Aldo Vidinha on AI Drug Discovery: Faster Molecules, Unchanged Attrition

A narrower claim than the rhetoric, a larger one than the critics allow

The argument over whether artificial intelligence has transformed drug discovery has settled into two unhelpful camps. One treats every candidate nomination as a productivity revolution. The other treats the field as computational chemistry with a new label.

Aldo Vidinha, a pharmaceutical engineering, quality and validation specialist with roughly two decades of international experience across pharmaceutical, biotechnology and life science environments, argues that both positions collapse on contact with the evidence.

What AI has credibly demonstrated, in his assessment, is compression of the discovery-to-candidate phase: fewer molecules synthesized, shorter timelines from project start to preclinical candidate. What it has not demonstrated is any effect on the variable that governs pharmaceutical economics, which is clinical attrition.

“AI is a real and increasingly essential tool that has changed how molecules are found, has not yet changed whether they work, and is being priced as though it had.”

The reference case is Insilico Medicine’s rentosertib, a TNIK inhibitor for idiopathic pulmonary fibrosis and the first AI-originated molecule to publish randomized human efficacy data. The Phase IIa results published in Nature Medicine in June 2025 covered 71 patients over 12 weeks, with the 60 mg once-daily arm showing a mean forced vital capacity improvement of 98.4 mL against a decline in the placebo group. The programme entered Phase III in July 2026.

Vidinha’s reading is deliberately restrained. That is one genuine data point, not a validated productivity curve. Set against it are discontinuations that receive far less attention: Recursion’s REC-994 stopped in May 2025 after longer-term data failed to confirm early trends, alongside two further programme cuts; Exscientia’s DSP-1181 halted after Phase I; BenevolentAI’s BEN-2293 failed Phase IIa, with the company later absorbed.

The category error at the center of the market

Before any verdict is possible, Vidinha argues, the sector has to stop treating AI drug discovery as one market. As marketed today, it comprises at least fourteen distinct activities, including target identification, disease modeling, protein structure prediction, molecular generation, virtual screening, lead optimization, toxicity prediction, repurposing, biomarker discovery, patient stratification, trial design, synthetic control arms, model-informed drug development and laboratory automation.

These differ in scientific maturity, evidence standard, regulatory relevance, capital intensity and defensibility. Protein structure prediction has undergone something close to a step change, turning a multi-year experimental campaign into a routine input. Molecular generation is a substantial incremental improvement over established generative and optimization chemistry, faster and more systematic, but operating on the same medicinal chemistry logic. Target identification using multi-omics and perturbation data is genuinely new in scale, yet its output is a hypothesis whose quality is revealed years later. Patient stratification and trial modeling are continuous with pharmacometrics and biostatistics, now better tooled.

Step change in some layers, incremental improvement in others, rebranding in a few. Any analysis that answers “is AI drug discovery real?” with a single verdict is answering the wrong question.

An evidence hierarchy the sector routinely skips

Vidinha’s most transferable contribution is a ten-level hierarchy that makes claim inflation visible. Ascending:

1. Computational benchmark on a curated dataset

2. Retrospective validation, where the model finds known answers in historical data

3. Prospective laboratory prediction, made before the experiment and tested

4. Candidate nomination

5. Preclinical proof in relevant models

6. Regulatory acceptance for clinical study, meaning a cleared IND or equivalent

7. Phase I evidence on human safety and pharmacokinetics

8. Proof of biological or clinical activity

9. Comparative development advantage against a conventional comparator

10. Approved product or demonstrated portfolio productivity

Levels one to four are commonplace across the sector. Level eight exists in a single well-documented case. Levels nine and ten have not been reached by anyone.

Two traps sit inside that ladder. Retrospective validation is treacherous because a model trained on the literature will rediscover the literature. And IND clearance is regulatory acceptance of one molecule’s study plan, never an endorsement of the algorithm that produced it. That distinction, Vidinha notes, is blurred routinely in investor communication.

Accelerating an activity is not improving a decision

Pharmaceutical R&D’s productivity problem was never that molecules are slow to make. It is that roughly nine in ten candidates entering human trials fail, mostly on efficacy and safety that only emerge in humans.

Generating a candidate in 15 months rather than 42 accelerates an activity that occupies a modest share of total development time and cost. If the underlying target is wrong, faster design produces faster failure. Cheaper failure has some value. It is not transformative.

Improving decisions means picking better targets, killing bad programmes earlier on defensible evidence, designing trials that detect real effects, and identifying patients who can respond. That is where the economics live, and it is the harder claim to evidence.

Where the regulators have actually moved

The evidentiary picture is uneven in a way that cuts against the loudest part of the market. The strongest regulatory standing belongs not to generative chemistry but to modeling and simulation, which have supported regulatory submissions for two decades.

ICH M15 on model-informed drug development reached Step 4 on 29 January 2026, establishing a harmonized framework for assessing MIDD evidence. The EMA Step 5 document was published on 9 February 2026 with effect from 23 July 2026, and FDA announced final guidance in the Federal Register on 3 June 2026.

By contrast, AI outputs used to support regulatory decisions sit under FDA’s risk-based credibility framework, which remains draft guidance issued in January 2025 and is organized around context of use rather than around approving algorithms. The FDA and EMA’s ten joint principles of good AI practice, published in January 2026, are explicitly non-binding foundations for future guidance. Further detail sits on FDA’s artificial intelligence for drug development hub.

Vidinha draws a hard line here. MIDD is mechanistic and statistical modeling of drug and disease behavior used to inform regulatory decisions. It is not generative AI, and conflating the two borrows credibility that generative chemistry has not yet earned.

The metrics that mislead

Several figures in circulation are, in Vidinha’s view, structurally misleading rather than merely optimistic.

Timeline claims against soft baselines. “Traditional discovery takes four to five years” is an industry average across heterogeneous programmes, compared against an AI company’s best programmes. The comparison is selection-biased in both directions.

Molecules generated. An input metric presented as an achievement. So are targets screened and patents filed.

Announced partnership value. The most misleading number in the sector. A headline collaboration worth two billion dollars or more typically comprises a modest upfront payment, research funding, and a long tail of contingent milestones and royalties, most of which will never be earned because most programmes fail. “Announced deal value is an option’s notional, not revenue.”

Selection bias across all of it. Successes are announced. Failures are disclosed quietly in quarterly reports, or not at all. There is no public registry of AI-derived programmes discontinued at preclinical stage, which is a real limitation on any assessment, including his own.

Where the defensible value sits

Vidinha’s position is that the strongest opportunity is not de novo molecular generation but the decision-quality layer: target prioritization grounded in proprietary biology, translational and model-informed development, patient stratification, and trial design. Alongside that sits the physical infrastructure of data generation.

The moat question resolves in a similar direction. Public data confers no differentiation. Licensed datasets are rentable by competitors. The defensible assets are internally generated experimental data produced systematically at scale, negative and failed-experiment data that almost nobody publishes and which are disproportionately informative for training, longitudinal patient data linking molecular profiles to outcomes, and multimodal data linked at patient level.

Quality and provenance matter more than volume. A large heterogeneous corpus assembled from inconsistent assays produces confident, unreliable models.

“The feedback loop, not the model, is the asset.”

Then there is the constraint market analyses persistently omit. A computationally optimized molecule can be synthetically intractable at scale, unstable, poorly soluble, difficult to formulate, or dependent on a supply chain that cannot support commercial volumes. CMC and process engineering realities have killed attractive candidates for decades and are indifferent to how the molecule was conceived. Any platform assessment that does not examine synthetic route feasibility, scale-up risk, and formulation strategy is incomplete.

Questions before capital

From a seventeen-point due diligence framework, the questions that most reliably separate platforms from pitches:

• What specific decisions does the platform improve, and who made those decisions before?

• What is the conventional comparator, defined precisely by modality, target class and company?

• Is performance measured prospectively, with predictions registered before experiments?

• Are failed predictions and discontinued programmes disclosed, with reasons?

• Has candidate quality improved, meaning developability, safety margins and CMC feasibility, or only candidate speed?

• What proportion of announced deal value is upfront and research funding versus contingent milestones?

• Are timeline claims measured from the same start point as the comparator, meaning project initiation rather than target selection?

• Which specific result would the company accept as invalidating its platform thesis?

Vidinha treats the last one as diagnostic. “A team that cannot name a falsifying result is describing a belief, not a platform.”

Three futures, and the test the field has agreed on

Looking to 2027 through 2030, Vidinha sketches three scenarios and declines to assign probabilities, on the grounds that the evidence base is too thin and too selection-biased to support numbers.

In the first, multiple AI-originated assets reach Phase II or III proof of concept and attrition in AI-derived portfolios measurably beats matched conventional comparators. Platform premiums return and data access becomes the strategic battleground.

In the second, AI becomes standard infrastructure across discovery and development, as bioinformatics did, while overall attrition stays close to historical norms. Value accrues to whoever owns data and execution rather than to method vendors.

In the third, weak platforms disappear, and value concentrates in data-rich pharmaceutical, health system and technology organizations. The Recursion and Exscientia combination, BenevolentAI’s absorption, pipeline pruning and severe valuation compression across the listed cohort all point this way.

His reading of current evidence is that the second and third scenarios are visibly underway, while the first remains open and will be settled by clinical readouts between 2027 and 2030.

The single development that would help most is unglamorous: systematic publication of negative results and discontinued programmes. Its absence is why nobody, including Vidinha, can yet compute the sector’s actual hit rate.

About Aldo Vidinha

Aldo Vidinha is a pharmaceutical engineering, quality and validation specialist with approximately 20 years of international experience across pharmaceutical, biotechnology, medical device and life science environments. His work spans Commissioning, Qualification and Validation, process and cleaning validation, Computer System Validation, Computer Software Assurance, pharmaceutical facilities and utilities, quality systems, regulatory compliance, and the implementation of artificial intelligence in regulated GxP processes. He has supported multidisciplinary programmes across Europe, the United States and Australia throughout the design, construction, qualification, validation and operational lifecycle of regulated facilities and systems.

Published originally on — https://marketsherald.com/aldo-vidinha-on-ai-drug-discovery-faster-molecules-unchanged-attrition/