Key Moments
Is AI Now Good Enough To Cure Cancer? | YC Paper Club
Y CombinatorKey Moments
AI models trained on DNA can identify disease-causing mutations, but their 'understanding' is limited to statistical patterns, not true biological comprehension.
Key Insights
The cost of DNA sequencing has decreased dramatically, leading to an exponential increase in biological data that current AI models are beginning to process.
Google's Med-Gemini model is outperforming human doctors in diagnosis, raising questions about explainability and deployment.
AI models like AlphaFold are now integral to the drug discovery process, used in some capacity for every drug entering the market.
Omanta is developing personalized research labs for each cancer patient, leveraging AI to analyze their unique data.
AI can model biological systems at the cellular and tissue level, moving beyond sequence-level analysis to predict physiological outcomes.
The Evo 2 and Minerva models, trained on vast amounts of genomic data, demonstrate emergent abilities like in-context learning, allowing for biological design tasks with minimal examples.
AI's emerging role in biological discovery and healthcare
The discussion highlights a significant shift in biology, driven by the collapsing cost of DNA sequencing, which has generated an overwhelming amount of data. AI models are now catching up, enabling the mining of this data for insights previously unattainable. This surge in data is transforming cancer care into a per-patient data problem, where AI can analyze individual genomic profiles to tailor treatments. The speakers emphasize that AI is not just assisting but is becoming foundational to biological discovery, with tools like AlphaFold being used in the development of nearly every new drug. The potential for AI extends to modeling entire cells rather than just molecules, learning from DNA sequences, and even exhibiting surprising capabilities like in-context learning in genome models without explicit training for it.
The data deluge in cancer genomics
Historically, cancer research relied on analyzing average trends across limited patient cohorts, as exemplified by the Cancer Genome Atlas (TCGA) project a decade ago. The high cost of sequencing restricted such studies to hundreds of patients. Today, with the plummeting cost of sequencing and insurance coverage for genomic testing, the paradigm has shifted. We can now generate massive datasets, allowing for the analysis of individual patients as unique datasets rather than just looking for statistical summaries. Each patient's genome can contain around 20,000 genetic data points, which, when combined with other multi-dimensional data like copy number variations, RNA expression, splicing, and cellular context, creates a 'big data' problem. The challenge now is to effectively mine this data to improve patient outcomes, a task for which current infrastructure is still catching up.
Personalized research labs and AI agents
The conversation introduces the concept of 'personalized research labs' for each patient, a vision being pursued by companies like Omanta. The traditional approach to large-scale genomic studies required massive teams of specialists, making such research accessible only to well-funded institutions like the Broad Institute or Sanger Institute. However, the advent of AI agents is democratizing this process. These agents can function as bioinformaticians, processing and analyzing vast amounts of data in real-time. This enables a shift from analyzing similarities across many samples to deeply mining a specific patient's dataset to optimize their individual treatment. The goal is to create an AI-driven infrastructure that can extract data in real-time, link it to the latest discoveries, and provide actionable insights rapidly, potentially shortening research timelines from years to real-time and transforming patient care.
Advancements in cell modeling and virtual cells
The discussion delves into 'virtual cell' research, which aims to use AI to model biological systems at the cellular, tissue, and organismal levels. While AI has shown immense capabilities at the sequence level (DNA, proteins), bridging this to the complex, dynamic environment of a cell presents a significant challenge. Researchers are developing models like the 'Universal Cell Embedding Model,' trained on millions of cells, to map cellular states and predict outcomes. This approach is inspired by the biological concept of 'Waddington's landscape,' where cell states are not random but follow specific trajectories. By using self-supervised learning on gene expression data, these models can create a unified representation of cells, revealing new relationships and structures. The goal is to create virtual spaces where different cell states can be mapped, and interventions can be tested to move from diseased states toward healthier ones, addressing a major cause of clinical trial failures: poor prediction of efficacy in humans.
AI models trained on DNA: Evo 2 and Minerva
The Evo 2 and Minerva models represent significant progress in training AI on DNA. Evo 2, trained on 9 trillion base pairs, uses a hybrid attention-convolution architecture to process long genomic sequences. The training process involves a data augmentation strategy that focuses on evolutionarily conserved regions, revealing that such focused data is crucial for model performance. The models learn to predict the next DNA base, and this ability translates into powerful predictive capabilities. For instance, Evo 2 can identify disease-causing mutations by assessing the likelihood of a given sequence, performing comparably to or better than human-specific supervised models. Beyond direct prediction, the models' internal representations, or embeddings, reveal learned features related to biological functions like bacteriophages and CRISPR systems, even when these were not explicitly labeled during training. This allows for exploring biological 'languages' that humans cannot directly read.
In-context learning and biological design
A surprising emergent property of large language models like Evo 2 is in-context learning (ICL), where the model can perform new tasks based on a few examples provided in the prompt, without any weight updates or fine-tuning. This capability is being leveraged for biological design. By providing a few examples of desired properties—such as specific binding affinities for proteins or desired activity in cell-specific enhancers—Evo 2 can generate novel sequences predicted to possess those traits. This is particularly powerful in biology, where obtaining large labeled datasets for specific design tasks can be challenging. The models can be prompted to design new bacteriophages, land modulins (proteins for rare earth metal binding), antibody designs, and CRISPR-Cas systems. This few-shot prompting approach significantly reduces the computational cost and data requirements compared to traditional fine-tuning methods.
Bridging the gap between virtual and physical biology
A critical challenge lies in translating AI's predictions and virtual models into real-world biological outcomes. For virtual cells, the aim is to accurately simulate physical systems, but noise and variability in experimental measurements make this difficult. The development of AI agents that can effectively interpret and generate biological data is key. Researchers are exploring reinforcement learning and automated machine learning techniques, borrowing from pure AI research, to create more efficient discovery loops. The goal is to build AI systems that can interact with physical labs, generate optimized data for specific biological questions, and accelerate the development of useful virtual cell models. Safety is a paramount concern, especially when designing novel therapies, necessitating robust validation and iterative testing to avoid unintended consequences.
The future of AI in biological discovery
The overarching vision is to shorten the timeline for biological discovery and therapeutic development. AI models can identify promising avenues for research, design novel molecules and systems, and even optimize experimental designs. The integration of AI with physical experiments, creating closed-loop systems, promises to accelerate breakthroughs. The speakers are optimistic that AI will not only help us understand complex biological processes but also enable us to engineer biological systems for specific purposes, ultimately leading to better patient outcomes in areas like cancer treatment. The ongoing work in self-improving models and agents suggests a future where AI plays an increasingly central role in scientific research, potentially shortening discovery cycles from decades to years or even real-time.
Mentioned in This Episode
●Software & Apps
●Companies
●Organizations
●Books
●Studies Cited
●Concepts
●People Referenced
Common Questions
The speaker conducted a live experiment suggesting that current AI chatbots, while advanced, may not consistently fool humans into believing they are interacting with another person. The results indicated a significant portion of humans were mistaken for AI, and vice versa, suggesting the test is not definitively passed.
Topics
Mentioned in this video
A company involved in large-scale AI investments in biosciences.
A protein component of the CRISPR-Cas9 system, designed using few-shot prompting.
An AI model mentioned for its understanding of the biological landscape through text.
A foundational model for biology that Garrick led, trained on 9 trillion base pairs.
A programming language used in AI development, confirmed to be used by the Locus system.
A gene-editing technology used in bacteria for defense against viruses, and as a basis for AI model features.
A predictive model used as a reward function in reinforcement learning for biological design.
A protein structure prediction tool used in drug discovery, with usage in a significant portion of marketed drugs.
A reverse enzyme with a CRISPR-like RNA matrix, discovered through DNA language model analysis.
An AI model mentioned for its understanding of the biological landscape through text.
An AI system developed by Intelligi, applied to molecular design and automated research.
A company or project that generates large datasets, mentioned in the context of AI investments.
A language model used as a reference for next-token prediction and generative tasks.
A new language model for DNA that predicts interactions between DNA parts, aiding in discovery.
An algorithm used to update the Evo 2 model in a reinforcement learning setup.
A large, open repository of web crawl data used for training language models.
A statistical measure used to evaluate the performance of AI models in predicting functional mutations.
A component of the CRISPR-Cas9 system that can be designed using Evo 2.
A gene associated with increased cancer risk, used as a benchmark for evaluating AI models in predicting functional mutations.
An emergent property of large language models that allows them to learn from examples provided in the prompt.
A method for optimizing functions that are expensive to evaluate, relevant to scientific discovery and AI research.
A technique used to guide AI model training based on human preferences and feedback.
An advanced AI training technique where AI models provide feedback to improve other AI models.
A test proposed by Alan Turing in 1950 to determine a machine's ability to exhibit intelligent behavior equivalent to, or indistinguishable from, that of a human.
A genus of bacteria whose genome was analyzed for DNA base-pairing interactions.
University where Alfredo studied early-acting transcription factors.
Divya's laboratory at MIT.
University where Yusha obtained his PhD and where Garrick is a PhD student.
The National Center for Biotechnology Information, a database where genomic data is stored and has doubled in the last three years.
An institute where Yusha founded the machine learning team.
University of California, Los Angeles, where Alfredo pursued his PhD in cancer bioinformatics.
A family of bacteriophages used for fine-tuning Evo 2 to generate functional phage genomes.
An institute where Alfredo worked on early CRISPR screens, and where he met his co-founder.
Massachusetts Institute of Technology, where Divya worked on structure-based methods before joining Hezi.
A repository of raw gene sequencing data accessed by agents for processing.
A startup focused on pure AI research and automated AI research systems.
A company from the speaker's lab, led by Eric Nguyen, developing AI-driven therapeutics.
A company co-founded by Alfredo and Renato, focused on treating cancer by building personalized research labs.
An organization where the YC Paper Club and other research initiatives are hosted.
A company where Yusha is Chief Science Officer, focusing on self-driving AI research in science.
A company involved in large-scale AI investments in biosciences.
A company that provided GPUs for training the Evo 2 model.
A company whose employees confirmed the widespread use of AlphaFold in drug development.
A speaker discussing virtual cells and AI models for cell states.
A British mathematician and computer scientist who proposed the Turing Test.
A supervisor for Divya's research on generative AI models for biological design.
Leader of Radical Numerix, a company developing AI-driven therapeutics.
Mentioned in relation to AlphaFold's widespread use in drug development.
More from Y Combinator
View all 643 summaries
45 minHow To Build The Future: Paul Graham & Amjad Masad
48 minBringing Supersonic Flight To The People
49 minWhat Happens When Everyone Can Build Software
111 minWhy Inference Matters 12x More Than Training | YC Paper Club
Ask anything from this episode.
Save it, chat with it, and connect it to Claude or ChatGPT. Get cited answers from the actual content — and build your own knowledge base of every podcast and video you care about.
Get Started Free