TL;DR: IA is starting to produce discovery candidates that can be tested, not just debated. Google’s Co-Scientist, AlphaFold, and AlphaEvolve show the shift: AI proposes, software ranks, experiments verify. The win isn’t autonomous science yet. It’s faster hypothesis generation with a clear proof trail.
IA and verifiable discovery candidates became impossible to ignore after Google’s Co-Scientist suggested vorinostat for liver fibrosis in a 2025 Advanced Science paper. Big claim. In human liver organoids, the compound reduced TGFβ-induced chromatin changes by 91% and promoted cellular regeneration.
That doesn’t mean the drug is ready for patients. Not even close. It means AI systems are moving from clever text generation toward candidates that scientists can test, reject, repeat, and improve.
We’ve seen the same pattern in business AI. After 50+ projects across fintech, healthtech, e-commerce, and legal workflows, we’ve learned that useful AI is rarely the model alone. It’s the loop around it: retrieval, scoring, human review, audit logs, and measurement.
What does IA mean for verifiable discovery candidates?
IA, used here as artificial intelligence, means systems that generate hypotheses with enough structure for independent testing. In scientific discovery, that may be a drug repurposing candidate, a protein structure prediction, a lab protocol, or a code-generated algorithm whose output can be checked against a known benchmark. The key word is verifiable. Without proof, it’s just a fluent guess.
According to Advanced Science, Google’s Co-Scientist proposed vorinostat as a candidate for liver fibrosis in 2025, and human liver organoid experiments showed a 91% reduction in TGFβ-induced chromatin changes. That is a stronger signal than a chatbot answer because the result moved into wet-lab validation.
Gary Peltz, Professor at Stanford University School of Medicine, states: “Co-Scientist feels like a collaborator.” I like that phrasing because it sets the right bar. Collaborators suggest. They don’t replace evidence.
The catch is simple: verification costs time, money, and attention.
Why are discovery candidates becoming more credible now?
The credibility jump comes from three forces arriving together: larger foundation models, domain-specific databases, and better evaluation loops. AlphaFold made protein structure prediction widely accessible. Agentic systems now plan experiments, search literature, and rank candidate mechanisms. Code agents can also generate algorithms that pass deterministic tests, which gives teams a cleaner proof path than open-ended prose.
According to Nucleic Acids Research, AlphaFold DB had more than 4.5 million users and more than 18,000 proteome file downloads by 2025. According to the AlphaFold Protein Structure Database, the platform offers access to more than 200 million predicted protein structures.
That scale matters. A biotech team can now begin with structural priors, literature search, and candidate ranking before committing scarce lab capacity. But it’s still easy to fool yourself. Models can overfit public data, repeat weak papers, or rank a candidate because the training set made it look familiar.
Fast isn’t the same as true.
How does the AI discovery stack compare with older R&D?
Older R&D depends heavily on expert search, lab screening, and slow feedback cycles. AI-driven discovery adds computational hypothesis generation before experimental work begins. That can make the funnel wider at the top, but it also creates a new burden: teams must track why a candidate was proposed and what evidence would disprove it.
According to McKinsey, AI could double the pace of R&D and unlock up to $500 billion per year in economic value. McKinsey, research team at McKinsey, states: “It can double the pace of R&D.” I’d treat that as a directional claim, not a guarantee.
| R&D mode | Main strength | Main weakness | Best verification method |
|---|---|---|---|
| Traditional screening | Direct experimental evidence | Slow and expensive | Assay replication and blinded review |
| Literature-led discovery | Uses human expertise | Can miss hidden patterns | Evidence grading and citation tracing |
| AI candidate generation | Produces more hypotheses faster | Can rank plausible nonsense | Benchmarks, wet-lab tests, and audit logs |
| Code-agent discovery | Generates testable algorithms | Scope can be narrow | Deterministic tests and production metrics |
When we implemented a RAG chatbot for a fintech client, support tickets dropped 40% in 3 months. Different domain, same lesson. The answer was not “AI knows.” The answer was retrieval, ranking, review, and measurement.
What can businesses learn from scientific verification?
Businesses should copy the scientific habit, not the scientific vocabulary. A candidate can be a drug, but it can also be a pricing rule, fraud signal, support answer, contract clause, or generated marketing brief. If an AI system proposes something, teams need a test that says whether it worked.
According to BCG, GenAI can reduce early drug discovery time by 25% or more through in silico discovery and improvement. According to BCG, end-to-end drug development success still sits near 10%, which shows why experimental verification remains the bottleneck.
Our team of 10+ specialists has worked with LangChain, LangGraph, CrewAI, and Agno in production ML systems for more than 8 years. We’ve learned to ask a blunt question before building: what would prove this candidate is wrong?
For a legal client, we built a document processing pipeline that automated 80% of contract review and saved 120 hours per month. The system still flagged uncertain clauses for lawyers. That limit was intentional. Some risks shouldn’t be hidden behind a confident score.
Five practices that make IA candidates testable
Testable AI candidates need more than a strong model and a clean demo. They need evidence records, rejection criteria, and a path from suggestion to measured outcome. According to Mordor Intelligence, AI in drug discovery was estimated at $2.58 billion in 2025 and could reach $10.29 billion by 2031, with identification and validation of targets representing 28.43% of the 2025 market.
That market growth is useful, but process quality will decide who gets value. After 50+ projects, we’ve learned that teams get better results when each AI output is treated like a candidate, not a final answer. Small shift. Large effect.
1. Define the rejection test first
Before asking an AI system for candidates, define what would eliminate them. In drug discovery, that may be toxicity, weak binding, poor assay replication, or no plausible mechanism. In a company workflow, it may be low precision, legal risk, or poor adoption by users. This keeps teams from defending outputs after the fact.
2. Keep the evidence attached
Every candidate should carry source documents, model version, retrieval traces, scoring rules, and reviewer notes. This sounds boring until something breaks. Then it’s priceless. We’ve had client reviews where the audit trail made the difference between trusting a workflow and pausing deployment.
3. Separate ranking from proof
Ranking helps decide what to test first. It doesn’t prove the candidate is correct. In AI content systems, for example, a high quality score may predict editorial fit, but it won’t prove audience impact. When we implemented an AI-powered content system for a marketing client, output rose 10x while quality scores stayed consistent. Traffic still needed separate measurement.
4. Use humans where judgment is expensive
Human review should sit where errors cost money, reputation, safety, or compliance. That doesn’t mean every output needs approval. It means the review layer should match risk. Low-risk drafts can move quickly. Medical, legal, and financial decisions need tighter gates.
5. Measure drift after launch
A candidate that works in February can fail in August. Data shifts. User behavior changes. Models change too. Set monitoring around precision, false positives, review overrides, user feedback, and business outcomes. Otherwise, a once-good system quietly decays.
What would a verification workflow look like in code?
A practical verification workflow turns AI output into a structured object, checks evidence, applies business rules, and records a decision. The point isn’t to make Python “do science.” It’s to make the candidate review process repeatable, inspectable, and hard to fake.
According to Google Cloud, AlphaEvolve recovered an average of 0.7% of Google’s global compute resources through better scheduling, accelerated a Gemini kernel by 23%, and reduced training time by 1%. That case matters because generated algorithms were judged by measurable performance, not vibes.
Here’s a small pattern we use when prototyping candidate review services:
from dataclasses import dataclass
from typing import list
@dataclass
class Candidate:
name: str
confidence: float
sources: list[str]
test_plan: str
risk_level: str
def verify_candidate(candidate: Candidate) -> dict:
checks = {
"has_sources": len(candidate.sources) >= 2,
"has_test_plan": len(candidate.test_plan.strip()) > 40,
"confidence_in_range": 0.0 <= candidate.confidence <= 1.0,
"human_review_needed": candidate.risk_level in {"medical", "legal", "financial"},
}
approved_for_testing = (
checks["has_sources"]
and checks["has_test_plan"]
and checks["confidence_in_range"]
)
return {
"candidate": candidate.name,
"approved_for_testing": approved_for_testing,
"checks": checks,
}
Crude? Yes. Useful? Also yes. The production version would add provenance, reviewer identity, model version, database snapshots, and rollback rules.
Where does Yaitec fit into this work?
Yaitec helps teams turn promising AI ideas into systems that can be tested in real operations. That often means RAG, agent orchestration, evaluation sets, workflow design, and production monitoring. We work with LangChain, LangGraph, CrewAI, Agno, and custom service layers when off-the-shelf tools don’t fit.
According to Stanford HAI AI Index 2026, private AI investment in the United States reached $285.9 billion in 2025, more than 23 times China’s private AI investment. The same report says generative AI reached 53% population adoption in three years, faster than PCs and the internet.
That speed creates pressure. I don’t recommend rushing straight into autonomous agents for high-risk decisions. Start with one workflow where the candidate can be tested: support answers, lead qualification, document review, research triage, or content scoring.
If you’re exploring this kind of system, contact us. We’ll help map the candidate, the proof step, and the operating model before anyone writes production code.
Conclusion
IA is now good enough to propose candidates that deserve serious testing, but not good enough to remove verification from the loop. The most valuable systems will not be the ones that sound most confident. They’ll be the ones that preserve evidence, expose uncertainty, and make it cheap to reject weak ideas early.
According to Grand View Research, the global AI drug discovery market was estimated at $2.3 billion in 2025 and projected to reach $13.8 billion by 2033, with a 24.8% CAGR. According to Cornelio et al. in The Need for Verification in AI-Driven Scientific Discovery, progress depends on “without scalable and reliable mechanisms for verification.”
That line should stick. AI can widen the search space. It can find patterns humans miss. It can even produce candidates that surprise experts. But science, business, and production software still move forward by proof.
Build the loop. Then trust the results.
Sources
- Stanford — retrieved 2026-09-01
