CEO and Founder Dr. Lurong Pan calls on the AI pharmaceutical industry to set boundaries for the algorithm, pairing executable safety rules with an honest, two-dimensional method for measuring what a model actually contributes
Ainnocence Inc., an AI-driven drug discovery company working across antibodies, GPCR targets, cyclic peptides, PROTACs, and drug-dye conjugates, today released an industry perspective from Founder and CEO Dr. Lurong Pan proposing a dedicated governance framework for artificial intelligence in pharmaceutical R&D. The framework couples seven executable rules designed to be written directly into corporate SOPs, collaboration agreements, and eventually regulatory practice with a two-dimensional "value scorecard" for evaluating a model's real contribution.
The position addresses what Dr. Pan describes as a regulatory vacuum: over the past decade, AI has moved from a peripheral tool to the core of the drug discovery pipeline driving protein structure prediction, generative molecular design, de novo design of antibodies and cyclic peptides, and patient stratification and endpoint prediction in clinical trials. Yet neither existing rulebook fully applies. Traditional drug regulation was built for wet-lab and clinical evidence and cannot assess how trustworthy a model-generated molecular hypothesis is, while general AI governance frameworks focus on bias, privacy, and transparency without setting an adequate safety margin for an erroneous prediction that may enter human trials.
"A fast-moving industry without rules will, sooner or later, write its rules for itself through one or two costly failures," said Dr. Lurong Pan, Founder and CEO of Ainnocence. "Rather than passively await that moment, those of us on the front line who understand both algorithms and drugs have a responsibility to put forward a framework that is executable and acceptable to regulators and investors alike."
Why AI Pharma Needs Its Own Rules
Drawing on Ainnocence's years of cross-border R&D and collaboration, the perspective identifies three features that distinguish AI pharma from general AI applications and make a dedicated set of rules necessary:
· The cost of error is asymmetric and delayed. When a recommender system suggests the wrong product, the consequence is immediate and reversible. When an AI model recommends the wrong lead compound, the consequence may not surface until the clinical stage two or three years later, after tens of millions of dollars and substantial patient opportunity cost have been consumed. Because the error is separated from the decision by years, the industry cannot rely on correcting, after the fact; it must rely on norms set beforehand.
· A model's "knowledge" comes from biased historical data. Public chemical and biological data are systematically skewed toward well-studied targets, druggable chemical space, and clinical data from particular populations. A model trained on such data makes confident predictions where it has coverage while quietly losing accuracy on rare-disease targets, novel scaffolds, and underrepresented populations, precisely the domains where AI ought to deliver its greatest value.
· The value chain is long, and responsibility diffuses easily. AI pharma projects span data providers, algorithm platforms, CROs, clinical teams, and regulatory filings, frequently across national borders. When a model-driven decision goes wrong, "whose responsibility is this" is easily diluted into "no single party's responsibility." A primary role of rules is to nail responsibility back onto every link.
The Seven Proposed Industry Rules
The seven rules are arranged along the lifecycle of an AI prediction, from generation to entering the clinic. They are framed not as abstract ethical declarations but as executable provisions:
1. Computation must serve TPP-driven multi-objective optimization, not single-point breakthroughs or blind screening. A real-world drug is never the champion of any single metric; it is the optimal balance among conflicting objectives: activity, selectivity, druggability, safety, metabolic stability, synthesizability, and cost. Before committing to compute, every project should explicitly define its Target Product Profile (TPP) and organize computation around it as a targeted, multi-objective optimization. AI's irreplaceable value lies in performing, within a high-dimensional and mutually constraining objective space, the collaborative optimization that human effort cannot exhaust.
2. Mandatory disclosure of model cards and data provenance. Any model used for critical decisions should carry a standardized "model card" much like a drug's package insert stating, at minimum: the source and time range of the training data; the chemical or biological space known to be weakly covered; the model's intended and explicitly non-intended uses; and the performance range on independent test sets. This turns a model's domain of applicability from the developer's tacit knowledge into explicit, verifiable information, and is the lowest-cost way to build technical trust in cross-border collaboration.
3. Predictions must carry uncertainty and distinguish interpolation from extrapolation. A model must not only output a prediction but also answer how certain it is and indicate whether the prediction falls within the interpolation region of its training data or extrapolates into space it has not adequately learned. High-confidence interpolation predictions let teams advance quickly; extrapolation predictions, however tempting the numbers, should be treated as hypotheses requiring priority wet-lab validation. The rule installs a brake against over-trusting the model.
4. Wet-lab closed-loop validation cannot be replaced by algorithms. AI can dramatically compress the number of candidates entering the wet lab but cannot eliminate the wet lab step itself. Any AI-generated candidate advancing toward animal or human studies must undergo independent experimental validation, and results of success or failure must flow back into the model for retraining. Failed validation data should be treated as an asset of equal importance to successful data; negative results must be recorded, structured, and used for iteration, not discarded.
5. Predictions of druggability and safety should be held to a higher standard than activity predictions. The industry carries a natural bias toward "positive" metrics such as activity and affinity, while "veto" metrics such as toxicity, off-target effects, and druggability are relatively neglected — yet the latter are leading causes of failure. Safety-related predictions should require more conservative confidencethresholds and stricter validation. When a model flags a cardiotoxicity risk, even at low confidence it should trigger mandatory experimental investigation: better to wrongly kill a candidate than to let a risk through.
6. The chain of responsibility must be traceable to specific links. Every critical decision in which AI participates should leave a record of which model version was used, on which batch of data, reviewed by whom and what judgment was reached. Across multiple entities and borders, this decision log is the only reliable basis for post-hoc review and the assignment of responsibility. A clear chain of responsibility protects honest practitioners by making due diligence a provable fact rather than an unverifiable excuse.
7. Fairness requirements toward underrepresented populations and rare diseases. If AI pharma merely serves already well-studied targets and populations faster, it will amplify rather than narrow existing health inequities. The rules should encourage and where possible require disclosure of a model's performance differences across populations and rare disease domains and make narrowing those differences a shared industry goal. AI's true value in data-scarce domains lies in its potential to make rare-disease research economically viable.
A Two-Dimensional Value Scorecard
The perspective argues that rules addressing what must not be done wrong require a matching, fair method of valuation, otherwise compliance becomes pure cost while genuinely valuable models receive no recognition. Ainnocence proposes evaluating a drug-discovery model beyond a single technical metric such as accuracy, across two dimensions.
Dimension One: Did it produce better molecules? A "better molecule" can be decomposed into measurable, comparable contributions:
· Endpoint quality, not intermediate metrics. Value must be tested by the performance of advanced molecules in the wet lab and clinic. Evaluation should anchor on endpoint metrics such as druggability, selectivity, and safety window, not merely the correlation coefficient of an affinity prediction.
· Increment over baseline. The right question is how much better the model's molecules are than not using a model, or than the previous generation of methods. This requires retaining an honest control baseline for traditional high-throughput screening, expert-designed candidates, or an existing model and quantifying differences in hit rate and hit-to-lead conversion.
· Novelty and patentable space. A truly contributing model produces scaffolds that are structurally novel yet still reasonably druggable. Novelty should be quantified (e.g., structural distance from known compound libraries) and evaluated jointly with druggability, since novelty without druggability is meaningless.
· Cycle compression. With molecular quality held equal, compressing the time from target to a credible lead compound from years to months is itself enormous value. Time is the most expensive resource in drug R&D, and cycle shortening should be explicitly counted as a contribution.
· Early exposure to failure. Accurately vetoing doomed candidates early contributes no less than producing good molecules. The detours avoided and wet-lab work saved are equally valued; an evaluation system should reward eliminating the bad as well as finding the good.
Taken together, a model's molecular contribution can be stated in a single verifiable sentence: while preserving an honest baseline, it produced within a shorter cycle and at a higher hit rate candidates that are more structurally novel and superior on endpoint metrics, while eliminating more doomed directions ahead of time.
Dimension Two: Energy efficiency and computational cost. As model scale balloons, the compute, electricity, and water consumed in training and inference are no longer negligible footnotes. Energy efficiency should be evaluated across several levels:
· Energy consumed per unit of scientific output to the compute and energy required, on average, to produce one experimentally validated, credible lead compound tying energy to scientific output and rebutting the fallacy that a bigger model must be better.
· Training and inference cost accounted separately, a model that is expensive to train but cheap to run suits entirely different scenarios than one that is cheap to train but heavy on every call; the two should be disclosed separately.
· Comparison against wet-lab energy consumption, the energy a model consumes to screen out ten thousand molecules versus the energy, reagents, and labor required to actually synthesize and test them. Only once this account is settled is "AI is greener" in an evidence-based conclusion rather than a slogan.
· The marginal curve of efficiency performance gains often comes with nonlinear growth in energy consumption, with the last few percentage points costing a multiple of the compute. Evaluation should help teams judge when adding compute is no longer worthwhile.
Incorporating energy efficiency is not about obstructing progress but about incentivizing more refined, specialized models rather than merely larger ones. In a vertical domain like drug discovery, a small model that deeply understands chemical and biological priors is often more efficient than a generic giant.
The two dimensions together form an honest value scorecard: validated molecular output and cycle compression on one side; energy and computational cost per unit of output on the other. Dividing one by the other yields a model's true return on investment turning value from a vague impression sustained by demos into a fact that investors can scrutinize, partners can verify, and teams can use internal decisions.
Who Sets the Rules, and How They Take Effect
The perspective outlines four types of actors required for implementation:
· Corporate self-discipline first. Most of the seven rules can be written into internal SOPs and collaboration agreements without waiting for legislation; leading companies' proactive adoption forms a de facto standard and provides validated templates for later regulation.
· Industry bodies set consensus standards. The fields of a model card, the reporting format for uncertainty, and the minimum requirements for a decision log are best fixed by industry alliances so that different companies' practices can be mutually recognized and cross-checked.
· Regulators set the floor and the review interface. Regulators need not dictate which models a company uses but should specify what information must accompany AI-involved evidence for a filing to be admissible focusing on the interface rather than internal implementation, holding the safety floor without stifling innovation.
· Investors make compliance a due-diligence dimension. When investors systematically examine whether an AI pharma company has genuine model-governance capability, compliance shifts from a cost item to a valuation item giving the rules their most direct commercial driver.
Dr. Pan emphasized the cross-border dimension: AI pharma is inherently global, with data, compute, target knowledge, and clinical resources distributed across countries. Fragmented, mutually unrecognized rules would raise the cost of cross-border collaboration enough to offset the efficiency the technology brings, making interoperability with international frameworks a long-term issue for every participant.
"In any emerging-technology industry, the arrival of rules is not the opposite of innovation but the mark of its maturity," said Dr. Pan. "Aviation, nuclear power, and traditional pharma itself only truly earned society's trust and only truly unleashed the value of scale after rules had been established. Setting boundaries for the algorithm is, in the end, about making this industry worthy of trust, and being worthy of trust is the only prerequisite for it to endure."
Ainnocence states it is willing to be an early practitioner of the framework and to work with industry peers, regulators, and investment partners to refine it into a widely adopted industry consensus.
About Ainnocence Inc.
Founded in 2021 and headquartered in Mountain View, California, Ainnocence Inc. is a next-generation biotechnology company uses its proprietary generative AI platform to screen up to 10 billion molecules within hours to accelerate drug discovery across antibodies, small molecules, cell therapies, and synthetic biology. Working directly from sequence data, without 3D structural modeling; the platform has been applied across 60+ therapeutic programs. The company partners with leading pharmaceutical companies, academic institutions, and global health organizations to accelerate the discovery of life-saving biologics. For more information, visit www.ainnocence.com.
Media Contact
Dr. Lurong Pan, PhD
Founder & CEO, Ainnocence Inc.
+1 205-249-7424
lurong.pan@ainnocence.com
service@ainnocence.com
Visit us on social media:
LinkedIn
YouTube
This release is based on an industry perspective representing the views of the author and the company; it does not constitute regulatory or legal advice.
