Publication Metrics

AI Transparency

Executive Summary:

Artificial intelligence (AI) has been heralded as a game-changer in biotechnology – promising faster drug discovery, personalized genomics, smarter diagnostics, and fully automated “self-driving” laboratories. Indeed, early successes like DeepMind’s AlphaFold (which solved the 50-year protein-folding problem) and AI-designed molecules entering clinical trials have fueled high expectations. But many of these promises have been overstated. Real-world results show that while AI is a powerful tool, it has not yet delivered a wholesale revolution in biotech R&D:

• Drug Discovery: AI was expected to cut drug development from a decade to mere months. It has accelerated some steps (e.g. virtual screening), but no new drug developed entirely by AI has reached the market to date, and overall R&D timelines remain long [1] [2].
• Genomics & Precision Medicine: AI can sift genomic data for patterns, but translating these insights into therapies has been slower than hoped. Successes in identifying gene targets haven’t yet yielded a burst of new cures.
• Diagnostics: AI algorithms match or beat doctors at narrow tasks like reading X-rays or pathology slides, and a few have regulatory approval, but they function best as physician assistants, not replacements [3]. Complex medical decision-making still requires human insight.
• Lab Automation: Highly automated “lights-out” labs exist, but widespread adoption is limited. As recently as 2020, nearly 90% of life science experimental protocols still involved manual steps [4]. Most research labs are far from fully autonomous.
  

Why the Gap?

Why the Gap? Early adopters discovered that biological complexity, data quality issues, and integration challenges have constrained AI’s impact. For example, IBM’s Watson Health initiative famously aimed to revolutionize cancer care with AI, but its partnership with MD Anderson was canceled in 2016 after $62 million spent without delivering a usable product [5]. The project revealed that AI struggled with messy, real-world clinical data and the nuance of medical reasoning. Similarly, many AI-driven drug discovery startups found that promising computer-designed molecules still failed in clinical trials – not because the algorithms were broken, but because the underlying data and biological knowledge were insufficient to capture the true complexity of human disease. In short, the first wave of AI in biotech often overpromised results by overlooking essential scientific realities.

kbDNA’s Perspective – Pragmatic Innovation: As an experienced innovator in life sciences, kbDNA has approached AI differently from the start. We believe that AI’s potential is real – but to avoid hype, it must be grounded in domain expertise and robust data. Our strategy hinges on three principles:

1. Data First: We have built proprietary backend databases through extensive manual data mining and custom automation. This curated data foundation (covering genomics, experimental results, protocols, etc.) helps ensure any AI models we use are learning from high-quality, relevant information. By assembling clean, context-rich datasets, we give AI a fighting chance to succeed where others failed. Our data platforms also serve as a benchmark: we can quickly validate or disprove AI-generated predictions by checking them against real-world results in our database. This safeguards against the classic “garbage in, garbage out” problem in AI.
2. Domain Expertise & Human-AI Collaboration: At kbDNA, we treat AI as an augmentation tool for our scientists, not a replacement. Our team of biologists, chemists, and data scientists work together so that machine learning models are developed and interpreted with deep scientific context. For example, if an AI suggests a potential drug candidate, our experts evaluate its practical feasibility (Can it be synthesized? Is it likely safe? Does it align with known biology?). This hybrid approach identifies false positives early and guides the AI towards more meaningful predictions. We’ve found that human intuition and experience are indispensable for steering AI – much like a skilled pilot flying with a high-tech autopilot. This collaboration improves the outcomes and also builds trust in AI across our R&D teams.
3. Rapid Iteration & Validation: We use automation not to create a “closed-loop” robot lab, but to expedite high-quality experimentation in support of AI. Our custom automated workflows – from sample prep to data capture – allow us to test AI-generated hypotheses quickly in the lab and feed results back into our databases. This “lab-in-the-loop” cycle means our AI models are constantly being recalibrated by up-to-date experimental evidence. It also means we can fail fast and learn, preventing weeks of wasted effort. Integrating AI with real-world lab feedback is crucial to convert in silico predictions into true scientific progress.
   

AI Promises vs. Reality in Biotech

AI Promises vs. Reality in Biotech: The table below summarizes several high-profile promises of AI in biotechnology and what has been achieved so far:

As the table indicates, the reality often lags behind the marketing. This doesn’t mean AI isn’t making a difference – rather, its contributions are more incremental and focused than the sweeping claims often suggested. For instance, several major pharma companies report that AI and machine learning are now routinely used to prioritize drug targets and analyze images or genomic data, resulting in measurable efficiency gains in those specific steps. But the overall process of bringing a new therapy to market still requires extensive laboratory work, clinical development, and regulatory approval, just as it always has.

The same pattern applies in other domains. Genomic analytics have been supercharged by AI algorithms that can scan millions of data points to find disease-related genes or predict protein structures (the success of AlphaFold being a prime example of AI’s value to basic science). Yet, identifying a genetic target or protein shape is usually the first step of many; drug discovery and clinical validation remain gating factors. Diagnostics AI has shined in analyzing images and lab tests faster or more consistently than some humans, but integrating those results into patient care calls for physician oversight. And in laboratory operations, automation driven by AI is helping with tasks like high-throughput screening and quality control, but labs must still be tailored and supervised by skilled researchers.

Why the Overpromise? Biotech and pharma are high-stakes fields where enthusiasm runs strong. It’s easy to see why AI has been viewed as a silver bullet: laboratories and R&D teams are overwhelmed with data, costs are skyrocketing, and breakthroughs are hard to find. Amid these pressures, the idea of super-intelligent algorithms unlocking miraculous insights was incredibly attractive – to scientists hoping to accelerate discovery, to executives seeking R&D efficiency, and to investors looking for the next big thing. In the mid-2010s, press releases and conferences often featured bold predictions: AI would “revolutionize” drug pipelines, decode every genome’s secrets, and replace many routine lab tasks. Some startups and large tech firms did little to temper these expectations, occasionally prioritizing marketing over practical results in the rush to claim “firsts” in AI-driven cures.

However, the biology “learning curve” for AI turned out steeper than anticipated. Key challenges include:

• Data Quality & Relevance: Biomedical data can be noisy, incomplete, and siloed. Early AI efforts often used whatever data was available (public databases, literature) which weren’t always representative of real-world conditions. As a result, algorithms made confident predictions that didn’t hold up experimentally. It became apparent that curating high-quality, context-specific data is a non-negotiable prerequisite for successful AI in biotech.
• Complexity of Biology: Diseases and biological systems involve layers of regulation and variability that single datasets or models can’t capture. For example, a machine learning model might flag a gene as a drug target based on statistical correlation, but it takes seasoned biologists to determine if modulating that gene will actually impact a disease without unacceptable side effects. Many AI models excelled in narrow tasks or with simulated data, but stumbled when faced with the full complexity of living systems and patients.
• Integration into Workflows: It’s not enough for an AI to make a prediction – that prediction must fit into existing R&D processes. One reason lab automation and AI tools haven’t spread faster is that labs are bespoke environments. Integrating new AI systems with legacy instruments, IT systems, and team workflows is hard. Early adopters underestimated the practical effort needed to retrain scientists, retool infrastructure, and ensure regulatory compliance for AI systems. The promise of autonomous labs crashed into the reality of operational and cultural inertia.
• Validation Requirements: In life sciences, experimental validation is the gold standard. No matter how “intelligent” an algorithm seems, its outputs must be verified in wet-lab experiments or clinical trials. This takes time and resources. Some AI predictions have proven valuable, while others added little, meaning R&D organizations have had to sift true signals from false leads. Only rigorous validation separates useful AI-driven discoveries from hype, and that vetting process slows down the pace of AI-driven change (appropriately so, given the human stakes).

kbDNA’s Approach – Building a Foundation for Productive AI

kbDNA’s Approach – Building a Foundation for Productive AI: The experiences above resonate strongly with our team at kbDNA. From the outset, we have championed a “data and domain first, AI second” philosophy:

• We invested in building comprehensive knowledge bases. Through painstaking manual data mining of literature and public databases – and by capturing our own experimental data via custom automation pipelines – we assembled high-quality datasets tailored to our R&D focus areas. These proprietary databases of genetic sequences, biomarkers, experimental results, and protocols give us an edge: when we deploy AI models, we can train and test them on reliable, context-rich data that we understand deeply. This minimizes the risk of spurious correlations and maximizes the chance of uncovering real biological insights. It also allows us to measure AI performance against an internal “source of truth.” In one case, our curated gene-editing dataset helped reveal that a predictive model was over-fitting; we retrained it with our data and significantly improved its accuracy.
• We maintain a strong human scientific oversight over AI projects. Every AI-derived result at kbDNA goes through review by experts in the relevant discipline. This combination of AI with human expertise is powerful: the AI can surface patterns or options humans might miss, while our scientists filter and refine those suggestions. For example, if our machine learning system flags 50 possible drug candidates, our chemists and biologists narrow them to the most promising 5 based on known medicinal chemistry principles and biological plausibility. This saves time and resources by focusing experimental validation on the best options. Far from replacing scientists, AI in our organization acts as a force-multiplier for their experience and intuition.
• We iterate and integrate into workflows carefully. Successful innovation in biotech often comes from incremental improvements and integration, not sudden disruption. We introduce AI-driven tools in a way that complements existing processes. A case in point: our automation team developed a robotic screening system that works in tandem with our AI models. The AI picks top candidates, the robotic system tests them in our assays, and the data feeds back to refine the model. By closing this loop, we ensure our AI remains calibrated to real results. This reduces hype because every AI prediction is followed by verification. It also makes our scientists more comfortable and fluent with AI tools, as they see direct, empirical evidence of how the models perform.

The Path Ahead – Realizing AI’s Potential, Responsibly

The Path Ahead – Realizing AI’s Potential, Responsibly: The initial over-hyping of AI in biotech is giving way to a more balanced understanding. Industry-wide, there’s a shift from grandiose claims to pragmatic deployment. Notably, more than three-quarters of life science R&D leaders in a 2025 survey said they plan to use AI in some form by 2027, but the focus now is on specific, high-impact use cases rather than broad disruption [7]. We at kbDNA are enthusiastic about what AI can do – when used correctly:

• Accelerate Research: AI can dramatically speed up data analysis and hypothesis generation. We’ve seen it in our own work, where tasks like genome analysis or high-throughput screening design that used to take weeks can now be done in days. This means scientists have more time to design experiments and interpret results, rather than crunching data.
• Optimize Decision-Making: In drug development, AI algorithms can evaluate vast chemical libraries or patient data to pinpoint leads and patterns that warrant pursuit. This can trim the “front end” of R&D projects – focusing efforts on more promising paths. For instance, AI helps triage hundreds of drug-like molecules to find a few with the ideal properties, which our chemists then refine further. This guided approach increases the odds of success down the line.
• Personalized Insights: With the right data, AI can identify subtle signals in genomic or clinical data that humans might overlook. These insights can guide more personalized medicine – for example, finding a biomarker that predicts which patients will respond to a new therapy. In one project, our machine learning analysis of past experiments revealed a patient subgroup that responded differently to a certain treatment, leading us to design a follow-up study for that specific genetic profile.
• Enhanced Automation & Efficiency: Rather than fully autonomous labs, the real win of AI in automation today is improving reliability and throughput. AI can monitor instrument data in real time to flag anomalies (preventing wasted experiments) and optimize scheduling (so equipment downtime is minimized). This results in more robust, efficient operations – a significant benefit for any research organization..

An Optimistic, Grounded Outlook

An Optimistic, Grounded Outlook: We firmly believe that AI will ultimately help unlock groundbreaking biotech innovations. The key is pairing AI with sound science. As we have learned, success comes from respecting the complexity of biology and the value of expert knowledge:

• Data and Infrastructure: Organizations should invest in data management, integration, and quality control. High-quality data is to AI what good experimental design is to a lab assay – fundamental for meaningful results. At kbDNA, this meant developing specialized data infrastructure and ontologies specific to our research domain. Others in the industry are making similar moves, recognizing that data preparation and governance are essential for AI readiness.
• People and Skills: Training and hiring are as important as technology. The most AI-enabled labs are those where scientists are well-versed in data science, and data scientists are well-versed in biology. Cross-training and collaborative culture will be critical so that AI isn’t a “black box” but rather a familiar part of the scientific toolkit for the whole team. The industry’s push for more AI training programs in life sciences is a promising sign that we are building this capability [7].
• Strategic Focus: Not every problem needs AI. In our experience, applying AI is most fruitful for challenges characterized by large data sets and well-defined objectives (for example, image analysis, genomic pattern-finding, optimizing known processes). Conversely, early exploratory research or situations with very sparse data may be better served by classical approaches. A strategic, selective application of AI ensures we allocate resources where they can truly make a difference.
• In conclusion, the initial hype around AI in biotech is settling into a more sustainable trajectory. The technology is proving to be a powerful accelerator and enhancer of R&D – but not a magic wand. Breakthroughs like AlphaFold show what’s possible, while setbacks like Watson for Oncology remind us of the pitfalls of overpromise. At kbDNA, our journey has reinforced that AI’s real value emerges when we combine cutting-edge algorithms with rich data, expert oversight, and continuous validation. By staying pragmatic yet optimistic, the life sciences community can harness AI to augment human innovation, leading to genuine progress in drug discovery, genomics, and health – even if it’s not as instant or automatic as early headlines suggested. The path to curing disease and advancing biology will always require hard work and smart science. AI won’t replace that, but used wisely, it will undoubtedly help us innovate better and faster than ever before.

Citations

1. Strickland, E. (2019). “How IBM Watson Overpromised and Underdelivered on AI Health Care.” IEEE Spectrum, 2 April 2019. (See: MD Anderson’s $62M Watson for Oncology project cancellation in 2016)
2. Brown, E. “Drugs discovered using artificial intelligence have not hit the market – yet.” Science News DK, 15 Oct 2024. (Summary of JAMA study: 164 AI-linked drugs in development, only one approved as of 2024, with minimal AI role)
3. Strickland, E. (2019). Ibid. (Noting that only a few AI diagnostic tools, mostly for image analysis, have regulatory approval and that AI struggles with complex medical data)
4. Laurent, A. (2025). “The Modern Biotech Lab: A Guide to Automation, AI & Data.” IntuitionLabs (report), 12 Dec 2025. (Citing a 2020 review that ~89% of published bioscience lab protocols were still manual or semi-manual)
5. Strickland, E. (2019). Ibid. (Quote from Dr. Robert Wachter on IBM “marketing first, product second,” highlighting Watson’s overhype and subsequent struggles)
6. MedDataX. “AI-Designed Drugs Achieve 90% Phase I Success Rate, Nearly Doubling Industry Average,” 25 Feb 2026. (Report on ~80–90% Phase I success rates for AI-designed drug candidates vs ~50% historical average)
7. Pistoia Alliance (2025 Survey). “Lab of the Future Survey 2025.” Technology Networks, Oct 13, 2025. (Finding that 77% of life science labs expect to use AI within 2 years; highlights trend of rising adoption and need for skills/training).