Compiled & Compiled by: Deep Tide TechFlow

Deep Tide Guide: During the 10th special event of "YC Paper Club" titled "AI for Biology" held on November 8, 2026, the visiting partner Francois Chaubard, who is about to leave Y Combinator to start a new venture, first introduced a live scanning blind test experiment on "Whether the Turing Test has truly been passed." This was followed by presentations and a Q&A session with four leading AI biology and computational oncology researchers, fully deciphering why now is a critical turning point for AI to reshape cancer treatment and life sciences.

The whole discussion lasted nearly 82 minutes and progressed around three main lines: First, Alfredo Gonzalez, co-founder and CEO of Omanta, revealed how to transform cancer from a "population average" into a "single patient data problem" after the cost of gene sequencing experienced a steep decline. Next, Yusuf Roohani, former head of machine learning at Arc Institute and head of science at Intology AI, analyzed the data bias traps exposed by a database of 500 million single cells and how to build a physical experimental feedback loop using large model self-training.

Finally, Stanford PhD student Garyk Brixi and Divya Nori demonstrated the Evo 2 genomic large model, the bidirectional co-evolution model Minerva, and few-shot context learning (ICL), detailing how the model can predict pathogenic mutations directly from massive DNA sequences and design gene editors without requiring fine-tuning, achieving empirical results in mouse brain wet experiments.

Guests: Alfredo Gonzalez (Co-founder and CEO of Omanta, PhD in Cancer Bioinformatics from UCLA), Yusuf Roohani (Head of Science at Intology AI, former head of machine learning at Arc Institute), Garyk Brixi (Stanford PhD student, team lead for Evo 2, author of the Minerva project), Divya Nori (Stanford computer PhD student, researcher at Brian Hie’s lab)

Host: Francois Chaubard (Visiting Partner at Y Combinator, Founder of YC Paper Club)

Podcast Source: Y Combinator / YouTube (https://youtu.be/-XL2gClQ2gg)

Original Title: Is AI Now Good Enough To Cure Cancer? | YC Paper Club

Broadcast Date: November 8, 2026

Key Points Summary

• Live Turing Test Blind Test: Francois Chaubard spent 12 hours using Claude Code to build a test webpage for blind testing the audience. After injecting a 32-year-old male persona and the live context, GPT-6 Astra tricked humans with the highest rate of only 36.4% (Claude Fable 5.1 was 30.0%, GLM 5.3 FlashX was 26.1%, Gemini 3.8 Flash was 23.8%, Grok 4.6 was 23.1%, Kimi K3 was 19.0%), far below the 70% passing threshold; while humans were also misidentified as AI at a rate of 33.3%.
• The Core Driving Force for AI Bio's Singularity: Sequencing costs have plummeted, and Evo 2's training set has expanded to 90 trillion base pairs; AlphaFold is now used in every new drug development phase going to market, AI protein binders achieve a success rate of 10% to 100% on the BindCraft benchmark, and phages designed by Radical Numerics have entered patient clinical applications. The entire industry is migrating a loop similar to verifiable reward reinforcement learning (RLVR) to biological experimental validation.
• Viewing Cancer as a "Single-Data Problem": Omanta utilizes each patient's multidimensional data of up to 20,000 genes, RNA expressions, and clonal evolution, replacing traditional bioinformatics labs with 30 to 40 people with a multi-agent pipeline, enabling rapid identification of new cross-tumor pathway solutions for rare tumors or patients failing frontline chemotherapy.
• The Data Bias Trap and Closed-Loop Experiments of "Cellular Common Crawl": Yusuf Roohani's team made agents run continuously for a month in the NIH original transcriptome database, boosting the cell data count from 35 million to over 500 million, only to find excessive over-sampling of immune cells and T cells; the solution lies in migrating "large model self-training (such as Intology AI's Locus system exploring PostTrainBench)" to AI-guided closed-loop experimental data generation rather than blindly piling data.
• Biological Understanding from Pure DNA Next-Base Prediction: Garyk Brixi's team trained 7 billion and 40 billion parameter models Evo 2 on 90 trillion base pairs after deduplicating genome sequences across all life domains. The model, having only seen one set of human reference genomes, performed comparably or even superiorly to specialized human supervised models in predicting pathogenic mutations in BRCA1 coding and non-coding regions; combined with the newly released bidirectional model Minerva, it can accurately identify co-evolution interaction surfaces and entirely new non-coding RNA and CRISPR reverse transcriptase systems.
• Few-Shot Context Learning (ICL) of the Genomic Large Models and Mouse Brain Empirical Evidence: Divya Nori and others found that with only 4 to 64 natural sequence examples as contextual prompts (without any parameter fine-tuning), Evo 2 could autonomously reorganize motifs and design rare earth binding protein lanmodulin, specific targeting antibodies, a complete set of CRISPR-Cas9 systems, and complete phage genomes; after combining AlphaGenome as the reward model and utilizing GRPO reinforcement learning for post-training, the designed cell-type-specific enhancer achieved successful specific region fluorescence activation in real mouse brain experiments.

Exciting Perspectives

• "In medical clinical practice, relying solely on the historical average of the population often has no meaning for an individual patient facing life and death; when each tumor itself carries two decades' worth of gene information, each patient is an independent big data set waiting to be mined."
• "In the past, only institutions like the Broad Institute or Sanger Institute could afford to maintain a top-tier team of 30 or 40 algorithm experts, bioinformaticians, and clinical medicine teams. Now we have automated this entire laboratory infrastructure and operate it in real-time around a single patient."
• "Even if you direct agents to scan all available NIH original cell data and expand the dataset to over 500 million cells, you would still realize that the vast majority are immune cells and T cells. Blindly piling data does not guarantee true biological diversity."
• "At the beginning of the leading cell biology textbooks, the first sentence states: In many ways, humanity's understanding of the structure of the universe exceeds our comprehension of the mechanisms of living cells; we have sequenced the complete genomes of many species but still cannot predict how a cell will behave after knocking out an unstudied gene."
• "Humans can read natural language, understand code and mathematical formulas, yet cannot directly comprehend DNA sequences. Genomic large models essentially translate a codebook written through millions of years of natural selection into a tool for humans and AI scientists to question."
• "Before training, we merely changed the ratio of randomly selected whole genome sequences to favor genes and conserved elements, while still predicting the next base. The model's performance on clinical mutation assessment benchmarks could jump from bottom of the pack to industry-leading (Spearman correlation increasing by 0.16)."
• "No one has specially trained Evo 2 for context learning, yet simply predicting the next base at the nucleotide level, the model naturally emerged with 'few-shot design capabilities.' Even with just 4 examples of CRISPR-Cas sites, it can generate a complete system directly."

Text Body

Live Testing of the Turing Test: GPT-6 Astra's Deception Rate at 36.4%, One-Third of Humans Misidentified as AI

Francois Chaubard: Good evening everyone! Welcome to tonight’s "YC Paper Club" AI for Biology special session. Today we have several top scholars and founders with whom I've worked in the lab and collaborated at YC, the lineup is fantastic. However, before diving into life sciences, I had a spontaneous idea to ask everyone present to take out their phones and scan the QR code on the big screen.

Recently, I had dinner with Greg from ArcAGI, and he firmly believed that "the Turing Test has been conquered". I challenged him on the spot: Alan Turing said in his classic paper in 1950 that about 50 years later (which means around the year 2000), we could write a program that would allow an average-level interrogator to distinguish between a machine and a human with a probability no greater than 70% after 5 minutes of continuous questioning (i.e., at least 30% of the time they would not spot it; and the live comparison chart marked the passing threshold at 70%). However, the reality is that every time I chat with an airline customer service robot, I can see through it in just a few lines.

So, over the past 12 hours, I used Claude Code to temporarily "Vibe Code" a blind test website. The rules are simple: half of the on-site audience is randomly paired with real human chat partners, while the other half is paired with an advanced model disguised as a 32-year-old male who is fully aware of the "context of sitting in the YC on-site room right now." Everyone has two minutes to probe each other and vote on their judgment.

Okay, the results of the two-minute test are in! You can see the bar graph on the big screen showing the performance of several mainstream models: the highest performing one was GPT-6 Astra, achieving an impressive 36.4% of testing rounds where people mistook it for a real person; closely followed by Claude Fable 5.1 (30.0%), GLM 5.3 FlashX (26.1%), Gemini 3.8 Flash (23.8%), Grok 4.6 (23.1%), and Kimi K3 (19.0%).

A very interesting comparison baseline is at the bottom of the chart: The proportion of real humans misidentified as AI by the audience here is as high as 33.3%! It seems that as long as everyone speaks a little more concisely and calmly, one in three people will be mistaken for a robot. Overall, even though I deliberately filled the model with on-site persona details and repeatedly adjusted the prompts to "favor" AI, the overall passing rate stayed around 30%, still a way from being undisputedly passed. By the way, a shout-out to Boris and the Claude Code team for enabling me to set up this real-time system in 12 hours.

Farewell Remarks: Why Now is the Best Time for AI Bio?

Francois Chaubard: Next, let's move on to tonight's central theme: Why should we pay attention to "AI for Biology"? During group meetings in the lab, I sit next to Yasa, who constantly pitches me on biological AI. His most persuasive line is: "When one day you actually find out you have cancer, you will deeply regret not spending more time researching AI for treatment." Although his rhetoric may not change people's careers immediately, over time I have truly been convinced, and I have begun to engage in related research myself.

On the slides, I have listed several core signals igniting the AI Bio era right now:

1. Unprecedented volumes of molecular and genetic data: Evo 2 has already completed training on 90 trillion base pairs, and the next generation of models is accelerating on a larger scale; the cost of gene sequencing continues to plummet, resulting in an exponential emergence of biological data.
2. Basic tools fully entering clinical drug pipelines: Demis revealed during his YC talk a few months ago that now every truly market-driven new drug uses AlphaFold at some stage of its development. At that time, I thought this was an exaggeration, but later, after privately consulting a dozen development experts and top PhDs from Genentech, everyone confirmed, "Yes, that’s about right."
3. Acceleration in protein design and real human therapies: AI protein binders have achieved a success rate of 10% to 100% on the BindCraft benchmark; the AI-generated phage therapy from Radical Numerics, founded by Eric Nguyen from the same lab background as mine, has also truly entered human patient testing; soon, Divya will demonstrate how only 4 to 64 examples for context learning (ICL) can guide biological sequence design accurately.
4. High-quality benchmarks and verifiable reinforcement learning (RLVR) loops forming: Whether it’s the CASP for structural predictions or the Virtual Cell Challenge for cell modeling, evaluation benchmarks are becoming increasingly clear; from a capital perspective, billions are flooding into this field from Lila, Chai, to Isomorphic. Moreover, the entire AI sector has figured out a methodology for recursive self-improvement (RSI / RLVR) aimed at "verifiable tasks", where the core flywheel has taken shape, i.e., using basic models to sample candidate drugs, picking Top N for efficacy testing, and recycling real data to fine-tune models for continuous iteration.

Moreover, I must bid a somewhat sentimental farewell: this episode is actually my last as the host, and it's the 10th episode of "YC Paper Club". Over the past six months, building this hardcore academic community with all of you has been a huge honor for me, with the previous nine episodes collectively receiving over a million views. This Friday, I will officially leave YC and embark on my next entrepreneurial journey. Next, I would like to welcome our first heavyweight guest of the evening, Alfredo Gonzalez.

Alfredo Gonzalez: When Tumor Sequencing Becomes Cheaper than Dinner, Transforming Cancer into a "Single-Patient Data Problem"

Alfredo Gonzalez (Co-founder and CEO of Omanta): Thank you, Francois, for your tremendous contributions to this community over the past six months! The sparks of thought generated from previous episodes have been truly shocking. I am Alfredo, one of the co-founders of Omanta, and what we are doing is building a comprehensive "dedicated R&D laboratory" centered on each individual cancer patient's specific tumor to help patients overcome cancer.

Let me briefly introduce my academic background: after my undergraduate studies, I first researched early developmental transcription factors at Harvard, then joined the Broad Institute, where I participated in the very first commercialized and academic CRISPR gene screening validation in cancer cell lines, and there I met Ranad Humeidi, who led the drug repurposing project and later became my co-founder and wife. Afterward, I pursued a PhD in cancer bioinformatics at UCLA, where I was responsible for building a large-scale automated analysis pipeline for tumor and normal tissue genomes, contributing to building one of the largest rare sarcoma tumor datasets globally, and my core doctoral focus was tackling the "N-of-1" rare difficult tumor cases.

Today when we at YC call out "using AI to conquer cancer," it's not just hot air; it's the first time in human history that we are genuinely standing at this critical juncture. I want to share around three points:

First, Why has cancer become a purely "data mining problem" today? As Francois just showed in the sequencing cost curve, the cost of a complete genome sequencing is now even cheaper than having dinner in the San Francisco Bay Area; more importantly, for the vast majority of late-stage and recurrent cancer patients, tumor sequencing has already been included in the reimbursement systems of the US health insurance, including Medicare and Medicaid.

Back then, a decade ago, humanity achieved a monumental milestone with the “The Cancer Genome Atlas (TCGA).” But at that time, under the then-technical conditions, it was already a limit to scan several hundred patients for each discrete tumor type. Due to limited sample size and data dimensions for a single patient, doctors and scholars were forced to look for "population averages" and the top-ranked high-frequency mutations from large cohorts, yet these population statistical averages often provide no practical guidance for specific patients.

Now the situation has changed: just one patient’s complete genome, copy number variation, RNA expression, variable splicing, and cellular states unfold and inherently form a huge independent dataset covering about 20,000 gene dimensions. The academic community in the past might have only utilized less than 0.1% of the surface data; now that we have complete data, what's truly missing is the infrastructure to thoroughly exploit this 20,000-dimensional treasure in real-time.

Second, How to build "personalized laboratories" using multi-agent systems? Looking back at the number of people needed to reach Broad Institute or Sanger Institute milestone research a decade ago? You needed a full roster of algorithm experts, infrastructure engineers, cluster operations technicians, variant detection specialists, geneticists, data visualization engineers, plus a clinical MD to interpret the results, easily a luxurious team of 30 to 40.

Now, everyone in the software industry talks about "software factories," and in the biomedical field, a similar intelligent agent framework signifies an "AI version of a dedicated laboratory factory": tasks that previously required 30 to 40 top bioinformatics experts months to perform—cleaning, clustering, and pathway inference—can now be decomposed into specialized AI agent clusters focusing on different roles (one focusing on clinical trial matching, one specializing in safety and toxicology evidence, one reading biology backgrounds and latest literature). As soon as new mechanisms that can be migrated from the latest papers get released on Tuesday morning, the system can pull them that afternoon to improve the treatment plans for patients being treated, compressing the months-long translational cycle into real-time.

Let me share two real cases where we at Omanta helped patients achieve remission:

• Breaking through for a rare brain tumor patient: Traditional medicine rigidly categorizes tumors by anatomical site, and when encountering rare cancers, physicians are often left with no literature. However, the underlying biological pathways of cells do not care about classification labels humans impose; neuroblastoma and lung adenocarcinoma could very well be driven by the same pathway. We once had a very rare brain tumor patient with no available drugs, and upon re-clustering his multidimensional dataset, we discovered that his tumor was molecularly very close to mesenchymal tumors; leveraging the truly activated cell surface markers and driving pathways, we successfully found a new therapy combination that would have otherwise never been mentioned in past literature but targeted the core mechanisms.
• Revising remedial treatment after multiple-line chemotherapy failures: Another patient had previously undergone three rounds of cytotoxic chemotherapy, all of which failed, drastically diminishing her quality of life. We conducted clonal evolution tracing and mining on her early original samples, and the data clearly proved that the three rounds of chemotherapy previously selected by the hospital using standard methods were biologically unsuitable for her given the high heterogeneity within her tumor. If this automated research infrastructure had been in place from day one, she would have completely avoided the torment of undergoing those three rounds of toxic chemotherapy.

Third, I want to discuss my vision for the future: Beyond drug repurposing with existing approved drugs, we will in the future real-time customize entirely new protein drugs and delivery peptide chains for specific patients. Just like the Bennett et al. paper from David Baker’s team has shown, AI can shorten the traditional R&D screening cycle by ten years; combined with Google's DeepMind's promotion of an AI scientist collaboration system (Co-scientist) in Mountain View last year, while the models may not necessarily surpass top human professors in every aspect, as long as their performance is close enough, they can save countless hours of trial and error time in laboratories on a global scale.

On-Site Questions: In your PPT, you mentioned using “organoids” as a crucial data source for establishing a personalized laboratory for a single patient. Do organoid plans truly have scalable feasibility when building these labs?

Alfredo Gonzalez: Great question. During my PhD, I was involved in several organoid projects. Objectively, there have been two waves of organoid startups in the past seven to eight years; from a pure scientific viability perspective, they absolutely excel: organoids can amplify extremely scarce live tissues of rare tumors and can be used directly for high-throughput drug perturbation testing, just like metastatic samples. The real bottleneck lies not in the science but in the systematic procedures of hospitals: currently, most hospitals' first reaction after receiving surgical slices is to dump them in formalin fixation, which completely kills the live cells and terminates the possibility of cultivating organoids afterward. What needs to be resolved is precisely enhancing the sample preservation infrastructure at the front end of the hospital.

On-Site Question: Even if you have successfully identified the actual pathogenic pathways of rare tumors in the lab or found the right drugs, how do you negotiate the "final step" in reality to enable clinical doctors to understand your reports and truly agree to prescriptions or even conduct N-of-1 clinical trials for individual patients?

Alfredo Gonzalez: Clinical doctors always hold the final medical decision-making power and responsibility boundaries. However, we find a very positive reality in our actual collaborations: today's medical oncologists are working extremely hard and under significant stress; they simply don't have time to set up computing clusters or run whole-genome differential analyses. Especially when standard treatments fail one after another, leaving doctors with no choice but to guess the next line regimens based on experience, if they can receive a tightly logical and well-evidenced multi-dimensional biological data report, most clinical doctors are more than willing to use it as a basis to argue for new solutions for their patients; as for initiating entirely new molecular N-of-1 trials from scratch, that falls under another compliance approval process.

On-Site Question: How do you view the recently much-discussed direction of "reversal of malignancy"? For instance, a South Korean institution claims to reverse cancer cells back to a normal cell state using organoid signal models?

Alfredo Gonzalez: I am not sufficiently informed about their specific experimental details to make a definitive comment. However, I believe that any discussion of "reversing malignancy" should consider two core hard indicators: first, how you strictly define what is meant by "reversal" at the molecular or phenotypic level (different subtypes of tumors may evolve completely differently); second, we must account for real clinical outcomes based on long-term tracking of patients as support.

On-Site Question: While exploring radical options in individual AI personalized laboratories, how can you ensure absolute safety and avoid fatal harm to patients due to algorithmic errors?

Alfredo Gonzalez: Safety always comes first, and in fact, just the opposite: this infrastructure significantly enhances treatment safety. Just like the patient who previously endured three rounds of ineffective chemotherapy, accurately filtering out ineffective toxic options is the most substantial injury reduction measure. If we prioritize matching pathways from an approved or already completed phase one clinical safety-verified drug library, the safety profile of those drugs is inherently clear; even if moving toward future personal drug molecules "de novo," we can completely follow the modern clinical trial's mature dose escalation safety control pathways.

Yusuf Roohani: The Bias Traps Exposed by 500 Million Single-Cell Databases and Accelerating Virtual Cells with "Autonomous Post-Training" Loops

Yusuf Roohani (Head of Science at Intology AI, former Head of Machine Learning at Arc Institute): Thank you for the invitation from the organizers! Today, I was originally assigned to talk about "Virtual Cells," but I plan to add a different perspective by combining my deep involvement in large models' Autonomous Post-Training work over the past few months and discussing its deep connections with building truly usable virtual cells.

What is "scientific discovery" and "virtual cells"? Scientific exploration is essentially a giant "search problem": facing an assumption space that is almost infinitely large, researchers wish to find the most effective information with the least experiments to unveil the causal laws of a certain phenotype or disease. The primary value of molecular models like AlphaFold or ESM is to significantly accelerate searches: without crystallizing all proteins in wet labs, we can first deduce their three-dimensional spatial structures from sequences via models, thereby efficiently prioritizing experiments.

The core ambition of virtual cell research is to elevate this computational capability from a purely "molecular sequence level" to a level more closely aligned with real disease symptoms and physiological laws at the cell, tissue, and even the entire organism level. In reality, the primary reason clinical trials often fail is not that molecular interactions do not work (we have become quite adept at designing high-affinity molecules for known targets), but rather that "clinical efficacy does not materialize": when originally selecting targets, we did not understand which switch to manipulate to regulate living cells. The reason living organisms are so unpredictably unstable is that cells are not static molecules but a mix of complex molecules in a state of chaotic interaction and constant motion.

To model cellular states, we previously developed and published the "Universal Cell Embedding" which employed a very intuitive approach: stitching together gene expression counts from thousands of single cells into a "cell × gene" transcript count matrix, randomly hiding parts of gene expressions and using self-supervised prediction for reconstruction. In fact, over 100 years ago, developmental biologists proposed the famous "Waddington Landscape": due to transcriptional and epigenetic constraints, live cells are unlikely to be in arbitrary disordered states, but rather are constrained within a "low-dimensional manifold" space resembling that of a large model.

After feeding 35 million single cells into an unsupervised autoencoder matrix, even without any artificial labels, the model’s output UMAP manifold space spontaneously clustered into large cell classes and subpopulations that closely matched over a century of biological literature!

But the fundamental question is: Natural language processing has completed Scaling Law miracles by using Common Crawl to scrape the internet; can we do something similar in cellular biology, sweeping through the entire public cell libraries amassed by humanity?

Earlier this year, in another study published in Cell, we made this bold attempt (long before Claude Code became widespread in 2025, we were among the first teams in the industry to deploy autonomous agents to mine data): We had agents continuously run for a whole month, automatically cleaning and reprocessing all single-cell records in the NIH public original transcriptome database, aggressively increasing the data volume from 35 million to over 500 million cells.

However, when we excitedly opened up this hot distribution map that transcended all tissue and cell types, a sobering result poured cold water on our enthusiasm: The vast majority of the 500 million cells available online were found to be heavily clustered around "immune cells," and among them, more than half (a full order of magnitude higher) were T cells! In other words, even if you turned over all existing open-source cell libraries, merely linearly piling up public data does not buy true biological breadth and diversity.

Therefore, the true way out does not lie in blindly increasing data linearly, but in switching to "AI-guided closed-loop experimentation (Scaling AI-guided experimentation)": utilizing cutting-edge models like Claude or GPT, which have acquired a global view from vast biological texts, to precisely decide what type of rare scarce cell data should be collected in the next round of wet experiments.

Once we treat R&D as a "hypothesis proposed → experimental validation → feedback updating" cycle structure, I was surprised to discover that it faces the same engineering challenges as the current cutting-edge artificial intelligence labs striving for autonomous large model post-training (Autonomous LLM Post-Training) and recursive self-improvement (RSI): whether for running wet experiments or training models autonomously, the core issues to address are three major bottlenecks: extremely high costs of single evaluations, noisy experimental/testing feedback, and an excessively vast hypothesis search space.

To verify whether this discovery loop can cross-industry, six months ago, I joined a startup focused on general automated AI R&D systems called Intology AI. In past months, our autonomous R&D system Locus achieved leading results on the authoritative list PostTrainBench, which assesses whether large models can complete autonomous post-training of base models such as Qwen, outperforming many common programming agents. The core reason isn't mystical: general programming agents on the market are not optimized for "scientific research retrieval and hypothesis trial-and-error"!

In the latest evaluation released just yesterday, we almost seamlessly migrated this Locus system optimized in a pure machine learning post-training scenario to the task of "AI molecular design and property prediction" (allowing the agent to perform high-efficiency sampling retrieval and control noise verification set overfitting based on the latent space of foundational models).

The results showed that under the strict premise of locking down the same GPU computing sandbox budget, as the computational budget (H100 GPU hours) continues to rise, Locus significantly distanced itself from Anthropic and OpenAI’s general programming agents, outperforming ready-made tools like BoltzGen and Codex in specific sampling efficiency tests (of course, these are currently computer simulation results, and the next crucial step must return to wet experimental verification).

Finally, I want to conclude with the first sentence from a classic cell biology textbook: “In many aspects, humanity's understanding of the structure of the universe even surpasses our comprehension of the mechanisms of living cells; although we have sequenced the complete genomes of many species, we still cannot predict how a cell will behave after knocking out even a single unexplored gene.”

On-Site Questions: What specific technology did you use during the post-training process (is it reinforcement learning RL)? How do you evaluate the gap between simulation (Virtual cell) and the real physical world?

Yusuf Roohani: To answer the latter question first: whenever we measure the same cell under completely identical laboratory conditions, biological fluctuations and noise from detection instruments are considerable, hence "how to accurately reproduce the physical world as much as possible amidst noisy evaluations" is the highest difficulty challenge; ultimately, all simulated evaluations must withstand the scrutiny of wet experimental realism. As for the first part, you’ve pinpointed the most critical issue: traditional programming agents aren’t optimized for scientific trial-and-error design; there aren’t dedicated intelligent agents responsible for long-term attention to hypothesis progression, controlling single experimental computation costs or funding, and strategically evaluating candidate hypotheses. It is precisely by reconstructing this multi-agent collaboration mechanism around the R&D workflow that automated machine learning achievements smoothly transitioned into the scientific discovery domain.

On-Site Question: Your last page's chart has "H100 computing time" on the horizontal axis; how did you ensure a fair comparison with Anthropic and OpenAI? Also, do you measure cost by token consumption or training steps?

Yusuf Roohani: What we did was a fully controlled side-by-side comparison: we placed our own Locus system and the ready-made autonomous agent systems from Anthropic and OpenAI into an identical sandbox with the same H100 GPU quota to run. Since this task involves training and fine-tuning scientific models driven by large-scale models, the true bottlenecks and major costs arise from GPU training compute hours rather than pure inference dialogue tokens.

On-Site Question: Could you share a little about Locus’s underlying architecture? For instance, how was it built?

Yusuf Roohani: It's inconvenient to disclose the specific proprietary architecture, but the core is built on a large language model as the engine foundation, with a series of research-specific engineering strategies refined around "formulating experimental ideas, avoiding noise overfitting in validation sets, and long-term memory retrieval." By the way, we still do the underlying coding in Python (laughs, just let that slip).

Francois Chaubard (follow-up): In the real world where sampling costs are extremely high, sample efficiency is vital. Recently, many impressive improvements in the large model world tend to stem from "Continuous Harness" frameworks and contextual memory updates that do not alter model weights. Do you think the emphasis should be on fine-tuning parameters or on frameworks and context engineering?

Yusuf Roohani: I fully agree with your judgment. Especially in the autonomous R&D and exploration phase in biotechnology, first optimizing external Harness (tool scaffolding and context updating mechanisms) is absolutely the highest cost-effective and quickest way to see results; parameter fine-tuning and post-training are more suitable when you need to consolidate high-frequency capabilities for rapid reasoning and retrieval. However, to ensure that general discovery capabilities do not blindly overfit to a single benchmark (bench maxing), we have shown remarkable restraint in modifying weights during the early exploration phase, prioritizing the intensification of Harness and inference framework design.

Garyk Brixi: Understanding Life Codes by Reading 90 Trillion Base Pairs as a Language, from Evo 2 to Discovering New RNA with Minerva

Garyk Brixi (Stanford PhD student, team leader of Evo 2): In these two years, everyone has been using language large models trained based on "next token prediction" daily. Humans can read English, understand code, and interpret mathematical formulas, yet when it comes to staring at DNA sequences, they find themselves utterly baffled. But in a certain sense, DNA is a universal language written over millions of years of natural selection: DNA encodes RNA and proteins, which through extremely complex mechanisms determine whether an organism is healthy or how tall it grows, while natural selection in turn filters which DNA can survive and be sequenced.

In the past three years, the amount of genomic data in the NCBI public database has doubled. However, modeling DNA has proven extremely tricky because it has a minimal vocabulary of just 4 base pairs, filled with parasitic sequences and "redundant junk" that replicate wildly inside genomes, with size spanning enormous distances: a single gene can contain hundreds of thousands of base pairs, yet a mere single base mutation might be fatal.

Thus, we decided to adopt the most straightforward and unrefined approach: to build a broadly applicable DNA generative pre-training large model Evo 2 (with 7 billion and 40 billion parameter dense models, based on a mixed architecture of attention and convolution). A core secret often overlooked during the training process is: We ruthlessly discarded the vast majority of public and fetched junk data!

Only the best single representative genome of each species was preserved, filtering out highly similar redundancy, shrinking the total scale down to 90 trillion base pairs; in the pre-training主体阶段，我们 intentionally performed data resampling amplification around functional regions with existing genes, RNA and evolutionarily conserved elements, effectively discarding tens of trillions of ineffective noise tokens, while first mastering basic semantic structures in short-context and then extending to 1 million base pairs for long-window learning of combinatorial rules.

After lowering the loss function, how does one converse with a "model that only outputs A/T/C/G"? Two main approaches exist:

1. Treating it as an extremely sharp "life causal discriminant (Decision model)": The model has read the DNA of all life domains that have naturally survived, thus understands what sorts of sequences "align with natural laws." Just by inputting a segment of sequence with genetic mutations, if the model outputs an extremely low probability (it deems that code highly abnormal), it often strongly suggests a loss of function or risk of disease. In clinical evaluation tests for mutations in the high-risk cancer gene BRCA1, Evo 2 having only seen one set of human reference genomes performed comparably or even superiorly to human-specific supervised models in predicting mutations in coding and noncoding regions! Moreover, ablation experiments further confirmed that should we lazily retain 50% of random whole genome sequences, the Spearman correlation would plummet down to 0.16 and end at the bottom of the list, clearly indicating how deadly the data mixing ratio can be.
2. Using sparse autoencoders (Sparse Autoencoder) for explanaory dissection, unlocking all new gene generation: After opening up the Evo 2 internal neuron characteristics, we discovered that, merely by "predicting the next base," the model internally grew neurons recognizing features of prophages before Escherichia coli, bacterial defense CRISPR arrays, and single genes commandeered by bacteria; these segments are entirely distinct in base sequences, yet the model naturally linked their biological functions through unsupervised self-learning. After tuning with functional classifiers, the team generated a new type of PhiX-like phage from scratch, which is now entering human clinical testing.

For Evo 2’s past shortcomings in predicting "co-evolution and spatial interactions beyond distance," we recently released an innovative bidirectional model Minerva. It can directly scan and outline hundreds of entirely new non-coding RNA structures that traditional biostatistics missed within genomes of species like Bacteroidetes, using the model’s internal long-range attention weights to pinpoint them. In one of the recent wet experimental validations, we robustly discovered a new system called UG27 reverse transcriptase (UG27 RT): it carries a programmable non-coding RNA array similar to CRISPR that can convert DNA to RNA and then reverse transcribe it back to clamped-structure DNA. Although we haven’t fully discovered all its purposes in nature yet, this proves that pure DNA language models now possess solid capabilities to directly act as engines for human biological discoveries.

On-Site Questions: Why did you ultimately choose to tokenize at the smallest granularity of a single nucleotide? Have you considered using multi-base vocabularies like BPE?

Garyk Brixi: Currently, based on mid to small-scale controlled experimental results, cutting at the lowest granularity of single nucleotides has proven significantly better for accurately identifying single-point pathogenic mutations (Mutation prediction). Of course, long-term, I also look forward to exploring how byte-level end-to-end modeling can balance resolution and computation length.

On-Site Question: Have you considered a "dual-tower" model like embodied physical AI, where one tower reads DNA sequences, and the other tower outputs natural language interpretations to biologists?

Garyk Brixi: Indeed, many teams are racing toward several competing routes: some leading large model companies have mixed training of DNA sequences directly into language pre-training, while other researchers are embedding DNA models into vector-aligned mappings within large language models. As for which architecture will ultimately prevail, there are currently no conclusions.

On-Site Question: Since data screening and mixing ratio can yield such massive performance improvements, how do you scientifically identify the optimal "pre-training data ratio mixture"?

Garyk Brixi: Honestly, training Evo 2 at the end of 2024 was part of a "writing while exploring" pioneering batch; at that time, there were almost no predecessors’ pre-print parameters available for direct copying. Our core approach was to construct a closely aligned evaluation set (Evals) and scaling law analysis while processing data; now, as the evaluation sets have matured, we can more effectively reverse-optimize the data ratios based on assessment shortcomings.

On-Site Question: Given that you've identified so many biologically meaningful neuron characteristics within the model, have you attempted to use these internal features for directed guidance to steer Evo 2, such as designing DNA sequences specifically targeting cancer cells?

Garyk Brixi: Just last week, we published a few latest preprints specifically discussing "how to guide Evo 2 for biological design." To be honest, at this stage, merely relying on "mechanical interpretability neuron interventions" for directed biological design has yet to match the stunning capabilities of the "contextual few-shot prompting" scheme that the next speaker Divya will present.

On-Site Question: You mentioned during training that for each species, you essentially retained only one genome set, which means you might have discarded a lot of individual genetic differences within human populations. Yet you can still accurately predict pathogenic mutations for human cancers, what's the logic behind that?

Garyk Brixi: The reasoning is fundamentally consistent with classic evolutionary comparative statistics: even without deep learning, simply aligning the genomes of over a hundred primate species together, those base points that have remained virtually unchanged and highly conserved over long evolutionary scales typically represent core vulnerabilities; any mutations occurring there are likely to have substantial consequences.

In contrast, if we input all the personal genomes of the audience present, a massive amount of token computation would be wasted on repetitive sequence redundancies within humans, thus failing to solidify foundational pathogenic characteristics. When budget for token computation is limited, prioritizing feeding diverse and conserved signatures that span the entire tree of life yields a higher cost-performance ratio; of course, to capture the complexities and regulations specific to human diseases, the next step must integrate multi-omics data from populations alongside the individual single-cell expression data mentioned by the previous guest.

Divya Nori: No Fine-Tuning Needed! Directly Prompting DNA Large Models with 4 Examples, Lighting Up Mouse Brains Onsite

Divya Nori (Stanford Computer PhD student, researcher in Brian Hie’s lab): Hello everyone! I’m here to unveil what Garyk just teased: how to design entirely new biological components efficiently by merely "writing a few prompts to the genomic large model" without altering any weights or running thousands of fine-tuning data? This work is primarily led by postdoc Alex Hao, with contributions from Garyk and other colleagues in the Brian Hie lab.

Over the past few years, "few-shot prompting" has completely reshaped how we utilize text large language models: even without updating model parameters, simply inserting a few email templates into the context allows the large model to instantly grasp directive formats and implicit rules through emergent "in-context learning (ICL)."

However, in the biological sequence design field, traditional paradigms have long been stuck in another extremely data-dependent framework: for example, early papers required tens of thousands of CRISPR sites to perform full fine-tuning on Evo 1 to generate an active CRISPR-Cas9 system; similarly, generating a complete small phage genome also required thousands of microvirus family genomes for heavy fine-tuning on Evo 2. The harsh reality of biology is that many entirely new targets and rare protein families simply do not yield thousands of annotated training samples.

To our surprise, we recently discovered an almost magical emergent property in Evo 2 that had only appeared before in human natural language models: After merely large-scale single-nucleotide next-character predictions training, without anyone teaching it context learning, Evo 2 possesses a surprisingly strong few-shot context learning capability! This means you only need to provide it with 4 to over a dozen functional DNA sequence examples in the prompt, and the model can autonomously infer recombination and mutation algorithms, seamlessly generating a new sequence that shares the same functionalities.

We validated this new paradigm across four biological scales ranging from 100 nucleotides to 100,000 nucleotides:

1. Rare Natural Proteins (Lanmodulins): This is a revolutionary high-affinity rare earth binding protein used for rare metal mining recovery, with one particular family in nature having only 14 known samples! We directly provided these 14 sequences in the context prompts, and without any fine-tuning, Evo 2 successfully prompted the generation of an entirely new functional candidate protein sequence.
2. Antibody Targeting Design: We simply borrowed a few hit sequences generated based on three-dimensional structures as textual examples for Evo 2, successfully generating high-activity new hit sequences for all tested antibody targets without any input of three-dimensional protein structure conditions.
3. Reproducing Complex Gene Editing Systems (CRISPR-Cas9): Previously, tens of thousands of data points were needed for full fine-tuning; now, merely inputting 4 CRISPR-Cas site examples into the context allows Evo 2 to design an entire end-to-end system including the CRISPR array, tracrRNA, and Cas9 proteins all at once.
4. Complete Phage Genome Design: At a fraction of the computational cost of a full fine-tuning scheme, we could directly prompt the model to generate a new type of phage with the expected complete genome architecture.

But we are not stopping at software simulations; we are genuinely pushing this paradigm into animal wet experiments, aiming to design cell-type specific enhancers (the crucial core switch that determines whether gene therapy drugs can express precisely in certain brain regions or pathological sites without harming healthy organs).

As we push the enhancing prompting capabilities to the fullest, we also address the prior audience's concerns about "how to post-train Evo 2": we used a sequence-to-function prediction model (similar to AlphaGenome, which predicts whether a regulatory element will activate in certain cell types) as the reward function (Reward model), while introducing constraints to prevent gamification, controlling DNA sequence entropy and GC content, and performed post-training fine-tuning on Evo 2 through GRPO reinforcement learning cycles to amplify the model's ability to leverage context clues for prompts!

Subsequently, Alex injected the newly designed enhancers, powered by reinforcement learning, into the real mouse brains in the wet lab. The clear fluorescent microscopy images on the right side of the presentation verified that the specific targeted brain region successfully emitted clear and highly specific fluorescent spots!

Looking ahead, this "data-efficient conditioned control" will usher biological design thoroughly into the "lab in the loop" era: scientists holding a small pool of natural sequence examples can allow Evo 2 to rapidly output various new solutions via ICL, quickly filtering through in silico filters to select the optimal designs for wet testing, and then directly reintegrating the real validated sequences discovered back into the next prompt example pool for continuous snowballing.

Moreover, the inference enhancements and scaffolding techniques accumulated by human language models can seamlessly transfer into the world of biological design. All the content shared today has been included in the two preprints we just announced last week.

On-Site Questions: Could you elaborate on the reinforcement learning post-training mechanisms, reward functions, and anti-cheating mechanisms you designed for enhancers?

Divya Nori: The overall environment design is quite standard and efficient: First, we present known active natural enhancer examples in specific target cell types as prompts to Evo 2, letting it batch sample and generate a pool of new candidate sequences; subsequently, we send them through a sequence-to-function oracle model similar to AlphaGenome to predict specificity activation scores in various cell types for each new sequence as the reward signal for updating Evo 2's weights in the GRPO algorithm; at the same time, to prevent the model from taking shortcuts or "gaming" the reward during optimization, we’ve included regularization constraints in the rewards to ensure normal DNA sequence entropy and reasonable GC base content.

On-Site Question: What do you think is the reason for Evo 2's emergence of context learning ability? For adjacent structures like gene clusters, I can understand, but why does the generalized effect even occur in antibody sequence design?

Divya Nori: Honestly, we were quite surprised at first because the model had not undergone specific ICL training. There are several mainstream explanations in academia for the emergence of ICL in natural language large models, and they fit remarkably well in genomic data: first, natural sequences similarly conform to typical long-tail power law distributions and burstiness, where certain rare k-mer segments cluster around particular contexts; second, the genomes are naturally filled with significant semantic relatedness and parallel symmetrical structures.

From the internal mechanism analysis we performed in our preprints, we found that Evo 2’s attention layer has actually grown a mechanism almost identical to that in natural language models known as the "Induction head" to execute pattern recognition and contextual analogy.

On-Site Question: In the task of "context prompting + reinforcement learning to design enhancers," how much of this effect is due to the classic expression patterns already ingrained in the pre-training weights, and how much truly arises from the new patterns in prompts contributing to combinatorial creativity?

Divya Nori: This precisely illustrates the splendid trade-off between "weight-internalized priors" and "contextual on-the-fly combinations." Specifically for enhancers, their representation within Evo 2's extensive pre-training corpus is indeed extremely sparse; thus I believe, while the model does utilize underlying physical priors of sequence data, a far greater portion of the breakthrough specificity in expression regulation indeed comes from Evo 2's ability to instantly interpret the core motifs in the few new templates presented in prompts and directly precise them in the designs.

Francois Chaubard: Extremely exciting! Again, thank you Divya, and thanks to all the guests and friends here tonight!