“Let’s not be scared of it (AI). Let’s recognize that we’re competing not against some mythical perfect healthcare system but the one we actually have, which doesn’t work very well” – Dr. Wachter.
The featured speakers were Dr. Robert Wachter, MD, Professor and Chair, Department of Medicine, University of California, San Francisco [He will be coming out with his 7th book soon, “A Giant Leap: How AI is Transforming Healthcare and What It Means for the Future”], and Dr. Laurah Turner, PhD, Associate Dean for Artificial Intelligence and Educational Informatics, Assistant Professor, College of Medicine, University of Cincinnati
Dr. Wachter said that the risk of moving too slowly outweighs the risk of moving too quickly when implementing AI in healthcare, emphasizing the need for thoughtful yet bold action.
He emphasized that strategic priorities should focus on “low-risk, high-yield” use cases, such as administrative support (i.e., summarizing charts and drafting prior authorizations) and non-clinical applications (i.e., improving communication and empathy in patient interactions) before moving to high-stakes clinical decisions. With all AI upsides mentioned by Dr Wachter, he strongly cautioned that over-reliance on AI can erode human expertise, as seen in aviation.
Dr. Wachter paraphrased Ethan Mollick’s 4 Rules of Co-Intelligence:
1. Always invite AI to the table
2. Be the human in the loop
3. Treat AI like a person
4. Assume this is the worst AI you’ll ever use.
Dr. Laura Turner discussed four essential pillars as the foundation for developing and deploying AI in medical education.
Pillars:
Transparency: Students and patients should know when AI is involved in their education or care.
Privacy: AI must respect and protect personal data. Learners should have control over how their data is used.
Fairness: AI must work for everyone, not just those well-represented in training data.
Accountability: Developers and institutions must ensure AI prioritizes safety over convenience.
Dr. Turner emphasized that these pillars are critical for ensuring AI enhances medical education without compromising its core human values. She noted that these pillars serve as a guiding framework for her projects at the University of Cincinnati, ensuring alignment with the needs of learners, educators, and patients.
Summary
Drs. Wachter and Turner highlighted AI’s immense potential to transform healthcare and medical education while underscoring the need for careful, strategic implementation to maximize benefits and mitigate risks.
A big thank you to Joey Bernal for including me in the symposium.
Despite $90 billion being spent annually on cancer treatments, only 10% of drugs are approved for sale by the FDA. The costs of cancer treatments are so high that many patients go bankrupt. We have got to do better. At the conference, many speakers showed how AI tools are improving drug discovery and patient recruitment in clinical trials.
Kevin Hacker, UCSF Catalyst Industry Advisor and Co-founder AI Nova Strategy, and Roopa Ramamoorthi, Director, UCSF Catalyst Program
Shout out to UCSF for co-sponsoring the PMWC.
Key Takeaways
1. Real-world data (RWD) and AI are being integrated into healthcare
AI and RWD, like electronic health records (EHR) and wearables (i.e. iWatches), are being leveraged to improve precision medicine, clinical trials, and patient outcomes. EHR often captures only brief snapshots of patient experiences, leaving significant gaps in understanding the patient journey. Phil Johnson from Evidation Health emphasized the importance of directly engaging patients to collect longitudinal data through wearables, surveys, and molecular data. He demonstrated how its platform uses AI to engage participants, enabling real-time data collection and insights. Johnson also highlighted the importance of ensuring participants’ consent, privacy, and fair compensation for data use, which was a recurring theme, particularly in discussions around wearable devices.
2. AI is being used to predict drug efficacy and identify complex biomarkers that correlate with drug response
AI-driven simulations, such as digital twins, are being used to predict drug efficacy, optimize trial designs, and reduce the time and cost of drug development.
Luca Emili from InSilicoTrials showcased the potential of virtual clinical trials to simulate patient populations and optimize trial parameters, significantly accelerating the drug development process. He highlighted a recent project with Merck where digital twins simulated an entire clinical trial in one day, saving years of research time.
Peter Ellman, CEO of Certis Oncology, discussed using predictive AI to match investigational drugs with specific patient tumor profiles. Certis’ AI platform recently received a U.S. patent, identifies complex gene expression biomarkers, and enables personalized treatment strategies, improving patient outcomes and reducing costs.
3. Advanced AI Models and Knowledge Graphs tools identify disease mechanisms, biomarkers, and drug targets
Knowledge graphs** integrated with large language models (LLMs, such as OpenAI’s GPT) create structured, interpretable insights from diverse datasets.
Janusz Dutkowski from Data4Cure and Paul Rejto from Pfizer demonstrated how these technologies enable researchers to contextualize and compare findings across studies, improving decision-making in drug discovery. Pfizer’s partnership with Data4Cure has resulted in a knowledge graph that integrates omics, clinical, and literature data to provide actionable insights for oncology research. Advances in AI models enable new biology applications, such as protein design, cell state modeling, and multi-omics integration.
Nicolo Fusi from Microsoft Research discussed their work on unified models that combine biological and natural language data, predicting that these models will eventually converge into a single, multi-modal system – a digital representation of humans.
AI is used to power adaptive clinical trials and in silico experiments, dynamically updating trial designs based on real-time data and reducing reliance on traditional static methods.
**Knowledge graphs are structured representations of information that organize data into entities and relationships to capture and model real-world knowledge. They enable machines to understand, reason, and derive insights from interconnected data.
4. Quantum Mechanics-Based AI, multi-omics, and single-cell data are being used in precision medicine
Orly Alter, CSO of Prism AI, highlighted their use of quantum mechanics-based AI to derive interpretable and actionable biomarkers from small, noisy datasets. This approach is being applied to predict patient outcomes and accelerate biomarker discovery.
Integrating multi-omics and single-cell data is critical for understanding disease mechanisms and developing personalized therapies. Yasin Senbaboglu from the Chan Zuckerberg Biohub emphasized using zebrafish and in vitro systems to study dynamic cellular processes. She also highlighted the importance of capturing temporal data (e.g., lineage tracing) to model biological processes better and improve predictive accuracy.
5. AI is revolutionizing medical imaging
AI enables early disease detection, analyzing histopathology data, and integrating imaging with omics data for more comprehensive insights.
Jensen Huang, CEO of NVIDIA, showcased how advancements in GPU-based AI accelerate progress in imaging, genomics, and digital twin applications. NVIDIA’s DGX systems have enabled breakthroughs in AI-powered medical imaging and drug discovery.
The Chan Zuckerberg Biohub demonstrated using self-supervised learning on imaging datasets to uncover new insights into cellular architecture and protein localization. Their work on the OpenCell dataset, which mapped over 1,300 proteins in live cells, highlights how AI can reveal previously unknown subcellular compartmentalization.
6. Industry Collaboration and Data Standardization Are Needed
Collaboration between technology providers, pharmaceutical companies, and regulators is essential for scaling AI applications in healthcare. Examples include partnerships between NVIDIA, Pfizer, and Data4Cure and multi-stakeholder initiatives like the FDA’s Sentinel system.
Efforts to standardize data collection, sharing, and analysis are critical for enabling reproducibility and trust in AI-driven insights. Luca Emili from InSilicoTrials mentioned the FDA’s Good Simulation Practice (GSP) framework, which aims to standardize the use of AI in regulatory submissions.
Summary
The 2025 Precision Medicine World Conference (PMWC) highlighted how AI is transforming cancer treatment and drug development to address the sobering reality that only 10% of drugs reach FDA approval despite $90B annual spending.
The convergence of AI with real-world data creates more efficient pathways for drug discovery and patient care through several breakthrough approaches:
Digital twins and AI simulations are accelerating drug development while reducing costs.
Advanced AI models, particularly those combining knowledge graphs with large language models, help decode complex disease mechanisms and identify promising drug targets.
Quantum mechanics-based AI enables the extraction of actionable insights from small datasets—crucial for rare diseases and personalized medicine.
Integrating multi-omics data and AI-powered imaging analysis provides unprecedented views into disease progression at the cellular level.
The industry is moving beyond the “more data is better” mindset toward ensuring high-quality data is collected from multi-omics (DNA and RNA sequencing, epigenetics, spatial transcriptomics, and live cell testing). The FDA’s GSP framework and similar guidelines are critical for standardizing AI applications in regulatory submissions.
A key theme throughout the presentations was that AI should augment, rather than replace, human expertise with tools designed to help researchers and clinicians make better-informed decisions.
2 million wells from 2 million cellular experiments analyzed/week – many terabytes of imaging data
The 35th fastest supercomputer on the top 500 list.
Above from Patrick Collins, Director of Automation at Recursion, talk at SLAS.
Recursion at leader in AI automation for drug discovery.
Dr. Collins said that speed is their most important driver for automation, and facilitating complex workflows is their second important driver. He likened their processes to those of a monk working, a printing press, and an ink jet printer: experiments with manual pipetting, automated repetitive tasks, and automated varied tasks. For speed, they have autonomous mode automation when no one is in the lab. The Recursion paradigm, like that of other companies at the conference, was design, make, test, and learn (or analyze).
Recursion
At the meeting, Recursion had hands down the most impressive AI and automation platform.
Key automation takeaways from Cellino Bio
Cellino is a personalized human cell company. Dr. Jesse Mulcahy, Head of Automation at Cellino, summarized their key takeaways in a slide that resonated with many other talks.
Robots trained with machine learning
The exhibit showroom had many vendors (like Gibson and Ginkgo Bioworks) with stationary robotic arms and autonomous mobile robots (AMR)(High-Res Biosolutions and Veon Scientific). The problem with AMRs is that they are slow so they don’t run onto people. High Res had a robot with manual mobility. The robots at the conference were trained with machine learning (ML) an AI subgroup. Machine learning enables computers to learn from data and make decisions or predictions without being explicitly programmed to do so.
The power of Generative AI (Gen AI) in digital connected labs.
Speaker Matt Gafenco, Customer Solutions Manager at AWS, said, “9 out of the 10 pharmaceutical organizations globally use AWS for generative AI and machine learning, and it’s more essential than ever that you have a foundational data strategy, bridging the gap between what you have already and a new architecture to enabling the lab of the future.”
He outlined the challenges many labs face and where they see the most benefit of GenAI and ML.
He said that GenAI can be leveraged to automate repetitive and time-consuming tasks in the lab, freeing up scientists to focus on more creative and innovative work. Some examples provided include:
Summarizing the latest scientific literature using natural language processing, allowing researchers to quickly process and analyze new information.
Automating instrument operational efficiency monitoring and troubleshooting, reducing the need for manual intervention.
Generating synthetic data for clinical trials and other experiments, reducing the burden of data collection.
Amazon Bedrock and AWS’s generative AI assistant, Cubby, were presented as tools to integrate AI into existing systems, fostering innovation and efficiency in lab operations.
Summary
In the healthcare and life science industry, automation and AI are speeding up the generation and analysis of data at a truly remarkable new rate as evidenced by the talks and exhibits at the SLAS conference this year in San Diego. Cutting-edge technologies, including robotics, GenAI, cloud computing, and IoT connected devices are converging to drive this new era of drug discovery.
As an AI business consultant specializing in biotech, biopharma, and fintech transformation, I attended the GenAI Summit at the Santa Clara Convention Center from November 1 to 3, 2024. My goal was to immerse myself in the latest GenAI developments and evaluate their potential for advancing life sciences and digital healthcare – key areas where AI Nova Strategy provides strategic consulting services.
The Exhibition Experience: In the exhibition room, I experienced firsthand how AI is transforming traditional industries.
The Tesla truck display, with its minimalist interior and AI-ready infrastructure, exemplified how artificial intelligence is reshaping even conventional sectors – a transformation journey that mirrors what we help our consulting clients navigate.
Summit Overview: The conference focused primarily on reducing GenAI hallucinations and showcasing AI startups. As an AI consulting firm, we noted that while there were numerous presentations on technical implementations, only one pitch addressed healthcare, and none covered biotech or biopharma – highlighting the opportunity gap our firm addresses.
I. Productionizing GenAI: Lessons for Business Implementation
Lukas Biewald, CEO, @Weights&Biases presented the first talk that I attended. He said that 30% to 40% of GenAI projects are stuck in production. Projects are “easy to demo but hard to productionize.” One of the biggest problems is disappointment. “Going from 90% to 92% accuracy is hard.” Management sees the money being spent and the progress and pulls the plug.
Software (SW) and AI development are very different. SW development is linear, whereas AI development is experimental. With SW, your code is your IP, whereas with AI, what you learn is your IP. You can’t just look at an AI model file. He also stated that you should be able to swap in a new foundation model. Your application’s infrastructure and workflow should be designed to be model-agnostic.
These insights align with what we’ve observed in our AI business consulting practice, particularly when helping companies move from proof-of-concept to production.
II. The Future of RAGs in GenAI
Tengyu Ma of Voyage AI talked about using Retrieval-Augmented Generation (RAG) to bring proprietary data into an LLM. RAG finds relevant information from a knowledge base and uses that information to create accurate responses.
Ma stated that multimodal embedding models are the future. They transform unstructured data from multiple modalities(PDF, PPT, and images) into a shared vector space. Vectors are the semantic meaning of data (such as words, images, or concepts) in a format that allows for efficient computation and comparison. AI multimodal embedding models can directly vectorize inputs containing interleaved text + images and exhibit near-perfect performance on mixed-modality searches. For biotechnology work, a researcher could search through microscopy images using technical descriptions, find relevant research papers based on diagram similarity, match patient scan images with written diagnostic criteria, and analyze relationships between visual lab results and written protocols. Voyage AI currently provides one multimodal embedding model [voyage-multimodal-3 https://docs.voyageai.com/docs/multimodal-embeddings].
Thinking further into the future, Ma showed two slides:
RAG today with lots of “tricks (in orange)” like GraphRAG, corrective RAG, etc.
The RAGs of tomorrow will have more powerful AI models and less “tricks”.
I have mentioned “embedding models”. “Rerankers” pictured in the slides are sophisticated components in the information retrieval pipeline that perform a second, more detailed evaluation of search results to improve accuracy.
For our biotech and biopharma consulting clients, these developments in multimodal embedding models represent significant opportunities to accelerate research and improve clinical outcomes.
III. AI Agents
Bhavin Shah, CEO of Moveworks, discussed their work with AI Agents. AI Agents plan, observe, and act autonomously. Getting the autonomous component to work is hard.
He showed a slide of where AI projects fail: 1. “Misaligned enterprise-wide strategy and goals 2. Lack of vision alignment across business and technology leaders, 3. Inconsistent monitoring of outcomes.”
At AI Nova Strategy, we can help businesses avoid these problems.
Arind Jain of Glean compared AI agents to Interns. A comparison I have heard many times before.
IV. From Seed to Scale: Navigating the Challenges of Building AI Startups with Venture Capital
This was the final panel I listened to at the meeting. It was a very distinguished panel that included Jon Shalowitz, Managing Director of Union Square Advisors LLC; David Hefter, Director at BlackRock; Rocky Yu, Founder & CEO of AGI House; Murray Newlands, Founder of Open Future Form; and moderator Scott Clark, CEO of Distributional.
Clark asked the group for mistakes they had seen. Shalowitz stated that founders don’t think enough about exits, Yu, don’t talk to users enough, Hefter, being in another company’s roadmap, Newlands, building for now and not the future (for example, it should be building for when communities have autonomous vehicles.). Shalowitz cautioned about investing in AI startups with a 20x evaluation while SaaS has typically a 2x evaluation. He stressed for startups to have a director of partnerships, because partnerships with large companies will be important, as going to production scale with 100% accuracy is very hard.
Clark asked the group for tips on how startups can stand out. Yu, talent. Regarding talent, Hefter commented GenAI is going on its 2 year anniversary so what was the team doing before then?
In closing, the group answered questions and offered closing tips. Tip: Solve the last-mile problem, don’t underestimate how dysfunctional large companies are, don’t bet on getting three home runs in a row, and don’t count on a one-year partnership being renewed. Particularly if the large company has two other partnerships similar to yours—they are just experimenting.
In conclusion, The GenAI Summit highlighted several challenges facing AI implementation today: the difficulty of moving AI projects from demo to production, the importance of being model-agnostic in AI infrastructure, and the growing significance of RAG and multimodal embedding models. Venture capitalists emphasized that AI startups need to focus on building sustainable partnerships and solving last-mile problems.
The recurring themes at the summit reflect what we see in our consulting work at AI Nova Strategy. Here’s how we help companies overcome these hurdles:
1. Set Realistic Expectations
– Help management understand that AI development is experimental, not linear like traditional software
– Set realistic accuracy goals (getting from 90% to 92% is hard)
– Plan for longer development cycles with experimental phases
2. Design for Production from the Start
– Build model-agnostic infrastructure
– Implement RAG strategies for incorporating proprietary company data
– Establish clear metrics for success
3. Build Strong Foundations
– Ensure proper data infrastructure is in place
– Create partnerships with established companies
– Develop clear use cases with measurable ROI
– Start with smaller, achievable projects before scaling
4. Address Common Pitfalls
– Help companies avoid getting stuck in the demo phase
– Guide them through the transition from proof-of-concept to production
– Establish processes for continuous evaluation and improvement
– Help identify and mitigate potential risks and limitations
5. Focus on Business Integration
– Align AI initiatives with business objectives
– Develop change management strategies
– Train staff and build internal capabilities
– Create sustainable maintenance and updating processes
I attended the AI Summit 2024 For Everyone last Saturday, October 26. I met students, developers, speakers, and AI consulting fellows AI consulting fellows who are revolutionizing business with AI. I also met Walter Greenleaf from Stanford, whom I greatly admire. I’m excited about the future of immersive visual reality (VR) in healthcare but still a little worried about VR’s possible misuse. The Summit had a positive vibe, with many breaks with cookies and a lunch with many vegetarian options.
Gopi Kallayil, Google’s Chief Business Strategist AI, started the Summit by asking, ‘Will you trust your life to AI? The life of your children? Well, that’s what you do when you and your family ride in an autonomous vehicle.”
Kallayil urged attendees to think of one big competitive thing to add to their company and how deploying AI could help. “Seek the business challenge first and then work back”- a fundamental principle of AI consulting for businesses.
Ran Nu from NVIDIA pointed out that computing power has increased 10 millionfold over the last 10 years. She discussed an AI factory where you put a token in and get intelligence out. She expressed that AI would augment human creativity and productivity by orders of magnitude.
Sharon Mandell, CIO of Juniper Networks, said that startups should begin their AI projects knowing the regulatory requirements, like ISO requirements. She also said that people change slowly, which will slow AI implementation across a company. When it comes to problems worth pursuing, Mandell suggests going after things that are hard without AI and impossible problems that you couldn’t do before AI, like big data problems. Ivan Lee, Datasaur, commented that just as one vehicle (i.e., minivan, sports car, pickup truck) doesn’t suit everyone, there will not be one LM with the highest ROI for every business. In the future, organizations will use many different models, some off the shelf and some built in-house. The accuracy of the model can be improved by prompting, RAG [Retrieval-Augmented Generation (RAG) is the process of optimizing the output of a large language model, so it references an authoritative knowledge base outside of its training data sources before generating a response], and then fine-tuning.
Philip Rathle, CTO Neo4j, talked about GraphRAG. and how it can improve LLMs’ accuracy and long-term reasoning capabilities.
Rathle compared ChatGPT to a parrot: learns random sentences from random people, talks like a person but doesn’t really understand what it is saying, and occasionally speaks absolute nonsense. But, a parrot, unlike a ChatGPT, is a cute little bird.
Rathle went into how GraphRAG can improve the accuracy of LLM results. GraphRAG improves the standard RAG methodology by utilizing a graph database for contextual information, enhancing the input for large language models.
Ram Menon, CEO of Avaamo (a multimodal AI healthcare service) and very engaging speaker talked about how 90% of people still prefer voice interactions to online interactions and that each consumer service call today costs the provider ~$20 dollars. Menon stated, “With AI agents, they will cost $0.02.” He presented the patient journey, something we all have probably experienced, and wished for something better, and broke it into a system-like component of parts.
Avaamo is working on AI agents for all the parts.
The medication AI assistant and test results agent need 100% accuracy. Making the perfect healthcare AI assistant is hard. In addition to requiring 100% accuracy, medical protocols, like colonoscopy, vary from hospital to hospital, and patients in different parts of the US like different tonal nuances and degrees of empathy.
The Education panel, which included Peter Noris, an Ed Fellow at Stanford, Simon Shim, a Professor at SJSU, Sadid Hasan, Microsoft, and moderator Walter Greenleaf, spoke about how education should prepare students for the future. AI can help personalize learning. There will also be the potential to lose privacy as personal AI assistants follow their emails, texts, calls, and schedules. Hasan quoted futurist William Gibson: “The future is already here; it’s just not evenly distributed.” He meant that today, AI, computers, and the internet are accessible to some but not all students. The panel agreed that, like in the development of aircraft, there will be many crashes along the way.
The healthcare panel comprised Ram Menon, Walter Greenleaf, Alaa Youssef, a Stanford postdoctoral researcher, and moderator Alan N McKellar, VP of engineering at JanusHealth. The panel mentioned that there will be many new ways of monitoring patients, like medical wearables, voice analytics, and sensors. These devices will give clinicians a window into patients’ thoughts and emotions. It will be essential to have strong privacy policies. They discussed the need for an AI roadmap for businesses to integrate these new technologies successfully. The panel noted that medical device engineers must learn to fit their products into the complex healthcare system. Greenleaf has worked with the VA for many years with his visual reality devices to improve mental health. The VA is the largest healthcare system in the US, with 180 medical centers and 25 centers for innovation.
Based on the presentations and discussions at the AI Summit, significant work remains before AI can be fully integrated into various industries, particularly in healthcare, where 100% accuracy is required. Privacy concerns, regulatory compliance, and the need for equitable access across different populations remain crucial challenges that must be resolved. AI business consultants and AI consulting firms will play a crucial role in helping organizations develop their AI roadmaps and make smarter decisions about AI deployment, particularly in fintech, biotech, biopharma, and healthcare sectors where the stakes are exceptionally high.
