AI Drug Discovery: Why Phase 2 Is the Real Test

AI Drug Discovery: Fast to the Clinic, Unproven in Patients

AI drug discovery has delivered on its first promise faster than many sceptics expected. Companies built around machine learning report moving from project start to a preclinical candidate in well under two years, and well over a hundred AI-derived assets have now entered human trials. The second promise, that those molecules are more likely to become approved medicines, is still unproven. Most of the pipeline sits in early safety studies. Phase 2, where a drug first has to show that it actually helps patients, is historically where the largest share of candidates fail, and only a handful of AI-originated drugs have come through it so far.

Where AI Speeds Up the Pipeline and Where It Does Not

AI compresses some stages dramatically and leaves the hardest scientific question almost untouched.

What AI Drug Discovery Has Actually Delivered So Far

The clearest gain is speed in the discovery phase. Companies built around machine learning commonly report moving from the start of a programme to a preclinical candidate in something like twelve to eighteen months, where a conventional programme might take three or four years. Generative chemistry lets teams explore far more molecular designs than a medicinal chemistry group could make by hand, and property prediction reduces the number of design and test cycles needed to reach a viable compound.

The clinical pipeline has grown accordingly. Industry analyses now count well over a hundred AI-derived assets that have entered human trials across dozens of companies. The most cited milestone came in 2025, when Insilico Medicine published mid-stage results for rentosertib in idiopathic pulmonary fibrosis, widely described as the first drug with both an AI-identified target and an AI-designed molecule to report Phase 2a data in a peer-reviewed journal. Against that, industry trackers reported that no fully AI-designed drug had received FDA approval by mid-2026.

Why AI Drug Discovery Looks Strongest in Phase 1

Several analyses have reported high Phase 1 success rates for AI-derived molecules, and that finding is genuinely encouraging. It is also easy to over-read. Phase 1 mainly tests safety, tolerability, and how a drug behaves in the body, which is exactly where computational design is strongest. A molecule can be clean, stable, well absorbed, and entirely useless against the disease it was designed for. Good Phase 1 numbers show the chemistry is working. They do not yet show the medicine is.

Why Phase 2 Is Where Biology Pushes Back

The deepest bet in any drug programme is the target hypothesis: the claim that changing the activity of a particular protein or pathway will meaningfully change the disease in people. AI can design a better key. It cannot guarantee that the lock matters. Phase 2 is typically the first time that hypothesis meets real patients at a dose intended to work, and across the industry it has historically been the stage where the largest share of candidates fail, most often on efficacy rather than safety.

The underlying reason is a translational gap that computation has not closed. Models learn from cell lines, animal studies, and published literature, and each of those is an imperfect stand-in for human disease. Animal models frequently fail to predict how people respond. Published findings carry well-documented reproducibility problems. A system trained on that evidence can inherit its blind spots while presenting conclusions with impressive precision.

None of this means AI programmes will fail more often than conventional ones. It means the evidence that would show they succeed more often, meaningful Phase 2 and Phase 3 efficacy data across many assets, is only now beginning to arrive.

Five Questions to Ask About Any AI-Discovered Asset

Whether you are an investor, a partner, or a scientist inside the company, these questions cut through the label.

1. Did AI find the target, design the molecule, or both? These are very different claims, and the target choice usually carries the larger efficacy risk.
2. What human evidence links the target to the disease? Genetic association, human tissue data, and earlier clinical signals count for far more than animal results.
3. Which stage has it completed, not merely entered? Announcements often mark trial starts, which say nothing about outcomes.
4. How were trial patients selected? Biomarker selection can improve the odds of success while shrinking the eventual market.
5. Were success criteria fixed before the data arrived? Endpoints chosen after the fact make any result look better than it is.

Pros and Cons of AI-Led Discovery Programmes

The balance sheet is positive, with one large item still unresolved.

• Pro: shorter discovery timelines. Reaching a candidate faster means more shots on goal for the same budget.
• Pro: wider chemical exploration. Generative design can reach structures that human teams would not have considered.
• Pro: better early property prediction. Fewer compounds fail late for avoidable reasons such as poor absorption or instability.
• Con: target risk is unchanged. A beautifully designed molecule aimed at the wrong biology still fails.
• Con: hype compresses patience. Investors expecting speed can lose interest precisely when trials need time.
• Con: training data carries its own flaws. Models learn from literature and assays with known reproducibility gaps.

The Literature Problem Inside Target Selection

Target selection leans heavily on synthesising published evidence: genetic studies, pathway biology, earlier trials, and competitor programmes. Teams increasingly use language models to digest that literature, and those summaries can overstate effect sizes, merge separate studies, reverse the direction of a finding, or cite papers that do not exist. We have covered how AI hallucinations are already polluting scientific research more broadly, and target dossiers are not immune.

It helps to separate two very different uses of AI. Generative chemistry models propose molecules that are then synthesised and tested in the lab, so their errors get exposed by experiment. Language models summarising literature produce prose that feeds decisions directly, so their errors surface only if someone reads the original paper. The second kind of error can quietly send an excellent molecule after the wrong biology. Regulated development work adds its own layer of obligations, which we discussed in validating GenAI in pharma.

Real Scenarios Worth Thinking Through

These scenarios are illustrative, showing how AI drug discovery risk plays out in practice rather than presented as verified case studies.

An investor reviewing a biotech sees an AI-discovered label and a clean Phase 1 readout and models the programme as largely de-risked. On closer reading, the target evidence rests mainly on mouse models. Phase 2 misses its primary efficacy endpoint, and the valuation had assumed the opposite.

A discovery team uses a language model to assemble a target dossier. One key human genetic association is summarised as supporting the target, when the original study actually reported the opposite direction of effect. A senior scientist catches it only because they happen to remember the paper from years earlier.

A company uses machine learning to select Phase 2 patients by biomarker profile. The trial succeeds in the selected group, which is real progress. The approved population, if the drug gets there, will be much smaller than the original market forecast assumed, and the business plan has to be rebuilt around it.

“After testing multiple AI models on coding, research, and business prompts, combined outputs produced more reliable results than any single model.” Internal multi-model evaluation, Talkory research team.

Why Talkory Wins

Talkory is not a molecule design tool, and it does not replace experimental validation. Where it fits is the evidence layer around expensive decisions: target dossiers, competitive landscapes, due diligence summaries, and investment memos. Running the same question across GPT, Claude, Gemini, Grok, Perplexity Sonar, and Kimi K3 shows where accounts of the evidence agree and where they diverge.

Divergence on a key claim, such as whether a genetic association supports a target or which trial reported which outcome, is the cue to open the primary paper before that claim goes into a go decision. In a field where one wrong assumption can cost years, that is a very cheap check.

Final Verdict

AI drug discovery is real progress, and the speed gains are not marketing. The honest scorecard, though, is still mostly written in Phase 1, where computational strengths show most clearly. Phase 2 is where target hypotheses meet patients, and it will decide whether AI changes the success rate of drug development or only its speed. Read pipelines by stages completed, ask what human evidence supports each target, and check literature summaries against primary sources before they drive decisions measured in years and hundreds of millions.