Can AI really predict which cancer drugs will work?

Last updated: 18 November 2026
Cover of NewMarketPitch's AI Drug Discovery: Reality Check report Can AI tell which cancer drugs will work for which patients? Somewhat, in small tests. Orakl says 90%, but its published check is 7 of 8 past trial matchups. Pangea’s AI scores 0.78 against 0.68 for today’s PD-L1 test. Chart from NewMarketPitch’s AI Drug Discovery report.How often do cancer drugs fail in clinical trials? Almost always. Only about 5 in 100 cancer drugs that start trials get approved, against 8 for all drugs. Three in four stop in phase 2. Chart from NewMarketPitch’s AI Drug Discovery report.Do AI-found drugs succeed more often in trials? Not proven. About 8 in 10 pass phase 1, but in a 2026 review of 129 drugs only 1 in 5 passed phase 2. Chart from NewMarketPitch’s AI Drug Discovery report.Which AI-discovered drugs are furthest along in human trials? Insilico’s rentosertib and Generate’s GB-0895 reached phase 3. None is approved, and Recursion and Iambic plan phase 3 trials in 2027. Chart from NewMarketPitch’s AI Drug Discovery report.How much do drug giants really pay AI drug discovery firms? Up front, $115M at most, like Lilly’s deal with Insilico. The billions announced come only if the drugs succeed. Chart from NewMarketPitch’s AI Drug Discovery report.Which AI drug discovery companies are worth the most? Isomorphic Labs, by far, at about $50B reported, with no drug in human trials yet. Insilico, with one in phase 3, trades at $2.5B. Chart from NewMarketPitch’s AI Drug Discovery report.What went wrong for AI drug discovery’s early stars? Recursion stopped two phase 2 drugs and cut about 20% of staff in 2025. BenevolentAI left the stock market and Exscientia was absorbed by Recursion. Chart from NewMarketPitch’s AI Drug Discovery report.Who has the most AI-found drugs in human trials? Insilico Medicine, with 10 programs in trials and one in phase 3. Isomorphic Labs raised $5B in November 2026 but has no drug in trials yet. Chart from NewMarketPitch’s AI Drug Discovery report.

Every funding round, every investor, and who is really ahead in AI drug discovery, in one document. Have a look ↓

Summary

Yes, AI can already predict some cancer-treatment results well enough to help researchers make choices. But it cannot yet pick the successful new cancer drugs across the board. The best models work on specific cancers, medicines and patient groups, and their accuracy depends on exactly what they are asked to predict.

That distinction is getting attention because Orakl Oncology has raised $34 million and says its technology reaches 90% accuracy. The startup grows patients' tumors in the lab, tests medicines on them and uses AI to estimate how future trials will turn out. Its investor Norrsken says the company has raised $50 million altogether.

There is a lot to gain from getting those choices right. In BIO's historical analysis, just 5.3% of oncology programs entering Phase I reached approval. A later study put the chance at 5.7% for its latest nine-year window. Different datasets give slightly different figures, but both suggest roughly 19 in every 20 cancer programs fall short.

Orakl has published more detailed evidence than the funding headline suggests. Its SCOPE preprint compared forecasts with 32 treatment arms from 23 published colorectal and pancreatic cancer trials. The preprint reports an average error of 0.82 months in predicting how long patients went before their cancer worsened. It also says SCOPE picked the better treatment in seven of eight comparisons, although nearly nine in ten evaluated arms used conventional chemotherapy.

Then came a useful real-world check. SCOPE had forecast that daraxonrasib would beat chemotherapy in pancreatic cancer. The drug later received FDA approval after a randomized trial in 500 patients found better survival and longer disease control. Orakl got the direction right, but its simulation involved different patient groups and comparison regimens, so the two sets of figures cannot be scored as a matched test.

Other models use quite different evidence. COMPASS reads patterns of gene activity in tumors to help sort immunotherapy responders. SCORPIO uses ordinary blood tests and medical records to estimate outcomes after immunotherapy. Researchers have also trained AI to read prostate cancer biopsy slides and find groups that benefit differently from chemotherapy. These models solve separate pieces of the problem.

Even collecting the right biological samples is difficult. In a prospective colorectal cancer study involving 232 enrolled patients, researchers successfully established organoid cultures from about half of biopsied cases overall. Only 42 cultures were included in the interim drug-response analysis. A powerful prediction model will help fewer people if usable tumor samples cannot be prepared reliably.

Predicting final drug approval brings another trap: most programs fail, so a model that always says 'failure' can appear remarkably accurate. One approval-prediction study correctly classified 43 of 44 later outcomes, but 41 of those 44 programs had failed. Its useful achievement was identifying all three approvals with one false alarm, not merely posting 97.7% accuracy.

Drugmakers are starting to test the tools on real decisions. Orakl has announced work involving four Sanofi oncology assets. Its claim that 18 of the top 20 pharmaceutical companies are in its commercial pipeline refers to prospects, not confirmed paying customers. We need to see which drug programs change course because of these forecasts and what happens next. Drugmakers are starting to test these tools. How much of pharma's deal money is really paid is a question we take up in our report on AI drug discovery.

Our answer is cautiously positive about the science and much firmer about its limits. AI drug discovery tools can help choose the next experiment, find patients likely to respond and flag weak trial candidates. To earn trust for new medicines, they still need predictions recorded in advance, independent tests and evidence that following their advice improves results for patients or drug developers.

Written by the New Market Pitch teamWe track new markets so founders and investors can move faster. About us

Can AI tell which cancer drugs will work for which patients? Somewhat, in small tests. Orakl says 90%, but its published check is 7 of 8 past trial matchups. Pangea’s AI scores 0.78 against 0.68 for today’s PD-L1 test. Chart from NewMarketPitch’s AI Drug Discovery report.
AI Drug Discovery reportThis page and 70+ morePDF Get the report

This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.

Why is Orakl Oncology making news in AI drug discovery?

Orakl Oncology is in the spotlight after raising $34 million to help predict cancer-trial results with AI.

Norrsken co-led the French startup's Series A with Singular and says the round brought Orakl's total funding to $50 million. The investor announced the deal on November 12, 2026. The company wants to produce five times more laboratory data, expand from colorectal and pancreatic cancers into lung and gastric cancers, and work on more than ten drug programs.

Orakl grows small, three-dimensional samples of patients' tumors, exposes them to medicines and combines the results with medical histories. It then tries to predict how groups of patients will respond in future trials. The big question behind the funding is whether drugmakers can rely on those forecasts before committing to years of clinical testing.

Why do so many cancer drugs fail in clinical trials?

Most cancer drugs still fail because promising lab results often turn into disappointing results in patients.

BIO, Informa Pharma Intelligence and QLS Advisors tracked 12,728 clinical and regulatory phase transitions involving 9,704 drug-development programs over 2011-2020. Across all diseases, just 28.9% of programs advanced from Phase II, when teams begin testing whether a treatment helps patients. Cancer programs were among the least likely to make it through development.

Two people can have the same cancer diagnosis and very different tumors. Mutations, immune cells, earlier treatments and other illnesses can all change what a drug does. A medicine that kills tumor cells in a dish may struggle in a person. This is the gap that AI drug discovery companies are trying to close. Cancer drugs fail for many reasons. In our AI drug discovery reality check, we look at where they fail most.

Are cancer drug approval rates getting better?

Not much. Recent estimates still put a cancer drug's chance of reaching approval from Phase I at about one in twenty.

BIO's study of 2011-2020 programs estimated 5.3% for oncology, against 7.9% across all therapeutic areas. A Nature Communications analysis published in October 2025 followed 20,398 development programs. Its adjusted cancer approval estimate edged down from 6.0% for the 2011-2019 window to 5.7% for 2015-2023. The datasets use different methods, so we compare their results without stitching them into one trend line.

Some treatments do better. BIO estimated 12.4% for immuno-oncology programs and 15.9% for programs using patient-selection biomarkers across all diseases. Those are historical subsets, and the biomarker result covers more than cancer. Better patient selection may help, but we cannot credit that whole difference to AI.

The table below shows the historical approval rates and the groups each study measured.

Study and window Group measured Phase I to approval
BIO, 2011-2020 All therapeutic areas 7.9%
BIO, 2011-2020 Oncology 5.3%
BIO, 2011-2020 Non-oncology 9.3%
BIO, 2011-2020 Immuno-oncology 12.4%
BIO, 2011-2020 Biomarker-selected programs, all areas 15.9%
Nature Communications, 2011-2019 Cancer programs, adjusted 6.0%
Nature Communications, 2013-2021 Cancer programs, adjusted 6.0%
Nature Communications, 2014-2022 Cancer programs, adjusted 5.8%
Nature Communications, 2015-2023 Cancer programs, adjusted 5.7%

Different datasets and estimation methods; these are not a single continuous series.

How often do cancer drugs fail in clinical trials? Almost always. Only about 5 in 100 cancer drugs that start trials get approved, against 8 for all drugs. Three in four stop in phase 2. Chart from NewMarketPitch’s AI Drug Discovery report.
AI Drug Discovery reportThis page and 70+ morePDF Get the report

Just a preview. If a number here looks old, no worries: your report comes with the latest data.

Is Orakl Oncology's 90% accuracy claim trustworthy?

Not yet. Orakl Oncology's headline cancer-trial accuracy is still a company claim, while its public SCOPE study reports narrower performance results.

Orakl's website says its engine achieves the headline figure across more than 30 clinical trials. Norrsken repeated the claim for trial outcomes in its November investment announcement. Orakl used 89% in a July announcement about four Sanofi cancer-drug assets. Those statements do not all define accuracy the same way, and the company has not published a full independent breakdown for the headline figure.

SCOPE, the public April preprint, reports errors in months when forecasting disease progression, differences in percentage points for tumor response rates and seven correct treatment rankings in eight comparisons. Its evaluation covers 23 published trials, fewer than the 'more than 30' on the company website. The paper also says independent checks of its individual-patient predictions are still underway. Before relying on the claim, we would want the exact patients, treatments, endpoints and wrong predictions behind it.

How close were Orakl's cancer-trial predictions?

According to its preprint, SCOPE came close to several published cancer-trial results, although Orakl tested only 23 earlier trials.

The Orakl and Gustave Roussy team developed the model using 54 treatment lines from 52 patients, alongside 81 tumor-organoid lines screened against nine compounds. It then compared predictions with 32 treatment arms from colorectal and pancreatic cancer trials. The authors report that, across 31 arms with usable data, the predicted median time before disease progression missed the published figure by 0.82 months on average. They put the average error for tumor response rate at 7.3 percentage points across 32 arms.

The research is a preprint, and 88% of the evaluated arms involved conventional chemotherapy. These are forecasts for average trial outcomes, so the average error should never be read as a personal survival prediction. Averages like these are easy to oversell, so we put AI claims next to published proof in our report on AI drug discovery.

The table below shows SCOPE's tested sample and reported results.

SCOPE test Reported figure What it covers
Development patients 52 Model training data
Treatment lines 54 Development observations
Tumor organoid lines 81 Laboratory screening bank
Compounds screened 9 Drugs in bank
Published trials 23 External comparisons
Treatment arms 32 Trial-level validation
Progression-free survival error 0.82 months 31 arms
Tumor response rate error 7.3 percentage points 32 arms
Treatment rankings correct 7 of 8 Head-to-head comparisons
Chemotherapy share 88% Treatment arms in validation

Results come from the April 2026 SCOPE preprint, not a completed independent prospective trial.

Is Orakl's seven-out-of-eight cancer-drug result convincing?

Only partly. Seven correct picks in eight cancer-drug comparisons is a promising result for Orakl, but the sample is still too small to trust for major trial decisions.

SCOPE compared pairs of treatments in eight situations. It chose the arm with the better reported outcome seven times. But only seven of those comparisons had a statistically confirmed superior treatment, and the model got six of those right. In the remaining comparison, researchers treated the arm with the numerically longer time before progression as the winner even though the trial had not established a clear difference.

With so few comparisons, a single extra comparison that went wrong would move the score from seven in eight to seven in nine. The set also focused on familiar treatments and two cancer types. Drugmakers would learn far more from dozens of predictions recorded before independent trials begin, especially for new medicines that work in unfamiliar ways.

Do AI-found drugs succeed more often in trials? Not proven. About 8 in 10 pass phase 1, but in a 2026 review of 129 drugs only 1 in 5 passed phase 2. Chart from NewMarketPitch’s AI Drug Discovery report.
AI Drug Discovery reportThis page and 70+ morePDF Get the report

This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.

What happened to the pancreatic cancer drug Orakl predicted would win?

Daraxonrasib went on to win FDA approval after beating chemotherapy in a major pancreatic cancer trial, matching Orakl's forecast about which treatment would perform better.

The April SCOPE preprint predicted that daraxonrasib, also called RMC-6236, would outperform chemotherapy in simulated pancreatic cancer comparisons. On August 26, 2026, the FDA approved Revolution Medicines' drug for certain adults with metastatic pancreatic adenocarcinoma. The decision followed RASolute 302, a randomized trial involving 500 patients.

Patients receiving daraxonrasib had median progression-free survival of 7.2 months versus 3.6 months with the control chemotherapies. Median overall survival reached 13.2 months against 6.7 months, and tumors shrank in 30% versus 11% of patients. Those are unusually clear real-world results in a hard-to-treat cancer. Orakl had picked the winning drug, even though the trial's patient mix and comparison treatments differed from its simulation.

Did Orakl predict daraxonrasib's trial results accurately?

Orakl got the better pancreatic cancer treatment right, but its predicted daraxonrasib numbers cannot be compared directly with the later trial's figures.

SCOPE simulated a second-line comparison in patients with KRAS G12-mutated pancreatic cancer. It predicted 8.9 months before disease progression on daraxonrasib and 5.1 months on FOLFIRINOX chemotherapy, with tumor responses of 29% and 13%. The later RASolute 302 trial used a broader overall population and allowed doctors to choose among several chemotherapy regimens. Its corresponding observed figures were 7.2 and 3.6 months, with responses of 30% and 11%.

The two comparisons used different patients and different control treatments. SCOPE also published a first-line simulation that was unrelated to the randomized study. We can credit Orakl for identifying the stronger option, but calculating a precise prediction error by subtracting these unmatched outcomes would give a misleading result. Naming the stronger drug is a good sign, but do AI-found drugs beat the usual trial odds? We ask in our AI drug discovery reality check.

The table below shows the simulated and observed outcomes, with the comparison groups kept separate.

Comparison Treatment or outcome Reported value
SCOPE simulated second line Daraxonrasib predicted median PFS 8.9 months
SCOPE simulated second line FOLFIRINOX predicted median PFS 5.1 months
SCOPE simulated second line Daraxonrasib predicted response rate 29%
SCOPE simulated second line FOLFIRINOX predicted response rate 13%
FDA RASolute 302, overall Daraxonrasib observed median PFS 7.2 months
FDA RASolute 302, overall Choice of chemotherapy observed median PFS 3.6 months
FDA RASolute 302, overall Daraxonrasib observed response rate 30%
FDA RASolute 302, overall Choice of chemotherapy observed response rate 11%
FDA RASolute 302, overall Daraxonrasib observed median overall survival 13.2 months
FDA RASolute 302, overall Choice of chemotherapy observed median overall survival 6.7 months

The simulated and randomized populations, control regimens and evaluation settings differ.

Does testing living tumors make AI cancer predictions better?

Probably, on early evidence. Testing medicines on living tumor samples can make cancer-treatment forecasts more useful than medical records alone.

Orakl grows patient-derived organoids, tiny three-dimensional tumor cultures, and watches how they respond to candidate medicines. SCOPE combines those experiments with patients' medical information. In the preprint, the authors report that the combined approach explained 85% of the variation in reported median progression-free survival, compared with 58% for the clinical score alone and 36% for organoid drug-screen results alone. For tumor response rates, the respective figures were 71%, 50% and 36%.

The two sources of information answer different questions. Medical history helps estimate how sick a patient is; exposing their tumor to a drug tests the treatment itself. A separate prospective colorectal cancer study reported discrimination scores of 0.78 to 0.88 for predicting responses to tested chemotherapies using organoids. Both studies are encouraging, though the samples, prediction tasks and measures differ.

Which AI-discovered drugs are furthest along in human trials? Insilico’s rentosertib and Generate’s GB-0895 reached phase 3. None is approved, and Recursion and Iambic plan phase 3 trials in 2027. Chart from NewMarketPitch’s AI Drug Discovery report.
AI Drug Discovery reportThis page and 70+ morePDF Get the report

Preview only. Some numbers may have changed since we made this page, but your report will be up to date.

Can doctors grow and test tumor samples from most cancer patients?

Not yet. Preparing a usable living tumor culture still fails too often for AI cancer-drug testing to reach most patients.

A prospective study of metastatic colorectal cancer enrolled 232 people and obtained 205 biopsies. Its overall organoid establishment rate was 52%, though the researchers improved from 22% to 75% as they refined the process. The interim chemotherapy-response analysis included 42 screened cultures. We should count those completed tests, not the full enrollment, when judging how widely the method could already be used.

The experiments remain small even with newer equipment. In a November 2026 organ-on-a-chip pilot involving 14 locally advanced colorectal cancer patients, researchers built eight patient-specific devices. Cultures may fail when samples contain too few living tumor cells, grow too slowly or lose important features of the original cancer. That practical hurdle comes before the AI model even makes a prediction.

Can AI tell which patients will respond to immunotherapy?

AI is starting to identify likely immunotherapy responders across several cancers, although the strongest current evidence still comes from existing patient records.

COMPASS, published in Nature Medicine in 2026, reads gene-activity patterns in tumors. Researchers trained it on 10,184 tumors from 33 cancer types and evaluated it on 1,133 patients in 16 clinical cohorts, spanning seven cancers and six immunotherapy treatments. Against 22 comparison methods, the authors report that it improved average response-prediction accuracy by 8.5% and the area under the precision-recall curve by 15.7%. Those are improvements in model performance, not observed increases in patients benefiting from treatment.

COMPASS tackles a different kind of question from Orakl's SCOPE, whose public trial-level tests mainly involved chemotherapy. SCORPIO takes yet another route using blood tests and medical information. Looking across them, we find useful ways of sorting patients, but they should not be combined into a single league table of 'AI accuracy.' Sorting patients is useful. Whether AI can really pick the cancer drugs that work is the bigger question our report on AI drug discovery takes on.

The table below shows what the main cancer-drug prediction studies actually tested.

Model or study Evidence tested Actual prediction task
Orakl SCOPE 31 trial arms Average progression-free survival
Orakl SCOPE 32 trial arms Average tumor response rate
Orakl SCOPE 8 trial comparisons Better treatment arm
COMPASS 1,133 patients Immunotherapy response
SCORPIO 9,745 treated patients overall Outcomes after checkpoint inhibitors
Prostate biopsy classifier Two Phase III trial cohorts Groups benefiting from docetaxel
Colorectal organoid study 42 screened cultures in interim analysis Chemotherapy response
Ono/Hitachi approval model 44 later trial outcomes Drug approval or failure
TrialTranslator 11 landmark trials Applicability of results to real patients

Different populations, outcomes and metrics; the rows are not a model leaderboard.

Can ordinary blood tests predict how cancer immunotherapy will work?

Routine blood tests already help AI estimate outcomes after cancer immunotherapy, even without a complex genetic assay.

SCORPIO, published by Memorial Sloan Kettering researchers and collaborators in Nature Medicine in January 2025, analyzed 9,745 immunotherapy-treated patients across 21 cancer types. The data came from medical centers and ten global Phase III trial cohorts. The model reads standard blood counts and chemistry panels together with basic clinical information.

In two internal test sets, its survival-prediction discrimination scores were 0.763 and 0.759. Tumor mutational burden, an established molecular measure, scored 0.503 and 0.543 in the same comparison. SCORPIO was better at ranking patients by later outcome. The next question for doctors is whether it can identify patients who would do better on immunotherapy than on an alternative, rather than simply predicting survival for people already taking it.

How much do drug giants really pay AI drug discovery firms? Up front, $115M at most, like Lilly’s deal with Insilico. The billions announced come only if the drugs succeed. Chart from NewMarketPitch’s AI Drug Discovery report.
AI Drug Discovery reportThis page and 70+ morePDF Get the report

This is just a preview page. Don’t worry if a number looks old: the report gets refreshed, and you’ll always get the latest version.

Can AI read a biopsy to find who benefits from chemotherapy?

An AI model reading prostate cancer biopsy slides found patient groups with different survival gains from docetaxel chemotherapy.

A Clinical Cancer Research study published in February 2026 examined stored tissue from two randomized Phase III trials, CHAARTED and NRG/RTOG 0521. In the CHAARTED group predicted to benefit, the hazard ratio for death with docetaxel was 0.52. In the group predicted not to benefit, it was 1.31, though that second estimate had wide uncertainty. The study also reported treatment-by-model interactions in both trials.

The researchers analyzed slides from 286 CHAARTED patients and 350 NRG/RTOG 0521 patients. Because those patients had been randomly assigned treatments, the team could study differences in the treatment's effect, which is a harder problem than predicting who has severe cancer. Doctors would still need a forward-looking test before using the classifier to choose today's prostate cancer treatment combinations.

Can AI predict whether a cancer drug will get FDA approval?

AI can sort oncology programs by approval risk, and one published model picked all three approvals in a later test of real outcomes.

Researchers at Ono Pharmaceutical and Hitachi published an oncology approval model in Scientific Reports in December 2025. Using information on drug development and trial design, they reported AUC scores from 0.86 to 0.98 in different tests. For a separate set of 44 later Phase II outcomes, their selected model correctly identified three approvals, classified 40 failures correctly and produced one false alarm.

That gives 43 correct predictions. The model can learn from trial length, recruitment and development history as well as drug-related information. So it is answering whether a development program looks likely to succeed, a different question from SCOPE's forecasts of tumor response or months without progression. With only three approvals in the test set, one additional wrong call would change its precision considerably. To see which AI drugs are closest to approval, and how far the rest still have to go, we turn to our AI drug discovery reality check.

The table below shows every outcome, including the simple baseline of predicting failure for every program.

Measure Result Interpretation
Trials evaluated 44 Later observed outcomes
Programs approved 3 Actual positive cases
Programs failing 41 Actual negative cases
Correct approval predictions 3 True positives
Incorrect approval predictions 1 False positive
Correct failure predictions 40 True negatives
Missed approvals 0 False negatives
Total correct predictions 43 of 44 97.7% accuracy
Always-predict-failure baseline 41 of 44 93.2% accuracy
Precision for predicted approvals 3 of 4 75%

The sample contains only three actual approvals, so precision is especially uncertain.

Why can cancer-drug prediction accuracy be so misleading?

A cancer-drug prediction model can look 95% accurate while missing every successful medicine, simply because failures are so common.

In BIO's historical oncology dataset, roughly five of every 100 programs entering Phase I reached approval. Imagine a model predicting that all 100 will fail. It would appear correct about 95 times, yet it would never identify a drug worth developing. That is a thought experiment based on the historical approval rate, not a test of Orakl's SCOPE system, which mostly forecasts trial efficacy.

The Ono-Hitachi approval model gives a real example of why the baseline matters. It got all but one outcome right, while a model that always predicted failure would have missed only the three approvals. The added value was catching three actual approvals with one false alarm. For any accuracy claim, we would want to see how many successes the model caught, how many it missed, and whether it locked its predictions before the outcome became known.

Which AI drug discovery companies are worth the most? Isomorphic Labs, by far, at about $50B reported, with no drug in human trials yet. Insilico, with one in phase 3, trades at $2.5B. Chart from NewMarketPitch’s AI Drug Discovery report.
AI Drug Discovery reportThis page and 70+ morePDF Get the report

Just a preview. If a number here looks old, no worries: your report comes with the latest data.

Will AI cancer-drug models work on entirely new treatments?

We still cannot count on AI cancer-drug models to perform as well on unfamiliar medicines as they do on previously studied treatments.

SCOPE's published trial-level test centered on conventional chemotherapy, although it also simulated daraxonrasib. COMPASS drew on multiple cancers and immunotherapies, but its reported evaluation did not compare treatment choices made by doctors using the model with usual care. SCORPIO predicted outcomes for people who had already received checkpoint inhibitors. Each is useful on its own terms; none supplies a universal test across new chemotherapy, targeted and immune-based drugs.

Orakl wants to move into lung and gastric cancer, requiring more samples and experiments beyond its initial gastrointestinal tumors. New treatment combinations can also weaken models trained on older care. A 2026 Nature Reviews Clinical Oncology assessment identified limited testing on future patients, changes in clinical data and uneven representation as serious challenges for AI simulations of cancer trials.

Are major drug companies already using Orakl's predictions?

Some major drugmakers are trying Orakl's cancer-drug predictions, but its reported commercial pipeline is much bigger than its disclosed customer list.

Orakl announced agreements involving four Sanofi oncology assets in July 2026. It said its O-Predict platform would evaluate likely clinical responses and help choose development paths. Whether those forecasts changed a drug's progress, trial design or eventual outcome has not yet been made public.

Norrsken's November announcement says 18 of the world's 20 largest pharmaceutical companies are in Orakl's commercial pipeline, with some early engagements becoming multiyear collaborations. A pipeline includes potential customers, so we would not call those 18 companies paying clients. The investor also mentioned more than 15 cancer centers in contract negotiations. The named Sanofi work is the clearer evidence of how predictions are entering pharmaceutical decisions. A pipeline is not a paying client, so we check whether valuations are running ahead of clinical results in our report on AI drug discovery.

Could AI stop companies spending billions on failed cancer trials?

AI could cut wasted cancer-trial spending, but savings will depend on which development decisions companies change after seeing its forecasts.

A drugmaker can save years of work if it drops a weak medicine before a costly late-stage trial. It might also redesign a study to focus on patients most likely to benefit. Orakl's work on Sanofi assets is an early example of that possibility. To measure actual savings, though, we would need a record of forecasts, changed decisions, trials avoided and subsequent outcomes; the company's model-error figures alone cannot produce a dollar total.

There are useful warnings from AI tools that target a different trial problem. A 2026 prospective study examined AI-assisted matching for 3,804 cancer patients, while a randomized JAMA Network Open study in 2025 found that AI-triggered notifications to doctors did not increase trial enrollment at one cancer center. Software can do an intermediate job well and still leave the final real-world result unchanged.

Will the FDA accept AI predictions instead of cancer trials?

No. FDA approval of cancer drugs still rests on clinical evidence of safety and benefit, even when AI helps researchers decide what to test.

The FDA issued draft guidance in January 2025 on using AI-generated information to support regulatory decisions about drugs and biological products. It called for models to be evaluated against a clearly defined purpose and level of risk. The document does not authorize companies to bypass randomized evidence of how a medicine affects patients.

Daraxonrasib is a concrete example. Orakl had forecast a positive result, but the agency approved the drug after examining randomized RASolute 302 findings on survival, disease progression, tumor shrinkage and safety in 500 patients. AI can help guide what sponsors bring forward; the FDA still needs data about what happened when people received the drug.

What went wrong for AI drug discovery’s early stars? Recursion stopped two phase 2 drugs and cut about 20% of staff in 2025. BenevolentAI left the stock market and Exscientia was absorbed by Recursion. Chart from NewMarketPitch’s AI Drug Discovery report.
AI Drug Discovery reportThis page and 70+ morePDF Get the report

This is a preview page. Spotted an outdated number? That’s normal here. The report you get is always the freshest version.

What would make AI cancer-drug forecasts dependable?

AI cancer-drug forecasts will earn trust when independent teams can check predictions made before trials report their results.

Developers should say exactly what they forecast: a patient's response, the average effect in a trial or eventual approval. Then they should lock the model version, predicted numbers, patient group and date before the actual results arrive. Independent researchers could test the system on medicines and hospitals left out of its development data, publishing every prediction, including the missed winners and false alarms.

For a tool meant to guide treatment, doctors would need a trial comparing AI-assisted decisions with standard care, including survival and side effects. For drug companies, the equivalent would track go/no-go decisions and what happened to the programs they changed. A 2026 review of AI in oncology trials called for exactly this kind of forward-looking testing with human oversight. Forward-looking tests with human oversight are one step. What must happen before AI drugs reach patients is covered in our AI drug discovery reality check.

The table below shows the checks we would want before trusting a major accuracy claim.

Test What should be disclosed Why it matters
Exact prediction Drug, patient group and endpoint Defines what success means
Prediction timestamp Date and locked model version Prevents hindsight
Full test set Every included trial or patient Limits selective reporting
Independent sites Hospitals outside development Tests generalization
New medicines Previously unseen drug mechanisms Tests novel-drug use
Negative cases False positives and false negatives Reveals costly mistakes
Calibration Predicted probability versus results Checks usable risk estimates
Standard comparator Clinical practice or simple baseline Measures added value
Decision impact Trials changed and later outcomes Connects forecasts to savings
Patient impact Survival, toxicity and quality of life Tests actual clinical benefit

Can AI really predict which cancer drugs will work?

Partly yes. AI can now predict some cancer-treatment outcomes quite well, but it still falls short of reliably picking successful new drugs across the board.

SCOPE deserves attention for its reported predictions of real treatment outcomes using living tumor samples, and its daraxonrasib forecast picked the treatment that later beat chemotherapy. Yet the main published test involved a small set of earlier gastrointestinal cancer trials, mostly with familiar medicines. Its results are promising within that range.

The wider evidence is encouraging in other, narrower jobs. COMPASS sorts likely immunotherapy responders using tumor gene activity, SCORPIO estimates outcomes using ordinary blood work, and a prostate cancer biopsy model picks out groups with different chemotherapy benefit. An approval-prediction study identified all three approved programs in a small later test. No one figure covers all these tasks.

Our verdict is to use AI to choose better experiments and spot promising patient groups now. We would let it make more consequential decisions only after independent researchers have tested predictions recorded in advance and shown that following them improves clinical care or drug development.

Who has the most AI-found drugs in human trials? Insilico Medicine, with 10 programs in trials and one in phase 3. Isomorphic Labs raised $5B in November 2026 but has no drug in trials yet. Chart from NewMarketPitch’s AI Drug Discovery report.
AI Drug Discovery reportThis page and 70+ morePDF Get the report

Preview only. Some numbers may have changed since we made this page, but your report will be up to date.

Our methodology

We reviewed information available through November 18, 2026. We started with Orakl Oncology's company and investor announcements, its July statement about Sanofi assets and the April version of its SCOPE preprint. We kept the company's headline claim separate from the paper's published tests. We compared the paper's retrospective trial-arm predictions, treatment rankings and forward-looking daraxonrasib simulation, while keeping individual-patient forecasting separate from trial averages. A preprint was treated as preliminary research, not as an independent clinical validation.

For the real daraxonrasib outcome, we used the FDA's RASolute 302 approval summary and kept its randomized patient population and chemotherapy controls separate from SCOPE's simulated groups. We compared historical approval rates from BIO and a Nature Communications study without merging their different datasets. We also read original COMPASS, SCORPIO, organoid, pathology and approval-prediction studies, plus FDA guidance and published trial-operation evaluations. We calculated the approval model's accuracy and always-predict-failure baseline from its reported results. We left out unverified drug-development savings, treated Orakl's pipeline figure as prospects rather than customers, and judged each prediction system on the specific outcome it tested.