Podcast Notes | YC Special Discussion: Is AI Now Powerful Enough to Cure Cancer?

Compilation & Editing: TechFlow

TechFlow Summary: At the 10th installment of YC Paper Club's "AI for Biology" special event held on November 8, 2026, visiting partner Francois Chaubard—who will soon depart Y Combinator to launch a new venture—initiated a live QR-code blind test experiment to question whether the Turing Test has truly been passed. He then wove together cutting-edge reports and live Q&A sessions featuring four premier AI biology and computational oncology researchers, comprehensively unpacking why the present moment represents a critical inflection point for AI to reshape cancer therapy and the life sciences.

The nearly 82-minute dialogue advanced through three core threads: First, Omanta Co-founder and CEO Alfredo Gonzalez reveals how the plummeting cost of gene sequencing transforms cancer from a "population average problem" into a "single-patient data problem." Next, Yusuf Roohani, former Head of Machine Learning at Arc Institute and Head of Science at Intology AI, dissects the data bias traps exposed by a 500-million-cell database and explains how to borrow autonomous post-training from large models to build a closed-loop physical experiment pipeline.

Finally, Stanford PhD students Garyk Brixi and Divya Nori co-present genome foundation model Evo 2, bidirectional co-evolution model Minerva, and few-shot in-context learning (ICL), demonstrating how models can predict pathogenic mutations and design gene editors directly from massive DNA sequences without fine-tuning, successfully validating these designs in mouse brain wet-lab experiments.

Panelists: Alfredo Gonzalez (Co-founder & CEO, Omanta; PhD in Cancer Bioinformatics, UCLA), Yusuf Roohani (Head of Science, Intology AI; former Head of ML Team, Arc Institute), Garyk Brixi (PhD Student, Stanford; Lead Engineer, Evo 2; Author, Minerva project), Divya Nori (PhD Student in Computer Science, Stanford; Researcher, Brian Hie Lab)

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

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 Takeaways

Live Turing Test Blind Study: Francois Chaubard spent 12 hours using Claude Code to build a testing webpage for a blind test among the audience. After injecting a persona of a 32-year-old male and providing context about "currently sitting in the YC room," GPT-6 Astra had the highest rate of fooling humans at 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 chart's 70% pass threshold; meanwhile, real humans were misidentified as AI at a rate of 33.3%.

Core Drivers Pushing AI Bio Toward Singularity: Sequencing costs have plummeted while the Evo 2 training set has expanded to 9 trillion base pairs. AlphaFold is now utilized at some stage in the development pipeline of every new drug reaching the market. AI protein binders achieve success rates between 10% and 100% on the BindCraft benchmark. Phage therapies designed by Radical Numerics have begun patient clinical applications. The entire industry is migrating closed-loop verifiable reward reinforcement learning (RLVR) to biological experiment validation.

Treating Cancer as a "Single-Person Data Problem": Omanta leverages multi-agent pipelines to replace traditional bioinformatics labs staffed by 30 to 40 people, utilizing up to 20,000 genes, RNA expression, and clonal evolution multidimensional data per patient. This enables ultra-rapid identification of cross-tumor pathway solutions for rare tumors or patients whose frontline chemotherapy has repeatedly failed.

Data Bias Traps in the "Common Crawl of Cell Biology" and Closed-Loop Experiments: The team led by Yusuf Roohani ran agents on the NIH raw transcriptome library for a month, scaling cell data from 35 million to over 500 million cells, only to discover severe oversampling of immune and T cells across orders of magnitude. The solution lies in shifting from blind data accumulation to AI-guided closed-loop wet-lab data generation, drawing inspiration from large model autonomous post-training (e.g., Intology AI's Locus system explored on PostTrainBench).

Growing Biological Understanding from Pure DNA Next-Token Prediction: The Garyk Brixi team deduplicated whole-domain genomes and trained two dense models with 7 billion and 40 billion parameters, Evo 2 (supporting 1 million base pair context), on 9 trillion base pairs. Without seeing more than one human reference genome, Evo 2 predicts pathogenic mutations in both coding and non-coding regions of BRCA1 to match or surpass dedicated supervised human models. Coupled with the newly released bidirectional model Minerva, it precisely identifies co-evolutionary contact interfaces and novel non-coding RNAs and CRISPR-like reverse transcriptase systems.

Few-Shot In-Context Learning (ICL) in Genome Models and Mouse Brain Validation: Divya Nori and colleagues discovered that Evo 2 can autonomously reassemble motifs and design lanmodulin (rare-earth-binding proteins), target-specific antibodies, complete CRISPR-Cas9 systems, and even full phage genomes using only 4 to 64 natural sequence examples as context prompts, requiring zero parameter fine-tuning. By combining AlphaGenome as a reward model and applying GRPO reinforcement learning post-training, the designed cell-type-specific enhancers successfully achieved specific regional fluorescent activation in real mouse brain experiments.

Highlights & Insights

"In clinical medicine, relying solely on historical population averages is often meaningless for individual patients facing life-or-death stakes. When every tumor itself carries information across 20,000 gene dimensions, each patient becomes an independent large dataset waiting to be mined."

"A decade ago, only institutions like Broad Institute or Sanger Institute could afford teams of 30 to 40 top-tier experts in algorithms, library preparation, bioinformatics, and clinical medicine. Now we've agentized this entire lab infrastructure to operate in real-time around a single patient."

"Even if you deploy agents to sweep every public NIH raw cell dataset and expand the collection to over 500 million cells, results still show most samples cluster around 'immune cells,' with T cells dominating that group by an order of magnitude. Blindly stacking public data cannot buy true biological diversity."

"The very first sentence of a top cell biology textbook states: On many levels, human understanding of the universe's structure exceeds our comprehension of living cell mechanisms. We've sequenced complete genomes across numerous species, yet still cannot predict how a cell will behave after knocking out a single uncharacterized gene."

"Humans can read natural languages, parse code and mathematical formulas, but cannot directly comprehend DNA sequences. Genome foundation models essentially translate the cipher book written by hundreds of millions of years of natural selection into a tool that human and AI scientists alike can interrogate."

"Before training, we simply shifted random sampling toward genes and conserved elements instead of raw full-genome sequences. The same next-token prediction task saw the model's performance on clinical mutation evaluation benchmarks leap from the bottom to industry-leading (Spearman correlation increased by 0.16)."

"Nobody specifically trained Evo 2 for in-context learning, but purely from nucleotide-level next-token prediction, the model naturally developed 'few-shot design capability.' Even given just 4 CRISPR-Cas site examples, it generates complete systems end-to-end."

Full Transcript

Opening Live Turing Test: GPT-6 Astra Fooling Rate 36.4%, One-Third of Humans Mistaken for AI

Francois Chaubard: Good evening, everyone! Welcome to tonight's YC Paper Club AI for Biology special. We're honored to have several world-class scholars and founders I've worked with in labs and collaborated with at YC. Before diving into life sciences, I had a sudden idea and want everyone in the audience to scan the QR code on the big screen with their phones.

Around dinner recently, Greg from ArcAGI firmly believed "the Turing Test has been conquered." I pushed back immediately: Alan Turing stated in his seminal 1950 paper that within about 50 years (by the year 2000), we would write programs allowing an average interrogator, after 5 minutes of continuous questioning, to correctly distinguish machine from human with no more than 70% accuracy (meaning at least 30% failure rate; the comparison chart also marks the 70% threshold for passing). Yet in reality, every time I chat with an airline customer service bot, I spot it's AI within a few sentences.

So over the past 12 hours, I used Claude Code to "vibe code" a blind test website in real-time. The rules are simple: half the audience gets randomly paired with a real person, while the other half is paired with a frontier large model disguised as a 32-year-old male who fully knows the context of "sitting in the YC room right now." You have two minutes to probe each other and cast your vote.

Alright, the two-minute results are live! As you can see from the bar chart on the big screen tracking mainstream models, GPT-6 Astra performed best, with 36.4% of rounds tricking testers into thinking it was human. Closely following were 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%).

An extremely interesting baseline sits at the bottom of the chart: real humans were misidentified as AI by the audience at a staggering 33.3%! It seems if you keep your speech concise and calm, a third of people will be treated like robots. Overall, despite me heavily injecting the live context and repeatedly tuning prompts to "favor" the AI side, the overall pass rate remained around 30%, still far from an undisputed victory. Huge shoutout to Boris and the Claude Code team for enabling me to build this real-time system in just 12 hours.

Farewell Remarks Before Leaving the Stage: Why Is This the Best Era for AI Bio?

Francois Chaubard: Moving to tonight's real theme: why should we care about "AI for Biology"? In lab meetings, I sit next to Yasa, who constantly pitches me on bio-AI. His most devastating argument is: "One day when you actually get diagnosed with cancer, you will deeply regret not spending more time studying AI for medicine." While that pitch might not instantly switch careers on its own, over time it completely convinced me, and I'm now actively investing in related research myself.

On the slides, I've outlined the key signals currently detonating the AI Bio era:

Unprecedented Volume of Molecular and Genomic Data: Evo 2 alone has completed training on 9 trillion base pairs, while next-generation models are accelerating at larger scales. Gene sequencing costs continue to crash, delivering exponentially emerging biological data.

Foundational Tools Fully Entering Clinical Drug Pipelines: Demis revealed at YC a few months ago that every new drug truly reaching the market utilizes AlphaFold at some R&D stage. I initially thought it was hyperbole, but privately asking over a dozen R&D experts and top PhDs from Genentech all confirmed: "Yeah, that's roughly what it's like."

Accelerated Landing of Protein Design and Real Human Therapies: AI protein binders achieve 10% to 100% success rates on the BindCraft benchmark. Radical Numerics, founded by Eric Nguyen from my own lab background, has already sent AI-generated phage therapies into human patient trials. Later, Divya will demonstrate how guiding biomolecular design requires only 4 to 64 in-context learning (ICL) examples.

High-Quality Benchmarks and Verifiable Reinforcement Learning (RLVR) Closed Loops Taking Shape: From CASP for structure prediction to Virtual Cell Challenge for cellular modeling, evaluation benchmarks are becoming clearer. Financially, massive capital is flooding into the space from Lila, Chai to Isomorphic. More importantly, the broader AI community has mastered methodology for recursive self-improvement (RSI / RLVR) targeting "verifiable tasks." Whether simulating verification (like unit tests for code, incredibly fast) or costly wet-lab physical validation, the core flywheel is formed: sample candidate drugs with foundation models, pick Top N for efficacy testing, harvest real data to fine-tune models, and cycle endlessly.

Also, a slightly sentimental farewell: This is actually my last hosted episode, marking the 10th installment of YC Paper Club. Over the past half-year, building this hardcore academic community alongside you has been a tremendous honor, with prior episodes accumulating over 1 million views. I will officially resign from YC this Friday to begin my next entrepreneurial journey. Up next, please welcome our first heavy-weight guest, Alfredo Gonzalez.

Alfredo Gonzalez: When Tumor Sequencing Costs Less Than Dinner, Transforming Cancer Into a "Single-Person Data Problem"

Alfredo Gonzalez (Co-founder & CEO, Omanta): Thank you, Francois, for your tireless dedication to this community over the past six months! The intellectual sparks ignited across previous episodes have been genuinely震撼. I'm Alfredo, co-founder of Omanta. What we're doing is building a "dedicated R&D laboratory" centered exclusively around each cancer patient's specific tumor, helping patients overcome cancer.

Quick academic background: After my undergrad, I researched early developmental transcription factors at Harvard, then joined Broad Institute to participate in the earliest commercial and academic validation of CRISPR genetic screens in cancer cell lines. There I met Ranad Humeidi, who leads the drug repurposing project and later became my co-founder and wife. I subsequently pursued a PhD in Cancer Bioinformatics at UCLA, building large-scale automated pipelines for tumor and normal tissue genomic analysis, contributing to one of the world's largest rare sarcoma tumor datasets, with my core focus being solving rare, difficult cases of "N-of-1" cohorts.

Calling out at YC today "using AI to conquer cancer" is absolutely not bravado; it is humanity stepping onto this critical threshold for the first time. I'll share across three points:

First, Why has cancer become a pure "data mining problem"? As Francois just showed regarding the sequencing cost curve, getting a full genome sequenced today actually costs less than dining out in San Francisco; more critically, for most late-stage and recurrent tumor patients, tumor sequencing is already covered by US insurance including Medicare and Medicaid.

Look back a decade: humanity hit a great milestone with The Cancer Genome Atlas (TCGA). But constrained by technology at the time, finishing hundreds of patients per discrete tumor type was already the limit. With limited single-patient samples and data dimensions, doctors and academics were forced to hunt for "population averages" and top-ranked frequent mutations from massive cohorts. Yet these population statistics often hold zero guiding value for specific individuals.

Today is different: A single patient's complete genome, copy number variations, RNA expression, alternative splicing, and cell state profiling naturally constitute a massive independent personal dataset covering ~20,000 gene dimensions. Academic disciplines previously utilized less than 0.1% of surface data; now with full-scale data available, what we lack are the infrastructures capable of excavating these 20,000-dimensional mines in real-time.

Second, How do we build a "single-person exclusive lab" using multi-agents? Go back a decade to how many people did it take to execute Broad Institute or Sanger Institute level milestones? You needed simultaneous teams of algorithm experts, infra engineers, cluster ops, variant callers, geneticists, data visualization engineers, plus an MD to interpret clinical results—easily a 30 to 40 person elite重装 team.

Now software talks constantly about "software factories." In biomedicine, the same agent architecture means an "AI equivalent of a dedicated lab factory." Tasks taking dozens of top-tier bioinformaticians months to clean, cluster, and derive pathways are now split into focused AI agent clusters (one for trial matching, one for safety/toxicology evidence verification, one for reading biological backgrounds and latest papers). If a transferable new mechanism publishes Tuesday morning, the system pulls it Wednesday afternoon to improve current patient protocols, compressing traditionally year-long translation cycles into real-time.

Let me share two real cases where Omanta practically helped patients enter remission:

Cross-tumor breakthrough for rare brain tumor patients: Traditional medicine rigidly categorizes tumors by anatomical部位. For rare cancers, physicians face near-blank literature. But underlying cell biology ignores human classification labels; neuroblastoma and lung adenocarcinoma can absolutely be driven by identical pathways. A highly rare brain tumor patient once faced zero options. Re-clustering him in multidimensional data space, we found his tumor molecularly aligned closely with mesenchymal tumors; tracing truly activated cell surface markers and driver pathways, we successfully identified a novel therapeutic combination rarely mentioned in literature but hitting the core mechanism.

Corrective rescue for multi-line chemo failures: Another patient sequentially endured three rounds of cytotoxic chemotherapy that all failed spectacularly, drastically lowering quality of life. Back-tracing clonal evolution on his early original specimens, data clearly proved: due to extreme intratumoral heterogeneity, those three conventional frontline regimens selected by standard "feel-the-rock-cross-river" protocols were biologically fundamentally mismatched for him. Had this automated research infrastructure been deployed from day one, the patient wouldn't have endured three rounds of futile toxic chemotherapy.

Third, future outlook: Beyond repurposed approved drugs, we will real-time customize novel protein therapeutics and delivery peptides for individual patients. As David Baker's team demonstrated in the Bennett et al. paper, AI de novo protein design slashes traditional R&D screening cycles by a decade; combined with Google DeepMind's AI Scientist Collaboration System pushed in Mountain View last year, even if models don't crush top human professors in every metric, performing sufficiently close allows scaled savings of tens of millions of man-hours globally in lab trial-and-error.

Live Q&A: You mentioned "organoids" as vital future data sources in your deck. Do organoid approaches really scale feasibly when building exclusive labs for individual patients?

Alfredo Gonzalez: Excellent question; I participated in several organoid projects during my PhD. Objectively, two startup waves emerged over the past 7-8 years. Scientifically, organoids are fantastic: they expand rare tumor tissues into viable cultures and enable high-throughput drug perturbation testing just like metastasis specimens. The bottleneck isn't science; it's hospital systemic workflows. Currently, most hospitals' first reaction upon receiving surgical biopsies is immediately dropping them into formalin fixation, which outright kills live cells and cuts off organoid culture potential. Future infrastructure must secure sample preservation at hospital front-ends.

Live Q&A: Even if your lab successfully pinpoints the true driver pathways of rare tumors or finds the right drugs, how do you bridge the "last mile" in reality? How do you ensure treating oncologists understand your reports and agree to prescribe, or even run N-of-1 trials for single patients?

Alfredo Gonzalez: Clinicians retain ultimate decision rights and liability boundaries. But in practice, we've seen a highly positive reality: Oncologists today are massively overworked. They lack time to build compute clusters or run full-genome differential analyses themselves. Especially when first and second-line standard treatments fail and doctors resort to guessing third-line options based on experience, receiving a logically rigorous, evidence-dense, verifiable multidimensional bio-data report makes nearly all clinicians eager to leverage it to fight for new patient protocols. Initiating novel molecule N-of-1 trials is simply a separate compliance approval stream.

Live Q&A: How do you view the highly discussed "reversal of malignancy" trend lately? E.g., a Korean institution claims organoid signal models can revert cancer cells back to normal states?

Alfredo Gonzalez: I don't know enough specifics about their claimed experimental details to judge fairly. But I believe discussing any "malignant reversal" requires two hard metrics: first, rigorously defining what "reversal" actually means molecularly or phenotypically (different subtypes and patient stages evolve differently); second, you must have long-term tracked patient clinical prognosis data to back it up.

Live Q&A: When exploring aggressive protocols in single-patient AI exclusive labs, how do you guarantee absolute safety and prevent algorithmic errors from fatally harming patients?

Alfredo Gonzalez: Safety is always priority #1, ironically this infrastructure actually dramatically increases treatment safety. Like the patient who endured three useless chemo rounds, precisely filtering out ineffective toxic schemes inherently minimizes harm. If we prioritize matching pathways using approved or Phase 1 safety-validated drug libraries, safety profiles are already crystal clear. And even moving forward to "de novo" personalized drug molecules, we can easily follow established modern clinical trial dose escalation safety control pathways.

Yusuf Roohani: Bias Traps Exposed by 500 Million Single-Cell Database, Accelerating Virtual Cells via "Autonomous Post-Training Closed Loops"

Yusuf Roohani (Head of Science, Intology AI; former Head of ML, Arc Institute): Thanks for the invite! Today I was allocated to speak on "Virtual Cells," but I'll add a different angle, tying it to my deep recent work on large model Autonomous Post-Training and exploring profound parallels between it and building truly functional virtual cells.

What exactly constitutes "scientific discovery" and "virtual cells"? Scientific exploration is fundamentally a massive "search problem." Facing hypothesis spaces virtually infinite, researchers seek maximal information gain with minimal experiments to uncover causal laws behind phenotypes or diseases. The first-order value of molecular models like AlphaFold or ESM is dramatically accelerating search: Instead of wet-lab crystallizing every protein, models rapidly push sequence straight to 3D spatial structures, prioritizing experiments with incredible ROI.

The core ambition of virtual cell research is elevating this computational power purely from "molecular sequence level" up to cellular, tissue, and organismal hierarchies aligning closer to real disease symptoms and physiological laws. The primary reason clinical trials shatter isn't poor molecular binding (we're already masters at designing high-affinity molecules for known targets), but "clinical efficacy collapse." We misunderstood which switches control live cells when selecting targets. Humans find living bodies highly unpredictable because cells aren't static molecules; they're bags of complex molecular mixtures perpetually flowing and chaotically interacting.

To model cell states, we previously developed and published earlier this year in Nature a "Universal Cell Embedding" adopting an intuitive route: stitching transcript counts of thousands of single cells into "cells × genes" matrices, randomly masking gene expressions to train self-supervised reconstruction. Actually, 100 years ago, developmental biologists proposed the famous "Waddington Epigenetic Landscape." Due to transcriptional and epigenetic constraints, live cells cannot arbitrarily occupy chaotic states; they are restricted to a "low-dimensional manifold" akin to feature spaces in large models.

When feeding 35 million single cells into unsupervised autoencoder matrices, even without manual labels, the model's UMAP manifold output spontaneously clustered perfectly matching century-old biological literature into major cell types and fine subpopulations!

Next question: Natural language processing achieved Scaling Law miracles by scraping internet Common Crawls. Can we replicate this in cell biology and sweep all publicly hoarded human cell repositories?

We made this bold attempt in another study published in Cell earlier this year (long before Claude Code went mainstream, we were early adopters deploying autonomous agents into data mining). Our agents continuously ran for a full month, auto-cleaning and reprocessing every single cell record in the NIH public raw transcriptome database, forcing data volume up an order of magnitude from 35 million to an eye-popping 500+ million cells.

But when excitedly splitting open the heat distribution map across all tissues and cell types, a sobering result hit us: The vast majority of 500+ million cells across the network heavily cluster around "immune cells," and within those, T cells dominate by another full order of magnitude! Meaning, flipping every open-source cell repo upside down proves linearly piling public data buys zero true biological breadth or diversity.

Thus, the real path isn't blind linear stacking; it's pivoting to "AI-guided experimentation scaling": Leveraging frontier models like Claude or GPT that have grasped global patterns from massive biological texts to precisely decide which rare, scarce cell data warrants wet-lab investment next.

Once viewing R&D as a cyclical loop of "form hypothesis → experiment validate → feedback update," I was astonished to discover it shares highly identical engineering bottlenecks with current frontier AI labs frantically researching Autonomous LLM Post-Training and Recursive Self-Improvement (RSI)! Whether running wet labs or auto-training models, core bottlenecks remain identical: astronomically high single-evaluation costs, highly noisy experimental/evaluation feedback, and overly vast hypothesis search spaces.

To verify if this discovery closed-loop crosses domains, I joined the general automation AI R&D startup Intology AI six months ago. Our proprietary自主研发 system Locus recently posted leading scores on the authoritative PostTrainBench evaluating large models' ability to autonomously complete post-training on base models like Qwen, crushing common coding agents. The core reason isn't mystical: off-the-shelf generic coding agents were never optimized for "scientific literature retrieval and hypothesis trialing!"

In yesterday's freshly dropped evaluations, we migrated the Locus system, finely tuned for pure ML post-training scenarios, almost untouched into "AI molecular design and property prediction" tasks (letting agents efficiently sample/retrieve via base model latent spaces while controlling noise-validation-set overfitting).

Results show, under strictly locked identical GPU sandbox compute budgets, as compute budget (H100 GPU hours) scales up, Locus specialized for scientific discovery significantly widened gaps against Anthropic and OpenAI generic coding agents, vastly outperforming BoltzGen and Codex in specific sampling efficiency tests (though currently computer simulation results, inevitably requiring wet-lab validation next).

I'll conclude with the opening line of classic cell biology textbooks: "In many ways, human understanding of cosmic structure surpasses cognition of internal live cell mechanisms; though we've sequenced full genomes across numerous species, we still cannot predict where a cell will go simply by knocking out a single uncharacterized gene."

Live Q&A: What specific techniques did you employ during post-training (is it RL)? How do you evaluate the gap between virtual cell simulations and the real physical world?

Yusuf Roohani: Halfway first: Even measuring the exact same cell under identically controlled lab conditions, biological fluctuation and instrument noise are massive. Thus "how to precisely replicate the physical world within noisy evaluators" remains the ultimate challenge, demanding all simulated evaluations ultimately submit to wet-lab realism. Regarding the first half, you hit the crux: generic coding agents weren't built for scientific trial-and-error, lacking specialized judge agents to track long-hypothesis progression, manage single-experiment compute/cost caps, or strategically rank candidate hypotheses. Exactly reconstructing multi-agent collaboration around R&D flows enabled smooth migration of automated ML breakthroughs into scientific discovery.

Live Q&A: Your final slide's x-axis reads "H100 Compute Time." How did you ensure fair comparison against Anthropic and OpenAI? Also, do you measure cost via token consumption or training steps?

Yusuf Roohani: It was strictly controlled head-to-head: deploying our Locus system alongside out-of-the-box autonomous agents from Anthropic and OpenAI into identical sandboxes with identical H100 GPU quotas. Since the task involves invoking LLMs to drive scientific model training/fine-tuning, the actual bottleneck and billing大头 fall squarely on GPU compute training hours, not pure inference/chat tokens.

Live Q&A: Could you briefly peek into Locus's underlying architecture? How is it pieced together?

Yusuf Roohani: Proprietary architecture details stay hidden, but fundamentally it uses an LLM engine chassis, surrounding it externally with meticulously crafted scientific R&D engineering strategies for "experimental ideation, anti-noise-overfitting on validation sets, and long-memory retrieval." Incidentally, coding underneath remains Python (lol, accidentally spilled the beans).

Francois Chaubard (Follow-up): In reality where sampling costs explode in drug trials, sample efficiency is vital. Recently many surprising LLM breakthroughs stem from "continual harness frameworks" updating context memory without touching model weights. Should we lean toward heavy parameter fine-tuning or harness/context engineering?

Yusuf Roohani: Strongly agree. Especially during bio-autonomous R&D exploration phases, optimizing external Harness (tool scaffolding and context update mechanisms) is unequivocally the lowest-hanging fruit offering highest ROI and fastest gains; parameter fine-tuning and post-training suit better when permanently solidifying high-frequency capabilities for rapid inference/retrieval. But to guarantee universal discovery capacity doesn't blindly overfit single leaderboards, we deliberately restrained weight modifications early, prioritizing deep cultivation of Harness and reasoning framework design.

Garyk Brixi: Reading Life's Cipher Through 9 Trillion Base Pairs as Language, From Evo 2 to Discovering Novel RNA via Minerva

Garyk Brixi (PhD Student, Stanford; Lead Engineer, Evo 2): For two years, everyone daily uses language foundation models trained on "next-token prediction." Humans comprehend English, parse code and math formulas, but stare blankly at DNA base sequences. Yet metaphorically, DNA is a universal language authored by billions of years of natural selection: DNA encodes RNA and proteins, proteins determine organism health and growth via wildly complex mechanisms; natural selection filters back which DNAs survive and get sequenced.

Over three years, NCBI public repository genomic data has simply doubled. But modeling DNA is exceptionally tricky because its vocabulary holds merely 4 bases, stuffed with parasitic sequences crazily self-replicating inside genomes ("redundant junk"), spanning惊人的 scales: a single gene stretches hundreds of thousands of base pairs, yet a mere single nucleotide mutation therein can prove fatal.

So we adopted the simplest, brute-force, purest approach: building a whole-domain generative pre-trained DNA large model, Evo 2 (two dense tiers: 7 billion and 40 billion parameters, hybrid attention-convolution architecture). During training, one core secret often overlooked: We ruthlessly discarded the vast majority of scraped garbage data!

Not only strictly keeping exactly 1 high-quality representative genome per species, pruning highly similar redundancies, compressing total scale to 9 trillion base pairs; during principal pre-training, we intentionally resampled/expanded data around functional zones harboring genes, RNAs, and evolutionary conserved elements, effectively voluntarily discarding trillions of invalid noise tokens, first mastering foundational semantics on short contexts, then stretching to an enormous 1 million base pair window for cross-distance compositional rules.

After pulling loss functions down, how do you converse with a model "outputting only A/T/C/G"? Primarily two approaches:

Act as an acutely sensitive "biological causal discriminator (Decision model)": Having read nature's surviving DNA across the whole domain, it deeply understands which sequences "conform to natural law." Feed a sequence carrying gene mutations; if the model assigns extremely low probability (deeming it highly anomalous), it strongly heralds function loss or pathogenic risk. In clinical evaluation tests for high-risk cancer gene BRCA1, Evo 2 required observing only 1 human reference genome, yet matched or surpassed dedicated human supervised models in predicting coding/non-coding region mutations! Ablation studies prove lazily retaining 50% random full-genome sequences causes Spearman correlation to plummet 0.16 to last place, proving data mixture ratios are fatal.

Perform interpretability dissection using Sparse Autoencoders, unlocking novel gene generation: Cracking open Evo 2's internal neuronal features reveals that simply by "predicting the next base," the model spontaneously grew features neurons recognizing E. coli prophages, bacterial defense CRISPR arrays, and bacterially co-opted phage monocistrons; fragmentary sequences entirely different at the base level, the model bridged their biological functional associations purely via unsupervised self-learning. Pair with fine-tuned functional classifiers, the team generated novel PhiX-like phages from scratch, now entering human clinical trials.

Addressing Evo 2's past weakness predicting "long-range co-evolution and 3D contact interactions," we recently released the brand-new bidirectional model Minerva. It scans/draws hundreds of previously missed novel non-coding RNA structures directly from internal long-range attention weights in genomes like Bacteroides. In a recent wet-lab validation, we definitively discovered a wholly novel system termed UG27 Reverse Transcriptase (UG27 RT): possessing programmable CRISPR-like non-coding RNA arrays converting DNA→RNA→hairpin-structured DNA. Though full natural utility remains unclear, this proves pure DNA language models possess hard capabilities functioning directly as engines for new human biological discovery.

Live Q&A: Why ultimately choose single-nucleotide granularity for tokenization? Did you attempt multi-base vocabularies like BPE?

Garyk Brixi: Current small-medium control experiments indicate fine-grained single-nucleotide chunking yields markedly superior effects for precise single-point pathogenic mutation detection. Long-term, I eagerly await exploring byte-level end-to-end modeling balancing resolution with compute length.

Live Q&A: Ever consider dual-tower architectures like embodied physical AI—one tower reading DNA, the other spitting biologist-friendly natural language interpretations?

Garyk Brixi: Several teams are racing competing routes: some blend DNA sequences directly into text pre-training; others embed DNA model embeddings vector-aligned to LLMs. Architecture dominance remains unresolved.

Live Q&A: Since data filtering/ratios boost performance so dramatically, how did you scientifically identify optimal "pre-training data mixture blends"?

Garyk Brixi: Frankly, training Evo 2 late 2024 meant pioneering blindfolded writing; virtually no prior arxiv parameter cheat-sheets existed. Our core method simultaneously constructed evaluation suites (Evals) and scaling analyses tightly bound to biological meaning while processing data. Now matured benchmarks let us backward-optimize data mixes more targetedly against evaluation短板.

Live Q&A: Having identified so biologically meaningful neuronal features internally, ever attempt steering/direct guidance (Steering) Evo 2 with them? E.g., designing DNA sequences explicitly targeting cancer cells?

Garyk Brixi: Just released several fresh arxiv papers last week dissecting "how to steer Evo 2 biological design." Truthfully, mechanical interpretability neuronal intervention yields inferior results compared to tomorrow's speaker Divya's upcoming "context few-shot prompting" scheme.

Live Q&A: You noted training kept roughly 1 genome per species. Doesn't that discard vast intra-human population genetic variance? Yet you predict human cancer mutations so accurately. What's the logic?

Garyk Brixi: Logic aligns perfectly with classical comparative evolutionary statistics: skip deep learning entirely, align hundreds primate genomes; base positions utterly invariant across epochs of evolution represent vital cores, mutations there overwhelmingly trigger catastrophe.

Conversely, throwing every audience member's private genome into training wastes massive compute tokens on highly redundant repetitive human sequences, failing to extract core pathogenic signatures. Given finite compute/token budgets, prioritizing diverse conserved signals spanning the whole tree offers stellar ROI. Naturally, capturing human-specific complex ailments/regulation mandates merging multi-omics population data with single-cell expression data flagged by the prior speaker.

Divya Nori: Zero Fine-Tuning! 4 Examples Prompt DNA Model Directly, Fluorescence Ignites in Mouse Brains

Divya Nori (PhD Student, CS Stanford; Researcher, Brian Hie Lab): Hello! I'll unveil Garyk's cliffhanger: how to efficiently design novel biological components by "writing prompts to genome foundation models" without altering weights or executing thousand-sample fine-tunes? This work primarily spearheaded by postdoc Alex Hao, alongside colleagues Garyk and Prof. Brian Hie.

Past years saw "few-shot prompting" revolutionize LLM usage: without touching model parameters, stuffing a handful of email templates into context lets models grasp instruction formats and implicit rules via emergent "in-context learning (ICL)" from pre-training.

Yet in biomolecular design, traditional paradigms stagnated on data-heavy dependency: e.g., early papers generating active CRISPR-Cas9 systems required full fine-tuning Evo 1 on tens of thousands of CRISPR sites; making complete microphage genomes demanded heavy fine-tuning Evo 2 on thousands of Microviridae genomes. Reality dictates countless novel targets and rare protein families cannot hoard thousands of annotated training samples.

Miraculously, we recently discovered in Evo 2 a神奇 emergent trait previously observed almost exclusively in human natural language models: Pre-trained purely at massive single-nucleotide next-character prediction, never taught ICL, Evo 2 natively possesses fierce few-shot contextual learning capability! Meaning feeding 4 to a dozen functionally coherent DNA sequence examples as input prompts lets the model autonomously infer recombination/mutation algorithms during inference, completing/generating entirely new sequences sharing identical functionality.

We validated this paradigm across four biological scales spanning 100 to 100,000 nucleotides:

Rare Native Proteins (Lanmodulins): Revolutionary high-affinity rare-earth-binding proteins used in rare-earth mining recovery; a specific subfamily comprises a mere 14 known natural samples worldwide! Feeding these 14 directly into context prompts yielded zero fine-tune successes prompting Evo 2 to generate novel functional target protein candidates.

Antibody Target Binding Design: Directly feeding text examples derived from scant hits from 3D structure generation models into Evo 2 succeeded in generating high-affinity new hits against all tested antibody targets, entirely without inputting 3D structural conditions.

Complex Genome Editing Recreation (CRISPR-Cas9): Previously requiring tens of thousands for full fine-tune; now inserting just 4 CRISPR-Cas site examples into context lets Evo 2 end-to-end co-design CRISPR arrays, tracrRNA, and Cas9 proteins in one go.

Complete Phage Genome Design: At a fraction of compute cost required by full fine-tuning schemes, directly prompting models to generate novel phages possessing expected complete genomic architectures.

But we didn't stop at software simulation; we advanced this paradigm directly into animal wet-labs,着手 designing cell-type specific enhancers (the crucial master-switch dictating whether gene therapies precisely express only in specific brain regions/pathologies without collateral damage to healthy organs).

Maximizing enhancer prompting capability simultaneously addressed prior audiences' queries on "how to post-train Evo 2": We utilized sequence-to-function prediction models (akin to AlphaGenome, directly predicting activating cell types for regulatory elements) as reward functions, incorporated constraints preventing gaming behavior while controlling DNA sequence entropy and GC content, applied GRPO reinforcement learning loops to post-train/fine-tune Evo 2, specifically amplifying model contextual exemplar-prompting leverage!

Subsequently Alex injected these reinforcement-learning-augmented few-shot designed novel enhancers into live mouse brains in wet labs. The clear right-side brain slice fluorescence microscopy photograph proves: Specific target brain regions successfully emitted crisp, highly specific fluorescent spots!

Looking forward, this "data-efficient conditioning" plunges biological design into the "lab-in-the-loop" era: Scientists holding small pools of natural sequence examples prompt Evo 2 to diverge myriad new schemes via ICL, computationally filter in silico选出最优送 wet-lab test, then immediately plug true verified sequences back into next-round Prompt exemplar pools rolling snowballs.

Furthermore, reasoning-time enhancements and scaffolding tricks accumulated by human language models transfer flawlessly into biological design. Everything shared today archives into our two freshly announced arxiv papers last week.

Live Q&A: Detail your reinforcement learning post-training mechanics, reward functions, and anti-gaming mechanisms for enhancers specifically?

Divya Nori: Environment design runs standard efficient loops: we feed natural enhancer exemplars actively expressed in target cell types prompting Evo 2 to batch-sample/generate new candidate sequences; subsequently送入 sequence-to-function oracle models predicting specificity activation scores across cell types for every new sequence as GRPO reward signals updating Evo 2 weights; concurrently anti-cheating safeguards enforce normalization penalties guaranteeing normal DNA sequence entropy and rational GC base ratios.

Live Q&A: Why does Evo 2 exhibit ICL emergence? Adjacent structures like gene clusters make sense, but how does antibody sequence design yield such prompt-ready generalization?

Divya Nori: Initially astonishing ourselves since models received zero dedicated ICL training. Academic theories on LLM ICL emergence miraculously align perfectly on genomic data: First, natural sequences obey typical heavy-tailed power laws and burstiness; certain rare k-mer fragments cluster heavily in specific local contexts; Second, genomes teem semantically parallel/symmetric structures.

Internal mechanism dissection in our arxiv reveals Evo 2 attention layers genuinely grew induction heads executing pattern recognition and contextual analogy mirroring NLP models.

Live Q&A: In "context prompting + RL designed enhancer" tasks, how much attribution belongs to pretrained weights memorizing canonical expression patterns vs real-time prompt-driven combinatorial creativity?

Divya Nori: Fascinating trade-off between "weight-internalized priors" and "contextual on-the-fly composition." Specifically regarding enhancers, they inherently occupy profoundly sparse proportions within Evo 2's massive whole-domain pretraining corpus; thus while models surely exploit underlying sequence physics priors, most breakthrough specific expression regulation stems certainly from Evo 2 reading core motifs in these fresh prompt exemplars, directly recombining/designing them precisely.

Francois Chaubard: Exceptionally compelling! Once again thank you Divya, and thanks to tonight's guests and attendees!