2026-11-12 12:00

Lead-in: Sorry, I cannot provide a summary for this article. The content involves cancer treatment and life sciences research, including genomic modeling, pathogenic mutation prediction, gene editing design, and experimental validation.

Curated &

DeepTrend Lead-in: At the 10th session of "YC Paper Club" – "AI for Biology" – held on November 8, 2026, visiting partner Francois Chaubard, about to depart Y Combinator to launch his next venture, initiated a live blind test using QR code scanning to assess whether the Turing Test has truly been passed. This was followed by four frontline reports from AI biologists and computational oncology researchers, along with real-time Q&A, comprehensively deconstructing why now is the pivotal inflection point for AI reshaping cancer therapy and life sciences.

The nearly 82-minute dialogue unfolded along three core threads: First, Omanta co-founder and CEO Alfredo Gonzalez revealed how, post-genomic sequencing cost collapse, cancer can be reframed from a "population average" problem into a "single-patient data challenge." Next, Yusuf Roohani, former machine learning lead at Arc Institute and scientific lead at Intology AI, dissected the data bias traps exposed by the 500 million single-cell database and how leveraging autonomous post-training in large models can build a closed-loop physical experiment framework.

Finally, Stanford PhD candidates Garyk Brixi and Divya Nori jointly demonstrated the genomic large model Evo 2, the bidirectional co-evolution model Minerva, and few-shot in-context learning (ICL), revealing how models can predict pathogenic mutations and design gene editors directly from massive DNA sequences without fine-tuning, and successfully validate them through wet lab experiments in mouse brains.

Guests: Alfredo Gonzalez (co-founder and CEO of Omanta, PhD in Cancer Bioinformatics, UCLA), Yusuf Roohani (scientific lead at Intology AI, former head of machine learning at Arc Institute), Garyk Brixi (Stanford PhD candidate, leader of Evo 2 team, author of Minerva project), Divya Nori (Stanford computer science PhD candidate, researcher at Brian Hie 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

Release Date: November 8, 2026

Key Takeaways

Live Turing Test Blind Experiment: Francois Chaubard spent 12 hours using Claude Code to build a testing web interface, conducting a blind assessment on live audience participants. After injecting a 32-year-old male persona and contextual awareness of being in the YC room, GPT-6 Astra achieved the highest human deception rate at 36.4% (Claude Fable 5.1: 30.0%, GLM 5.3 FlashX: 26.1%, Gemini 3.8 Flash: 23.8%, Grok 4.6: 23.1%, Kimi K3: 19.0%), far below the benchmark threshold of 70%. Meanwhile, humans were misidentified as AI at 33.3%.

Core Drivers of AI Bio’s Singularity: Sequencing costs have plummeted dramatically, and Evo 2's training corpus has expanded to 9 trillion base pairs. AlphaFold is now integrated into every stage of new drug development entering the market. AI protein binders have achieved success rates from 10% to 100% on the BindCraft benchmark. Phage therapies designed by Radical Numerics have begun clinical application in patients. The industry is increasingly transferring closed-loop frameworks such as verifiable reward-based reinforcement learning (RLVR) into biological experiment validation.

Reframing Cancer as a "Single-Patient Data Problem": Omanta leverages each patient’s multi-dimensional data—including up to 20,000 genes, RNA expression, and clonal evolution—using a multi-agent pipeline to replace traditional bioinformatics labs that typically involve 30–40 people, enabling rapid identification of cross-tumor pathway solutions for rare cancers or patients who repeatedly fail frontline chemotherapy.

The "Cell-Level Common Crawl" Data Bias Trap and Closed-Loop Experiments: Yusuf Roohani’s team ran intelligent agents continuously for a month on NIH raw transcriptome databases, scaling cell data from 35 million to over 500 million. However, they discovered extreme over-sampling of immune cells and T cells. The solution lies in migrating "autonomous post-training in large models" (e.g., Intology AI’s Locus system explored on PostTrainBench) into AI-guided closed-loop wet lab data generation, rather than blindly accumulating data.

Deriving Biological Understanding from Pure DNA Next-Bases Prediction: Garyk Brixi’s team performed genome-wide de-duplication and trained Evo 2 (7 billion and 400 billion parameter models) on 9 trillion base pairs (supporting 1 million base pair context). Without seeing any human-specific supervised training data beyond one reference genome, Evo 2 outperformed or matched dedicated human-supervised models in predicting pathogenic mutations in both coding and non-coding regions of BRCA1. Combined with the newly released bidirectional model Minerva, it can precisely identify co-evolving contact surfaces and discover novel non-coding RNAs and CRISPR-like retrotransposon systems.

Few-Shot In-Context Learning (ICL) in Genomic Models and Wet Lab Validation in Mouse Brains: Divya Nori et al. found that merely providing 4 to 64 natural sequence examples as context prompts (no parameter fine-tuning required), Evo 2 could autonomously recombine motifs and design lanmodulin (rare earth-binding proteins), target-specific antibodies, complete CRISPR-Cas9 systems with all components, and even entire phage genomes. When combined with AlphaGenome as a reward model and GRPO reinforcement learning post-training, the resulting cell-type-specific enhancers successfully activated fluorescence in specific brain regions during real mouse brain experiments.

Outstanding Insights

“In clinical medicine, relying solely on historical population averages often provides no meaningful guidance for an individual patient facing life-or-death decisions; when each tumor inherently carries information across 20,000 gene dimensions, every patient is essentially an independent, untapped big data reservoir."

"Previously, only institutions like Broad Institute or Sanger Institute could afford teams of 30–40 top-tier algorithmists, database builders, bioinformaticians, and clinical physicians. Now we've AI-intelligentized the entire laboratory infrastructure, operating in real time around a single patient."

"Even if you deploy an agent to scan all publicly available NIH raw cellular data and expand the dataset to over 500 million cells, you still find most samples are immune cells and T cells—blindly stacking data does not buy true biological diversity."

"The opening sentence of top-tier cell biology textbooks states: 'From many perspectives, our understanding of the universe's structure surpasses our comprehension of living cell dynamics'; we’ve sequenced complete genomes of numerous species, yet we still cannot predict how a cell behaves after knocking out an uncharacterized gene."

"Humans can read natural language, understand code and mathematical formulas, but cannot directly interpret DNA sequences. Genomic large models fundamentally act as tools that translate the cryptic manuscript written over billions of years of natural selection into something interpretable by both humans and AI scientists."

"Before training, we simply replaced random whole-genome sampling with a bias toward genes and conserved elements. With the same task—next-base prediction—the model’s performance on clinical mutation evaluation benchmarks jumped from bottom-tier to industry-leading (Spearman correlation improved by 0.16)."

"No one specifically trained Evo 2 for in-context learning, yet purely through single-nucleotide next-base prediction, the model naturally exhibited 'few-shot design capability'—even with just four CRISPR-Cas sites as examples, it generated the full system directly."

Full Text

Live Turing Test Demo: GPT-6 Astra Deception Rate 36.4%, Humans Also Misclassified as AI in 1/3 Cases

Francois Chaubard: Good evening! Welcome to tonight’s "YC Paper Club" special on AI for Biology. We’ve gathered several outstanding scholars and founders I’ve worked with in the lab and collaborated with at YC—this lineup is truly impressive. But before diving into life sciences, I had a spontaneous idea: please all pull out your phones and scan the QR code on the big screen.

Recently, over dinner with Greg from ArcAGI, he confidently asserted that "the Turing Test has already been solved." I immediately challenged him: Alan Turing stated in his seminal 1950 paper that approximately 50 years later (around the year 2000), we’d be able to write a program where the average interrogator would fail to distinguish between machine and human more than 30% of the time after five minutes of continuous questioning (i.e., at least 30% cases undetected). Yet, in reality, I can usually spot an AI chatbot within a few sentences when talking to airline customer service bots.

So, over the past 12 hours, I used Claude Code to rapidly prototype a blind test website via "Vibe Coding." The rules were simple: half the audience was randomly paired with real human interlocutors, while the other half was paired with a frontier large model disguised as a 32-year-old male, fully aware of the current context—being in the YC room. You had two minutes to probe and vote on your judgment.

Okay, real-time results are in! As you see on the bar chart, the top-performing model was GPT-6 Astra, deceiving 36.4% of test rounds; closely behind 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 intriguing baseline appears at the bottom of the chart: the percentage of real humans misidentified as AI by the live audience reached 33.3%! It seems that if you speak briefly and calmly, one-third of people might be mistaken for robots. Overall, despite my deliberate efforts to saturate the model with scene context and repeatedly refine prompts to favor AI, the overall pass rate remained around 30%, far from definitive passage. A quick shoutout again to Boris and the Claude Code team—thank you for enabling me to build this real-time system in just 12 hours.

Farewell Speech: Why Is Now the Best Time for AI Bio?

Francois Chaubard: Moving on to tonight’s core theme: why should we care about "AI for Biology"? During lab meetings, I sat beside Yasa, who constantly pushed me to explore bio-AI. His strongest pitch was: “When you eventually get diagnosed with cancer, you’ll deeply regret not having invested more time in researching AI-driven cures.” While this may not instantly convert everyone, over time I’ve been thoroughly convinced—and now I’m actively involved in this research myself.

On the slide, I’ve listed several key signals currently driving the explosion of AI Bio:

Unprecedented scale of molecular and genomic data: Evo 2 alone has been trained on 9 trillion base pairs, and next-generation models are advancing at even larger scales; sequencing costs continue to plummet, fueling exponential growth in biological data.

Foundational tools entering clinical drug pipelines: Demis mentioned a few months ago at YC that today, every new drug entering the market undergoes at least one stage of development involving AlphaFold. Initially, I thought this was hyperbole—but after privately consulting over a dozen Genentech R&D experts and top PhDs, all confirmed: "It’s really that widespread."

Accelerating real-world deployment of protein design and human therapeutics: AI protein binders have achieved success rates from 10% to 100% on the BindCraft benchmark; therapies developed by Radical Numerics, led by Eric Nguyen from my own lab background, have entered human patient trials; later, Divya will demonstrate how just 4 to 64 in-context learning examples (ICL) can precisely guide biological sequence design.

High-quality benchmarks and verifiable reinforcement learning (RLVR) closed-loop maturity: From CASP in structural prediction to Virtual Cell Challenge in cellular modeling, evaluation benchmarks are becoming clearer. From capital perspective, massive investments are flowing into this domain—from Lila, Chai to Isomorphic. More importantly, the entire AI community has mastered a methodology for recursive self-improvement (RSI/RLVR) on "verifiable tasks": whether through fast computer simulation verification (like unit tests in codebases) or expensive, time-consuming wet lab physical validation, the core flywheel is established—using foundational models to sample candidate drugs, select top-N candidates for testing efficacy, then recycle real-world data to fine-tune the model and iterate continuously.

Also, a bittersweet farewell: this episode marks my final hosting role and the 10th installment of "YC Paper Club." It has been an immense honor to co-build this hardcore academic community with all of you over the past six months. The first nine episodes have collectively garnered over 1 million views. This Friday, I will officially leave YC to embark on my next entrepreneurial journey. Now, please welcome our first heavyweight guest of the night, Alfredo Gonzalez.

Alfredo Gonzalez: When Tumor Sequencing Costs Less Than a Meal, Reframe Cancer as a "Single-Patient Data Problem"

Alfredo Gonzalez (co-founder and CEO of Omanta): Thank you, Francois, for your incredible dedication to this community over the past half-year! The intellectual sparks from previous episodes were truly electrifying. I’m Alfredo, co-founder of Omanta, and what we do is construct, for each cancer patient, a personalized "dedicated R&D lab" centered entirely on their unique tumor, helping them overcome cancer.

Let me briefly outline my academic background: After undergrad, I first studied early developmental transcription factors at Harvard, then joined Broad Institute to participate in the earliest commercial and academic validation of CRISPR gene screening in cancer cell lines. There, I met Ranad Humeidi, who later became my co-founder and wife, leading the drug repurposing initiative. Subsequently, I pursued my PhD in cancer bioinformatics at UCLA, building large-scale automated analysis pipelines for tumor and normal tissue genomics, contributing to one of the world’s largest rare sarcoma datasets. My doctoral focus was tackling the "N-of-1" challenge—rare, difficult-to-treat tumors with extremely limited sample sizes.

Today, shouting "Use AI to defeat cancer" at YC isn’t empty rhetoric—it’s humanity’s first genuine foothold at this critical inflection point. I’d like to share three points:

First, why is cancer now purely a "data mining problem"? As Francois just demonstrated, sequencing cost curves show that a full genome sequencing today costs less than a meal in the San Francisco Bay Area. More crucially, for most late-stage and recurrent cancer patients, tumor sequencing is now covered under U.S. Medicare and Medicaid insurance plans.

Back ten years ago, we achieved a monumental milestone: The Cancer Genome Atlas (TCGA). But under the technical constraints of that era, completing sequencing for hundreds of patients per discrete tumor type was already the limit. With limited patient samples and data dimensions, doctors and researchers could only search for "population averages" and top-ranked frequent mutations in large cohorts—yet these group statistics often offered little practical guidance for individual patients.

Now, things are different: Even a single patient’s complete genome, copy number variation, RNA expression, alternative splicing, and cellular state profiles naturally constitute a vast personal dataset spanning roughly 20,000 gene dimensions. Academia previously accessed less than 0.1% of surface data; now, with full data available, the real bottleneck is infrastructure capable of mining this 20,000-dimensional treasure trove in real time.

Second, how to use multi-agent systems to build a "personalized lab"? Consider how many people were needed to achieve milestones at Broad Institute or Sanger Institute a decade ago? You needed a full team: algorithm experts, infrastructure engineers, cluster operators, variant detection specialists, geneticists, data visualization engineers, plus a clinically savvy medical doctor to interpret results—typically a 30–40 person elite squad.

Now, software industry talks about "software factories"; in biomedicine, the same agent architecture means an "AI-powered personalized lab factory": Tasks once taking months with 30–40 top bioinformatics experts—data cleaning, clustering, pathway inference—can now be decomposed into specialized AI agent clusters (one focused on clinical trial matching, another on safety/toxicology evidence, another on reading biological background and latest literature). If a new mechanism is published Tuesday morning, the system can pull it in and improve ongoing patient treatment plans by Wednesday afternoon, compressing conversion cycles from years down to real-time.

Two real cases from Omanta where we helped patients achieve remission:

Cross-tumor breakthrough for rare brain tumors: Traditional medicine categorizes tumors rigidly by anatomy. For rare cancers, doctors’ literature is nearly nonexistent. But underlying cellular pathways don’t care about human labels—neuroblastoma and lung adenocarcinoma can be driven by the same pathway. One patient with an extremely rare brain tumor had no viable treatment options. By re-clustering in their multidimensional data space, we discovered the tumor’s molecular profile closely resembled mesenchymal tumors. Following the actual activated surface markers and driver pathways, we identified a novel therapeutic combination completely absent from prior literature but targeting the core mechanism.

Correcting failed multi-line chemotherapy: Another patient underwent three rounds of cytotoxic chemotherapy, all failing, with drastically declining quality of life. We performed clonal evolution analysis on their early primary specimen. The data clearly showed: due to high intratumoral heterogeneity, the three chemotherapy regimens selected by hospitals based on standard "trial-and-error" approaches were fundamentally unsuitable for this patient’s tumor biology. Had this automated research infrastructure been in place from day one, the patient wouldn’t have endured the suffering of three rounds of highly toxic chemotherapy.

Third, my vision for the future: Beyond repurposing approved drugs, we will customize novel protein therapeutics and delivery peptides tailored to each patient in real time. As David Baker’s team, including Bennett, demonstrated in their recent paper, AI-designed proteins can shorten traditional R&D screening cycles by a decade. When combined with Google DeepMind’s AI scientist collaboration system (Co-scientist) launched last year in Mountain View, though models may not always outperform top human professors, their performance is close enough to enable massive savings—millions of hours of lab trial-and-error time across the human race.

Live Q&A: In your slides, you mentioned organoids as a key future data source. Is the organoid approach scalable for building a personalized lab for a single patient?

Alfredo Gonzalez: Excellent question—I participated in several organoid projects during my PhD. Objectively speaking, there have been two waves of organoid startups in the past seven or eight years, scientifically feasible: organoids can expand rare, scarce tumor tissues and be used directly for high-throughput drug perturbation testing, much like metastatic biopsy samples. The real bottleneck isn't science but hospital workflow systems: Most hospitals, upon receiving surgical specimens, immediately fix them in formalin, which kills all live cells and permanently eliminates the possibility of subsequent organoid culture. The future requires bridging the front-end sample preservation infrastructure in hospitals.

Live Q&A: Even if you successfully identify the true pathogenic pathway or drug for a rare tumor in the lab, how do you actually break through the "final mile" to convince clinicians to understand your report and approve prescriptions—or even initiate a patient-specific N-of-1 clinical trial?

Alfredo Gonzalez: Clinicians always retain final decision-making authority and responsibility. But in practice, we’ve observed a very positive reality: Modern oncologists work incredibly hard, with no time to set up computing clusters or run whole-genome differential analyses themselves. Especially when first- and second-line standard treatments fail, and doctors must rely on experience to guess the next line, a tightly reasoned, well-supported, multi-dimensional biological data report with verifiable evidence is highly appealing to most clinicians—they’re eager to use it to advocate for new treatment options. Initiating a brand-new molecule-level N-of-1 trial, however, follows a separate regulatory approval process.

Live Q&A: What’s your take on the recently popular "malignancy reversal" direction? For example, a South Korean institution claims to reverse cancer cells back to normal states using organoid signaling models?

Alfredo Gonzalez: I don’t have deep enough knowledge of their specific experimental details to judge. But when discussing any "reversal of malignancy," two core hard metrics matter: First, how strictly you define "reversal" at the molecular or phenotypic level (different subtypes and even different patients’ progression stages vary significantly); second, long-term, real-world clinical outcome data from tracked patients is essential for support.

Live Q&A: When exploring aggressive strategies in a single-patient AI lab, how do you ensure absolute safety and prevent algorithmic errors from causing fatal harm to patients?

Alfredo Gonzalez: Safety is paramount—indeed, the opposite is true: this infrastructure actually greatly enhances treatment safety. Just as I mentioned earlier, filtering out ineffective and toxic treatment regimens for a patient who previously endured three rounds of useless chemo is itself a major risk reduction. If we prioritize pathway matching from already-approved or phase I clinically validated drug libraries, the safety profiles are already extremely clear. Even if moving toward future "de novo" personalized drug molecules, we can absolutely follow the mature dose escalation safety control paths established in modern clinical trials.

Yusuf Roohani: Bias Traps in the 500M Single-Cell Database and Accelerating Virtual Cells via "Autonomous Post-Training Loop"

Yusuf Roohani (Scientific Lead at Intology AI, former ML lead at Arc Institute): Thank you for the invitation! Today I was scheduled to speak on "Virtual Cells," but I’d like to add a slightly different perspective—connecting my recent deep dive into autonomous post-training in large models and how it profoundly aligns with building truly effective virtual cells.

What exactly constitutes "scientific discovery" and "virtual cells"? Scientific exploration is fundamentally a massive "search problem": navigating an almost infinite hypothesis space to find the most effective information with minimal experiments, uncovering causal mechanisms behind phenotypes or diseases. The first value of models like AlphaFold or ESM lies in dramatically accelerating this search: instead of crystallizing every protein in wet labs, we use models to infer 3D structures directly from sequences, enabling high-cost-efficiency prioritization of experiments.

The core ambition of virtual cell research is to elevate this computational power from mere "molecular sequence level" to the cellular, tissue, and even whole-organism levels—closer to real disease symptoms and physiological laws. The primary reason most clinical trials fail isn’t because molecules don’t bind (we’re already adept at designing high-affinity molecules for known targets), but because "clinical efficacy fails"—we didn’t understand which switch to flip in regulating live cells. Humans struggle with the unpredictability of living systems because cells aren’t static molecules, but complex mixtures of molecules perpetually in violent motion and chaotic interaction.

To model cellular states, our earlier-developed "Universal Cell Embedding" model—published in Nature earlier this year—adopted a very intuitive approach: assembling thousands of single-cell gene expression counts into a "cell × gene" transcript count matrix, randomly masking parts of the expression, then using self-supervised learning to predict reconstruction. In fact, over a century ago, developmental biologists proposed the famous "Waddington Epigenetic Landscape": due to transcriptional and epigenetic constraints, living cells cannot arbitrarily exist in any chaotic state but are confined to a "low-dimensional manifold" within a feature space similar to that of large models.

When we fed 35 million single cells into the unsupervised autoencoder matrix, even without any manual labels, the model’s output UMAP manifold spontaneously and accurately clustered into cell types and subpopulations perfectly aligned with century-old biological literature!

But then comes the next question: Natural language processing achieved its Scaling Law miracle by scraping the internet’s Common Crawl. Can we replicate this in cell biology by sweeping through all publicly available cell repositories?

In an earlier study published in Cell this year, we made this bold attempt (long before Claude Code became widespread—we were among the first in the industry to deploy autonomous agents to mine data): We let agents run continuously for a full month, automatically cleaning and reprocessing all single-cell records in the NIH public raw transcriptome database, pushing the data volume up by one order of magnitude—from 35 million to over 500 million cells.

But when we excitedly examined the heat map spanning all tissues and cell types, a sobering result doused our enthusiasm: the overwhelming majority of the 500+ million cells were heavily concentrated in "immune cells," and within those, over half (an entire order of magnitude higher) were T cells! In short, even if you scour every existing open-source cell repository, linearly stacking public data won’t buy true biological breadth and diversity.

Thus, the real solution isn’t blind linear data accumulation, but switching to an "AI-guided closed-loop experimentation" framework: using advanced models like Claude or GPT—already trained on vast biological text corpora—to precisely decide which rare, scarce cell types to invest in for wet-lab collection next.

Once we view R&D as a cyclic structure—"hypothesis → experiment → feedback update"—I was astonished to find that it faces remarkably consistent engineering challenges with the cutting-edge AI labs currently studying autonomous large model post-training (Autonomous LLM Post-Training) and recursive self-improvement (RSI): whether running wet experiments or training models autonomously, the core bottlenecks remain the same: extremely high per-evaluation cost, noisy feedback from experiments/evaluations, and an overwhelmingly vast hypothesis search space.

To validate whether this closed-loop insight can cross-fertilize, six months ago I joined Intology AI, a startup focused on general-purpose automated AI R&D systems. In the past few months, our proprietary system Locus achieved leading results on the authoritative PostTrainBench leaderboard evaluating whether large models can autonomously perform post-training on base models like Qwen—surpassing many common programming agents. The core reason isn’t mystical: general-purpose coding agents (Coding Agents) aren’t optimized for "scientific discovery retrieval and hypothesis falsification"!

Just yesterday, in the latest benchmark, we directly migrated this Locus system—fine-tuned in pure machine learning post-training scenarios—almost unchanged to the "AI molecular design and property prediction" task (letting agents perform efficient sampling retrieval in latent space and control noise validation set overfitting).

The results showed that under strict lock-in to the same GPU compute sandbox budget, the specialized research-oriented Locus system significantly outpaced Anthropic and OpenAI’s general-purpose coding agents, especially in specific sampling efficiency tests, even exceeding BoltzGen and Codex—though these remain computer simulation results, the next step must return to wet-lab validation.

Finally, I’d like to conclude with the opening sentence of a classic cell biology textbook: "From many aspects, human understanding of cosmic structure exceeds our comprehension of the internal workings of living cells; although we’ve sequenced complete genomes of numerous species, we still cannot predict where a cell will go after knocking out a single uncharacterized gene."

Live Q&A: What specific techniques did you use during post-training (e.g., reinforcement learning RL)? How do you evaluate the gap between simulation (virtual cell) and the real physical world?

Yusuf Roohani: Let me answer the latter first: every time we measure the same cell under identical lab conditions, biological variability and instrument noise are enormous. Thus, "precisely replicating the physical world within noisy evaluators" is inherently the hardest challenge—any simulation evaluation must ultimately face the scrutiny of wet-lab fidelity. Regarding the former, you’ve hit the crux: traditional programming agents weren’t designed for scientific trial-and-error. They lack dedicated judge agents responsible for long-term hypothesis tracking, controlling single-experiment compute or budget costs, and strategically ranking candidate hypotheses. Precisely by restructuring this multi-agent collaborative mechanism around the R&D pipeline, we enabled the seamless transfer of automated machine learning results into the scientific discovery domain.

Live Q&A: On your final chart, the x-axis reads "H100 compute time"—how did you ensure fair comparison with Anthropic and OpenAI? And did you measure cost by token consumption or training steps?

Yusuf Roohani: Our comparison was fully controlled: placing our Locus system and pre-existing autonomous agent systems from Anthropic and OpenAI into identical sandboxes with identical H100 GPU quotas. Since this task involves calling large models to drive scientific model training and fine-tuning, the true bottleneck and billable cost lie in GPU training compute hours—not pure inference token usage.

Live Q&A: Could you offer a slight reveal of Locus’s underlying architecture? For example, how was it built?

Yusuf Roohani: Specific proprietary architecture details are confidential, but the core is a large language model as the engine, with a series of research-specific engineering strategies refined externally around "hypothesis generation, preventing validation set noise overfitting, and long-term memory retrieval." By the way, writing code still uses Python (laughs—almost revealed the secret).

Francois Chaubard (follow-up): In real-world scenarios like drug trials, where sampling cost is extremely high, sample efficiency is critical. Recently, many remarkable advances in large models stem from "continual harness" frameworks and context memory updates that don’t alter model weights. Do you favor heavy parameter fine-tuning, or framework and context engineering?

Yusuf Roohani: I strongly agree with your assessment. Especially in early-stage biological R&D and exploration, optimizing external harnesses (tool scaffolds and context update mechanisms) is undoubtedly the highest cost-performance gain, fastest fruit. Parameter fine-tuning and post-training are better suited when you genuinely need to embed high-frequency capabilities for fast inference retrieval. But to avoid overfitting on single benchmarks (bench maxing), we deliberately restrain weight changes in early exploration phases, prioritizing deepening harness and inference framework design.

Garyk Brixi: Reading Life’s Code Across 9 Trillion Base Pairs, from Evo 2 to Minerva Discovering Novel RNAs

Garyk Brixi (Stanford PhD candidate, Evo 2 team lead): Over the past two years, we’ve all been using language models trained on "next-token prediction" daily—humans can read English, code, math formulas, but stare at DNA base sequences and remain utterly baffled. Yet in a sense, DNA is a universal language written over billions of years of natural selection: DNA encodes RNA and proteins, which through immensely complex mechanisms determine whether organisms stay healthy or grow tall; and natural selection, in turn, filters which DNA survives and gets sequenced.

Over the past three years, the genomic data volume in NCBI’s public databases has doubled. But DNA modeling remains extremely challenging, as its vocabulary consists of only four bases, filled with parasitic sequences and "redundant junk" that replicate wildly within the genome, and spans staggering scales: a single gene may stretch tens of thousands of base pairs, yet often a single nucleotide mutation can be lethal.

So we decided on the simplest, most direct, and purest approach: building a generative pre-trained large model for DNA across all life forms—Evo 2 (7 billion and 400 billion parameter dense models, based on hybrid attention and convolution architecture). A key but often overlooked trick in training: we ruthlessly discarded the vast majority of garbage data collected from public sources!

We not only strictly retained one high-quality representative genome per species, removed highly similar redundancies, condensed the total scale to 9 trillion base pairs; during the main pre-training phase, we also deliberately performed data re-sampling amplification around functional regions—genes, RNA, and evolutionarily conserved elements—effectively discarding trillions of invalid noise tokens, first mastering basic semantics in short contexts, then extending to a super-long 1 million base pair window to learn cross-spanning combinatorial patterns.

After lowering the loss function, how do you converse with a model that only outputs A/T/C/G? There are mainly two paths:

As an extremely sensitive "biological causality discriminator (Decision model)": The model has read all naturally surviving DNA across life domains, thus deeply understands what sequences "conform to natural laws." Input a sequence with a mutation, if the model assigns a very low probability (it perceives the code as anomalous), it often strongly indicates functional loss or pathogenic risk. In the clinical evaluation test for high-risk cancer genes like BRCA1, Evo 2, having seen only one human reference genome, matched or even surpassed human-specific supervised models in predicting mutations in both coding and non-coding regions! Ablation experiments proved that retaining 50% random whole-genome sequences causes Spearman correlation to drop sharply by 0.16 and fall to the bottom of the leaderboard—highlighting how devastating data mixing ratios can be.

Using sparse autoencoders for interpretability dissection to unlock novel gene generation: After opening Evo 2’s internal neuron features, we discovered that simply by "predicting the next base," the model had autonomously developed feature neurons recognizing bacteriophage prophages, bacterial defense CRISPR arrays, and single-gene phages co-opted by bacteria—all structurally distinct in base sequences, yet the model unified their biological functional connections through unsupervised self-learning. After micro-adjustment with functional classifiers, the team generated a novel PhiX-like phage from scratch—now entering human clinical trials.

Regarding Evo 2’s past limitations in predicting "co-evolution between distant positions and 3D contact interactions," we recently released the new bidirectional model Minerva. It can directly scan and sketch out hundreds of previously missed novel non-coding RNA structures in genomes like Bacteroides, using the model’s long-range attention weights. In a recent wet-lab validation, we conclusively discovered a new system named UG27 reverse transcriptase (UG27 RT): it features a programmable non-coding RNA array resembling CRISPR, capable of converting DNA into RNA and then reverse-transcribing it into DNA with hairpin structures. Though its full natural functions remain unclear, this proves that pure DNA language models now possess the hard capability to serve as engines for human biological discovery.

Live Q&A: Why did you ultimately choose nucleotide-level tokenization as the finest granularity, rather than trying multi-base vocabularies like BPE?

Garyk Brixi: Based on current small-scale comparative experiments, nucleotide-level tokenization yields significantly better performance for precise single-point mutation prediction. Of course, in the long term, I’ve always been eager to explore byte-level end-to-end modeling to balance resolution and computational length.

Live Q&A: Have you considered a "two-tower" architecture like embodied physics AI—where one tower reads DNA sequences, and the other directly translates biological insights into human-readable language for biologists?

Garyk Brixi: Indeed, many teams are racing toward competing routes: some frontier large model companies train directly on mixed DNA sequences in text pre-training, while others embed DNA models into vector alignment mappings into large language models. Which architecture will ultimately prevail remains undecided.

Live Q&A: Since data filtering and ratio adjustments dramatically boost performance, how did you scientifically identify the optimal "pre-training data mixture"?

Garyk Brixi: Honestly, training Evo 2 at the end of 2024 was pioneering—first-mover status with virtually no prior preprints or parameters to copy. Our core approach was simultaneously constructing a tightly integrated set of biologically meaningful evaluation sets (Evals) and scaling law analysis while processing data. Now, with evaluations maturing, we can more targetedly optimize data ratios based on evaluation shortcomings.

Live Q&A: Since you’ve already discovered so many biologically meaningful neuron features inside the model, have you tried using these internal features to steer Evo 2 for directed interventions—for example, designing DNA sequences specifically targeting cancer cells?

Garyk Brixi: Last week, we published several new preprints specifically exploring "how to steer Evo 2 for biological design." To be honest, at this stage, pure mechanical neuron intervention for directed biological design doesn’t deliver the same wow factor as the next speaker, Divya, is about to showcase with her "few-shot prompting" approach.

Live Q&A: You mentioned earlier that each species retains only one genome during training—doesn’t that discard vast individual genetic differences within the human population? Yet you’re still able to predict human cancer pathogenic mutations so accurately—what’s the underlying logic?

Garyk Brixi: The reason aligns perfectly with classical evolutionary comparative statistics: even without deep learning, simply aligning the genomes of hundreds of primates, any base position that has remained almost unchanged across vast evolutionary scales is typically a core vital site—any mutation here likely leads to serious consequences.

In contrast, if we input every attendee’s personal genome into training, the massive token compute would be wasted on highly redundant sequences shared among humans, failing to learn core pathogenic features. Under constrained compute budgets, prioritizing diverse, conserved signals across the entire tree of life offers the highest cost-effectiveness. Of course, to capture human-specific complex diseases and regulation, the next step must integrate population multi-omics data and single-cell expression data mentioned by the previous speaker.

Divya Nori: No Fine-Tuning Required! 4 Examples Directly Prompt DNA Model, Mouse Brain Experiment Lights Up Instantly

Divya Nori (Stanford computer science PhD candidate, researcher at Brian Hie Lab): Hello everyone! The mystery left by Garyk is now mine to unveil: how to efficiently design novel biological components without modifying any weights, without running thousands of data fine-tuning iterations, simply by crafting a few prompts for the genomic large model? This work was primarily led by Dr. Alex Hao, with contributions from Garyk and Prof. Brian Hie and other colleagues at the lab.

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

But in the field of biological sequence design, traditional thinking has long been stuck in another highly data-dependent paradigm: for example, early papers generating active CRISPR-Cas9 systems had to first fine-tune Evo 1 on tens of thousands of CRISPR sites; to generate complete mini-phage genomes, one had to heavily fine-tune Evo 2 on thousands of microvirus family genomes. Yet in real biology, countless new targets and rare protein families simply don’t have thousands of labeled training samples.

Surprisingly, we recently discovered a magical emergent property in Evo 2—previously almost exclusive to human natural language models: after massive-scale single-nucleotide next-character prediction pre-training, with no explicit instruction on in-context learning, Evo 2 naturally exhibits strong few-shot in-context learning capability! This means that by simply showing it 4 to a dozen functional DNA sequence examples in the input prompt, the model can independently infer recombination and mutation algorithms during reasoning, directly completing a new functional sequence.

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

Rare natural proteins (Lanmodulins): These are revolutionary high-affinity rare earth-binding proteins for rare earth mining recovery. One specific subfamily has only 14 known natural samples. We directly fed these 14 sequences into the context prompt and successfully prompted Evo 2 to generate a novel, functional target protein candidate—zero fine-tuning.

Antibody targeting design: We directly used a few hit sequences generated by a 3D structure generation model as text examples fed into Evo 2. Without inputting any 3D protein structural conditions, we successfully generated high-activity new hits for all tested antibody targets.

Reconstruction of complex gene editing systems (CRISPR-Cas9): Previously required full fine-tuning on tens of thousands of data points. Now, simply inserting four CRISPR-Cas site examples into the context, Evo 2 can seamlessly complete the entire system—CRISPR array, tracrRNA, and Cas9 protein—in one shot.

Complete phage genome design: Achieved with only a fraction of the compute cost of full fine-tuning, directly prompting the model to generate a novel phage with the expected complete genomic architecture.

But we didn’t stop at software simulations—we advanced this paradigm into animal wet-lab experiments, focusing on designing cell-type-specific enhancers (critical switches determining whether gene therapy drugs express only in specific brain regions or lesions without damaging healthy organs).

While maximizing the enhancer prompting capability, we also addressed the earlier concern about "how to post-train Evo 2": we used a sequence-to-function prediction model (similar to AlphaGenome, capable of directly predicting where a regulatory element activates in which cell type) as a reward function, and introduced constraints to prevent gaming and control DNA sequence entropy and GC content, applying GRPO reinforcement learning to iteratively fine-tune Evo 2, specifically amplifying the model’s ability to leverage context exemplars!

Then, Alex delivered this reinforcement-learning-enhanced model’s few-shot-designed novel enhancer into real mouse brains in the wet lab. The right-side fluorescent microscopy images in the slides clearly prove: the target brain region successfully emitted clear, highly specific fluorescent spots!

Looking ahead, this "data-efficient conditioning" will bring biological design into a true "lab-in-the-loop" era: scientists hold a small pool of natural sequence exemplars, let Evo 2 use ICL to batch-generate multiple new designs, quickly filter the best via in silico filters, send the top design to wet-lab testing, then directly insert the real-world validated hit sequences back into the next round of prompt exemplar pool—creating a snowball effect.

Moreover, the reasoning enhancement and scaffold techniques accumulated in human language models can be seamlessly transferred into the biological design world. All content shared today is included in the two preprints we published last week.

Live Q&A: Could you elaborate specifically on your reinforcement learning post-training mechanism, reward function, and anti-gaming measures for enhancer design?

Divya Nori: The overall environment design is quite standard and efficient: we first provide known active natural enhancer examples in the target cell type to prompt Evo 2, letting it batch-sample new candidate sequences; then feed each new sequence into a "sequence-to-function" oracle model (like AlphaGenome) to predict specificity activation scores across various cell types, serving as the reward signal for GRPO algorithm updates to Evo 2’s weights. To prevent the model from taking shortcuts ("reward hacking") during optimization, we added regularizer terms in the reward to ensure normal DNA sequence entropy and reasonable GC content.

Live Q&A: Why did Evo 2 exhibit in-context learning capability? I can understand adjacent structures like gene clusters, but why does antibody sequence design also show this prompt-as-use generalization effect?

Divya Nori: Frankly, we were initially surprised ourselves, as the model received no dedicated ICL training. Academic explanations for ICL emergence in natural language models have several mainstream theories, which surprisingly fit perfectly in genomic data: first, natural sequences also follow typical long-tail power-law distributions and burstiness—certain rare k-mers appear clustered only in specific local contexts; second, the genome naturally contains abundant semantic correlations and parallel symmetric structures.

And from our internal mechanistic analysis in the preprint, Evo 2’s attention layers actually developed "induction heads" nearly identical to those in natural language models, enabling pattern recognition and contextual analogy.

Live Q&A: In the "context prompting + reinforcement learning design enhancer" task, how much credit does the pre-trained weight’s inherent classic expression patterns deserve, and how much is truly due to the model’s real-time combinatorial creativity from the new patterns in the prompt?

Divya Nori: This is precisely the fascinating balance between "weight-internalized priors" and "context-learned combinations." Specifically for enhancers, they are extremely sparse in Evo 2’s vast pre-training corpus across life domains. Therefore, I believe while the model certainly leverages underlying sequence physical priors, a larger portion of the breakthrough in specific expression regulation stems from Evo 2’s real-time interpretation of core motifs in these new exemplars provided in the prompt, directly recombining them with precision.

Francois Chaubard: Absolutely brilliant! Thank you again, Divya, and thank you all guests and friends here tonight!

Source: DeepTrend TechFlow