new approach methodologies (March 2026); an April 2026 Year 1 progress report claiming all first-year goals met. In January 2026 the FDA and EMA jointly published ten guiding principles for AI in drug development. They are non-binding.
For small molecules, in May 2026, little has changed in practice. Sponsors still file standard tox packages. The FDA’s January 2025 draft guidance on AI in regulatory decision-making is still draft, with no finalization date. Fossler and Garner make the unsexy point in Clinical Pharmacology in Drug Development (March 2026): we still lack the biology to predict human toxicity without animals for most therapeutic areas. The roadmap is a direction of travel, not a regime change.
The direction matters. For decades, regulators required animal data regardless of in silico accuracy. That default is now up for revision once the science is good enough. The science isn’t there yet.
What AI is not yet fixing
The long tail is where AI still loses. Idiosyncratic DILI is the obvious example: labeled cases are rare, mechanisms are heterogeneous, the positive class is too small to learn. Most models settle on predicting “safe,” hit 99% accuracy on a meaningless metric, and miss every case that matters. Long-term safety in special populations and drug-drug interactions at the edges are largely untouched. So is the “beyond rule-of-five” problem (macrocycles, PROTACs, molecular glues), where oral bioavailability is the dominant failure mode. Lowe’s two big ones, target selection and human tox, sit firmly in this tail.
When we’ll know it’s working
The early signal is real. The late signal is not in yet. Three readouts will tell us when AI has cracked the modality.
First, two or three independent teams shipping clinical-stage oral small molecules on hard targets (allosteric, cryptic, mutant-selective) at a 50% reduction in both compound count and time, with clean PK and ADMET. Insilico has done it once. Whether anyone outside their platform can reproduce it is the open question.
Second, an idiosyncratic DILI predictor that prospectively flags compounds that go on to cause hepatotoxicity in humans. Sensitivity above 50%, specificity above 90%, validated across three independent cohorts. Axiom Bio is the closest credible attempt I’ve seen. Independent prospective validation will tell.
Third, a Phase 3 approval for a fully end-to-end AI-discovered drug where both target and molecule were AI-originated. Rentosertib is the candidate. The readout will arrive in the next two to three years.
Until then, the field is where it has been for the last three years: the early-stage parts are getting cheaper and faster, and the late-stage parts (which is where the cost and the failure live) are not.
The frontier
None of this is solved. Oral small molecules aren’t commoditized, and probably won’t be for another decade. Plenty of fundamental work remains. Generative models conditioned jointly on selectivity and ADMET. Pocket discovery for targets currently written off as undruggable. Idiosyncratic tox prediction from structure. Autonomous closed-loop optimization at scale. Each is a real scientific problem with a real clinical payoff.
We keep being told that AI is about to commoditize drug discovery. For antibodies, it plausibly already is. For orals, the frontier is still wide open. The teams that figure out how to stack a dozen probabilistic models into a reliable oral-drug design engine will change what is druggable.
That is a problem worth spending a decade (or two) on.
Appendix: Sources and further reading
On the modality landscape
• Lowe, D. In the Pipeline, commentary on AI in drug discovery, 2022–2025. Multiple posts on target selection and human toxicity as the unsolved problems.
• Hay, M. et al. Clinical development success rates for investigational drugs. Nature Biotechnology, 2014.
• Thomas, D. et al. Clinical Development Success Rates and Contributing Factors 2011–2020. BIO, Informa, QLS Advisors, 2021.
Oral GLP-1s
• Novo Nordisk, Rybelsus and oral semaglutide / Wegovy regulatory and commercial disclosures, 2022–2026.
• Eli Lilly, ATTAIN Phase 3 program for orforglipron (Foundayo) and FDA approval materials, 2025–2026.
AI clinical pipeline analysis
• Jayatunga, M.K.P., Ayers, M., Bruens, L., Jayanth, D., Meier, C. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discovery Today, 2024.
• Editorial, The AI drug revolution needs a revolution. npj Drug Discovery / Nature Biotechnology commentary, 2025.
Rentosertib (Insilico Medicine)
• Xu, Z., Ren, F., Wang, P., et al. A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial. Nature Medicine, published online June 3, 2025.
• Insilico Medicine 2025 Annual Results, March 2026.
• Insilico Medicine, Case Study: Insilico’s Transformation (companion to Harvard Business School case study on rentosertib), insilico.com/casestudy, accessed May 2026. Source for the 78-molecule and 18-month rentosertib benchmarks, and for the 60–200 molecules-per-program portfolio average.
Zasocitinib (Schrödinger / Nimbus / Takeda)
• Takeda, LATITUDE-PsO-3001 and LATITUDE-PsO-3002 Phase 3 topline release, December 2025; full data presented at AAD 2026.
• Schrödinger Inc., physics-plus-ML platform technical disclosures and FEP+ benchmark publications, 2020–2024.
ADMET and DILI
• Axiom Bio, in silico DILI prediction from primary human hepatocytes, technical materials, 2024–2025.
• Inductive Bio, ADMET foundation models and Polaris blind benchmark results, 2025.
• Eli Lilly TuneLab and federated ADMET partnerships (Insitro, Apheris consortium with Lundbeck, Orion, Servier, Recursion), 2024–2026.
Regulatory
• FDA, Considerations for the Use of Artificial Intelligence to Support Regulatory Decision Making for Drug and Biological Products, draft guidance, January 2025. Comment period closed April 7, 2025; remains draft as of May 2026.
• FDA, Roadmap to Reducing Animal Testing in Preclinical Safety Studies, April 2025. Initial focus: monoclonal antibodies. New chemical entities to follow.
• FDA, draft guidance on the reduction of nonhuman primate testing for monoclonal antibodies, December 2025.
• FDA, draft guidance on validating new approach methodologies (NAMs), March 2026.
• FDA, Reducing Animal Testing in Nonclinical Studies: Year One Progress and the Path Forward, April 2026.
• FDA-EMA, Guiding Principles of Good AI Practice in Drug Development, January 14, 2026. Non-binding.
• Fossler, M.J. and Garner, C.E. The FDA Roadmap to Reducing Animal Testing in Preclinical Safety Studies: Where Will It Lead Us?Clinical Pharmacology in Drug Development, March 2026.
Fasiglifam and DILI as a structural modality risk
• Marcinak, J.F. et al. Liver Safety of Fasiglifam (TAK-875) in Patients with Type 2 Diabetes: Review of the Global Clinical Trial Experience. Drug Safety, 2018.
• Menon, V. et al. Fasiglifam-Induced Liver Injury in Patients With Type 2 Diabetes: Results of a Randomized Controlled Cardiovascular Outcomes Safety Trial. Diabetes Care, 41(12):2603–2609, December 2018.
• Otieno, M.A. et al. Fasiglifam (TAK-875): Mechanistic Investigation and Retrospective Identification of Hazards for Drug-Induced Liver Injury. Toxicological Sciences, 163(2):374–384, June 2018.