Transcript

Machine Learning in Drug Discovery: Oren Kraus

From Broad Institute · Episode page · Watch on YouTube

· 19 min · 3,210 words · 24 timestamped sections

This is the full, timestamped transcript of "Machine Learning in Drug Discovery: Oren Kraus", an episode of Broad Institute published January 5, 2026 (19 min, about 3,210 words). It mentions 15 people, works and ideas, including Recursion, Dr. Orurin Kas, SARS-CoV-2, Open Phenom, RXRX1.

Who speaks in this episode?

What is discussed, and when

  1. – 3:00: Recursion, Dr. Orurin Kas, Google LLC, Roen Tech
  2. – 6:00: SARS-CoV-2, Recursion
  3. – 12:00: Recursion, RXRX1, CPPR
  4. – 15:00: Recursion, ICML
  5. – 18:00: Hugging Face, Open Phenom, RXRX3

Notable quotes from "Machine Learning in Drug Discovery: Oren Kraus"

  • Next speaker is Dr. Orurin Kas. He is the director of machine learning at recursion Toronto. His group focuses on advancing machine learning for tech enabled drug discovery and…
    about Recursion · at
  • In oncology and one in rare disease and partnerships with leading tech companies like Nvidia and Google as well as leading biioarma companies like Synopia and Roen Tech.
    about Google LLC · at
  • Those embeddings of images, we can start to compare different conditions. So in this example, we're looking at disease model of SARS co2 and we have a we can generate a vector…
    about SARS-CoV-2 · at
  • So we spent about four weeks screening all the potential treatments in a variety of cell lines. two well-known treatments that we screened were remos and hydroxychloricquin.
    about hydroxychloroquine · at
  • This was done with a data set that we released publicly called RXRX1. this was part of a part of a nurse challenge in 2019 for the computer vision community. so that data set has about.
    about RXRX1 · at
  • And obviously that took a lot more compute. and then this work was presented at CPPR last year.
    about CPPR · at
  • So this work was presented at ICML this year. what we showed pretty quickly was that this approach with this approach we continue to see improved scale improved performance with.
    about ICML · at
  • We've actually made these models and some of the data sets publicly available. so open phenom is a smaller version of the phenom models I was mentioning. it's actually trained…
    about Open Phenom · at

Who and what is mentioned in this episode?

People

  • Dr. Orurin Kas — Person · Main speaker · 1 mention · first at
    Introduced as the next speaker and delivers the talk (remotely) on foundation models for phenomics and early-stage drug discovery at Recursion.
    Our next speaker is Dr. Orurin Kas. He is the director of machine learning at recursion Toronto. His group focuses on advancing machine learning for tech enabled drug discovery and today he.

Companies & orgs

  • Recursion — Company · Business focus · 7 mentions · first at
    Primary company discussed, focusing on machine learning for drug discovery
  • Hugging Face — Company · Platform · 2 mentions · first at
    Platform where Recursion released their models
    Up to 12 channels, you can still compute embeddings from it. and then we released it completely open source on hugging face and it's been downloaded over 115,000 times since…
  • Google LLC — Company · Partner · 1 mention · first at
    Tech company partnering with Recursion
  • Synopia — Company · Partner · 1 mention · first at
    Biopharma company partnering with Recursion
    Partnerships with leading tech companies like Nvidia and Google as well as leading biioarma companies like Synopia and Roen Tech.
  • Roen Tech — Company · Partner · 1 mention · first at
    Biopharma company partnering with Recursion
    Like Nvidia and Google as well as leading biioarma companies like Synopia and Roen Tech.
  • Broad Institute — Company · Referenced organization · 1 mention · first at
    Mentioned as a place where the speaker is a fan of symposiums
    First I want to say I'm a huge fan of the symposiums at the Broad Institute.

Products & tech

  • Open Phenom — Product · Secondary owned · 1 mention · first at
    Open-source version of the phenom models for scientific community
  • RXRX1 — Product · Secondary owned · 1 mention · first at
    Public dataset released by Recursion for computer vision community
  • RXRX3 — Product · Secondary owned · 1 mention · first at
    Genome knockout screen dataset released by Recursion
    Hugging face and it's been downloaded over 115,000 times since about last year. RXR 3, that's a genome knockout screen that also includes about 1,600 compounds. so it's much…
  • hydroxychloroquine — Product · Incidental · 2 mentions
    Mentioned as one of the compounds screened
  • remdesivir — Product · Incidental · 2 mentions
    Mentioned as one of the compounds screened
    So we spent about four weeks screening all the potential treatments in a variety of cell lines. two well-known treatments that we screened were remos and hydroxychloricquin.

Concepts & topics

  • SARS-CoV-2 — Concept · Case study · 2 mentions · first at
    Example of disease model used in phenotypic screening

Places & events

  • CPPR — Event · Referenced event · 1 mention · first at
    Event where Recursion presented their work
  • ICML — Event · Referenced event · 1 mention · first at
    Event where Recursion presented their work

Full transcript with timestamps

Our next speaker is um Dr. Orurin Kas. He is the director of machine learning at recursion uh Toronto. His group focuses on advancing machine learning for um tech enabled drug discovery and today he will be talking about you know foundation models in early stage drug discovery. So with that uh we'll let uh you know so Orin he cannot uh come in person so he will be uh sharing his talk on Zoom.

Everyone um really sorry I could be in person at MLB today. I was at the airport most of the afternoon yesterday but uh because the flight disruptions in Boston Logan I wasn't able to come in. Um first I want to say I'm a huge fan of the symposiums at the Broad Institute. uh back in 2016. I'm at my co-founder of my previous startup at a cyto that ancar lab organized. Uh so it's really a great opportunity for for everyone there to take advantage of it and and network. Um so my talk today will be about foundation models and early drug discovery. Um so I'll introduce platform first uh and what we call maps of biology. uh and I'll specifically talk about how we build foundation models to analyze the high content data that we generate at recursion. Uh and finally I'll plug that into how we can use those foundation models for virtual screening.

Um so recursion started about 12 years ago. Uh the initial idea was focused around using deep learning uh that was really taking off for computer vision at the time for microscopy data. um and the idea that we could really extract uh very high level insights from from phenotypic based images. Uh we've expanded a lot since then. So much more than phenomics going on at recursion. Uh we have high throughput transparic assays. Uh also a lot of work going on in computational chemistry and just general MLAI research. uh and currently have four clinical stage programs, three in oncology and one in rare disease and partnerships with leading tech companies like Nvidia and Google uh as well as leading biioarma companies like Synopia and Roen Tech.

So the core concept that recursion has is to change the shape of the drug discovery funnel. So we actually want to expand the scope of therapeutic areas that we're exploring and potential targets that we're exploring in those therapeutic areas with unbiased genomewide screen up front. Um and then we want to use predictive models to quickly assess which of those are are not really feasible so we can fail them fast. Uh and finally we're using cutting edge computational chemistry tools to accelerate how quickly we can develop compounds against those potential targets.

Uh this is powered by our robotic wet lab which can conduct 2.2 million wet lab experiments every week across approximately 50 different cell lines and we use whole genome crisper for most of these uh maps that we built. Uh we screened over two million compounds totaling over 65 pabytes of proprietary data mostly imaging based. Um so how can we use phenotypic models for drug discovery? So I'll start with a a straightforward approach. So in this case we have a clear disease model uh and healthy controls. So when we run uh imaging experiments with these we can look for compounds that revert the the disease model back into the healthy phenotype. Um and to do that we need really good representations of what's going on in the images. ideally ones that can um that can deal with all the batch effects that that show up between different experiments. Uh and that's where I'll spend the bulk of the talk today is how

we generate those. Um but once we have those embeddings of images, we can start to compare different conditions. So in this example, we're looking at disease model of SARS co2 and uh we have a we can generate a a vector in high dimensional space that indicates the direction of the phenotype from the control condition to the disease condition and then as we start swinging compounds we can try to see which ones revert the phenotype in the in the direction of the disease. uh so we can project those vectors onto the uh healthy disease vector and um ideally find compounds that have a a strong on target effect.

Uh so for CO specifically uh the team at recursion uh was a mega effort back when CO was was just starting uh I think tribute to the scientific community. Uh so we spent about four weeks screening all the potential treatments in a variety of cell lines. uh two well-known treatments that we screened were remos and hydroxychloricquin. Um so in these examples you can see how different compounds have on disease and off disease scores. So remosphere and this GS compound have positive uh responses with increasing doses but the drugs like rimosphere seem to have very high offtarget effects. uh and when we expand this to all the compounds that have been considered in clinical trials um you know it took a few years for out of all the readouts for these but the platform essentially predicted nine out of the 10 clinical trials correctly uh in terms of e efficacy so it really shows the p the power of doing

phenotypic screening for for drug discovery um so moving on I'll I'll explain another interesting application of phenotypic screens especially when we start adding genomewide crisper knockout strands. So in this case, we can start to recover known functional relationships in in genomic networks. Um in this example, we're looking at the mtor pathway. Um and if we look at the same genomic vectors in one of these knockout screens, we can see that um genes that are part of the same complex will typically have high similarity. So in this case TSC1 and TSC2 seem to have like similar phenomic embeddings um and things that are inhibitory of that complex uh have have an opposite uh effect. So we call these phenal opposites typically uh so this lets us recover recover known interactions but we can also start to uncover novel interactions. So in this case, if we're screening compounds in the same space,

we can also find compounds that have a similar effect to a G knockout or an opposite effect to a G knockout. Um, one of the applications of this is we can start to go after targets that are hard to uh develop compounds against in traditional settings. So one example is Mick, which is disordered. So it's hard to do like a typical um binding assay against it. uh but when we look at in the phenotypic space we can identify a handful of compounds that seem to have uh dose response relationships with meg this doesn't necessarily mean that these compounds are directly interacting with me but they seem to have a similar phenotypic effect um and then when we do this across the whole prote or less uh we can start seeing proteins with similar annotations clustering together and similarly for compounds with no annotations they cluster together in these phenotypic spaces.

And if we extend to genomide screens, that's really where the magic happens where we can start to identify completely novel interactions between um genes and genes or genes and compounds that weren't previously known. Um so that enables us to really open up the hypothesis space for new targets in in therapeutic areas. uh to benchmark these what we call maps uh we developed a a framework along with collaborators at gelantech. So this can be applied to self painting images or even RNA seek data. Um but we start with the embeddings I was mentioning. First we filter out any QC issues from from those samples. Uh then we use batch alignment techniques to align all the experiments so we can get good representations. Uh and finally we aggregate those to the perturbation level and that's when we can start to relate pertabbations to each other. Uh for benchmarking there's two main categories of benchmarks we have. One is

to identify how many perturbations have a signal that's differentiated from the background effect. Uh and the other one is looking at interactions between pertations. So here we rank all the similarities and difference opposites um between all the perturbations and we're looking to see in the extremities of that distribution how many known interactions are recovered based on databases like string DB and Cororm. Um so with that background I'll explain how we train these embedding models to extract information from the images. um our initial approaches which were the first for deep learning after we moved away from cell profiler uh was using what we call weekly supervised learning.

So this approach would take image crops from the microscopy wells uh and then we tried to predict what the perturbation was that uh was applied to those wells. Uh this was done with a data set that we released publicly called RXRX1. Uh this was part of a part of a nurse challenge in 2019 uh for the computer vision community. Uh so that data set has about a thousand SRNA treatments across 50 batches. So it's a good way to get the model to learn um how different perturbations look versus the batch effects which are kind of labeled the same way over and over again. Uh once this model is trained, we can use the second to last layer typically as the the feature representation. Um and that's what was used to power the maps in our platform for a while.

One issue we faced with this as our data set grew is that when we wanted to train larger models with with more diverse data, we found that performance typically degraded and that's what uh motivated us to start looking at um self-supervised learning approaches. So these were becoming very popular 2002 sorry 2022. Um so in this setting instead of using the the labels of what was uh what was applied to the the wells and the microscopy data we're just using the raw microscopy images as their training data. Um and the task here is to reconstruct masked patches in the original image. Um so typically this is done with a high mask ratio of around 75%.

uh and then our transformer-based encoder only really sees the visible patches and it's tasked with reconstructing all the pixel values from the mass patches. Um the nice thing is thing about this is it can scale really well. So we can train much larger models of martial art data sets and uh that tended to increase performance which I'll show in the next few slides and the embedding we use here is like the output of the u transformer encoder for the decoder layers. Uh so just for some stats on this approach um we were able to switch from a densenet backbone to a vision transformer background bone that allowed us to increase the model size from 30 million parameters to 300 million. So a 10x improvement um and then you know the biggest challenge was the data set size.

So instead of using just those um 100,000 wells that we had for RXRX1, we expanded to nearly like all the genomic wells that recursion had that we were able to train this on. Uh so that was a thousandx increase in the data set size. Um and obviously that took a lot more compute. Uh and then this work was presented at CPPR last year. So overall the resulting model was a VIT large with a patch size of eight 330 million parameters uh trained on those 93 93 million wells for 40 epox and it really showed that um you know the same recipe that's being applied in the general NLP and computer vision communities also applies to these biological data sets where if we scale compute and scale the model size and the data set size we're we're seeing improved performance um specifically on that recall benchmark I was mentioning.

So what we did next, so this this happened last summer. Recursion upgraded our cluster and for a while it was number I think 35 or 36 of the top supercomputers in the world. Um and this was this work was a bit uh driven by one hawk was in the audience. Uh but we we wanted to push this even further. So we expanded the model size to a VIT gigantic which has 1.8 billion parameters. So it's a 6x improvement over the phenom one model. Um and then it was trained for four times longer on the the new cluster that we installed.

Um and interestingly it actually was trained on a smaller data set. So we took a subset of the 93 million data set and I'll explain why in the next few slides. So this work was presented at ICML this year. Um what we showed pretty pretty quickly was that this approach uh with this approach we continue to see improved scale improved performance with scale. Uh so here you're looking at a log plot of the model parameters um against the relationship recall in string DB. Uh so that 1.8 billion parameter model uh is once again improving over feon one um almost linear in log space. Uh and additionally on the the replicate consistency metric uh it also had a pretty significant improvement. So um about 25% more um knockouts were found to be consistent with this model.

Um in addition to just scaling we actually had to make quite a few changes to this approach to get it to work better. Uh one is what I mentioned about data creation. So the original 93 million data set had mostly like everything we could scrape from creation. Um but what we did here uh was try to reduce redundancy in the data set so it was more diverse and could provide stronger signal over the training batches. So we reduced we removed a lot of uh negative controls which tend to look similar to each other and uh prioritized wells that were that had consistent signal using other models like the weekly supervised model or the feed on one model. Um and the other key insight we had is the final layer of the encoder doesn't necessarily have the best uh feature space for the downstream tasks. So we came up with some methods to find which layer from the encoder uh actually had the the best performance

downstream. Um so the plot here on the right shows how each one of those contributions uh contributed to the performance across both metrics. Um, diving deeper into that uh blockwise search I was I was mentioning. Um, it's really impractical to to run the the uh recall metrics across uh every single one of the 48 layers for the largest model across all the checkpoints

um because we need to inference 80 million images each time because we we have to compute the embeddings for the entire genome walkout. Uh so what we came up with was two proxy tasks that are based on um on linear classifiers at intermediate layers. So one of these is a set of 40 genetic functional groups. Um so just a 40-way classification and the other one is the the RXRX1 data set I mentioned previously. So these are much smaller um data sets to work with and we can see that the performance kind of degrades toward the later layers um especially for the larger models. So that let us select like the ideal layer for downstream tasks. Uh and then we showed that the proxy tasks actually well well correlated with the more compute intensive uh recall and replicate consistency tasks. So these have a rank correlation of 97 and 91% for the string DB recall and the replicate consistency recall respectively.

Um so overall uh the best model was a what's called an ME gigantic model with a batch size of eight and the the train model um which uh only computes up to the 38th layer instead of the 48th layer had the best performance on um on the replica consistency metric and on the recall metro and you can see the effect of individual changes. So the data separation has a big impact on the replicate consistency. Uh the layer selection does as well and uh same with the scaling. Um and then finally we we also validated this model on external data sets. So we computed the same recall metric on the jumpcp data set. Uh this is a different set of G knockout.

So it's not directly comparable to the RXRX3 experiments. Um but there we also saw that the um with improve with the larger model sizes like the the recall generally improves even on these external data sets that actually have a different set of channels on them. So that was quite impressive. Um to support the scientific community we've actually made these models and some of the data sets publicly available. Uh so open phenom is a smaller version of the phenom models I was mentioning. Um it's actually trained with a channel agnostic architecture. So whether you have one channel or I think up to 12 channels, you can still compute embeddings from it. Um and then we released it completely open source on hugging face and it's been downloaded over um 115,000 times since about last year. Um RXR 3, that's a genome knockout screen that also includes about 1,600 compounds. Uh so it's much larger. It's

about 83 terabytes, but it's also been downloaded 6,000 times. Um and then uh RXRX3 core is a subset of uh that data set that's on hugging face. Um and it's more designed for benchmarking and it has an associated benchmarking task. Uh so that one was published at a workshop at IAR last year. Um I only have one minute left so I might have to skip over the the other application about virtual drug screening. Uh but this was work that was done at Veilance where we combined um graph encoder models specifically one called GPS with these feed on one embedded to be able to train contrastive models uh that can uh virtually screen compounds against uh phenotypic uh phenotypic profiles.

Um so the takeaway there is that the foundation model for phenomics uh is a pretty key component in getting contrasted models to work with uh with graph neural network based approaches on molecules. Um so yeah with that I'll I'll wrap up. Um the key takeaways are that self-supervised learning combined with data creation and scaling compute and uh data set sizes uh is a really good recipe for training foundation models for phenomic images but also generally for for scientific data. Um and that as you start to combine foundation models it really unlocks new capabilities um for multimodal learning. And with that I'll take any questions and thanks for listening.

Questions about this episode

Who speaks in "Machine Learning in Drug Discovery: Oren Kraus"?

Dr. Orurin Kas (Main speaker).

What is discussed in "Machine Learning in Drug Discovery: Oren Kraus"?

In order: 0:00 Recursion, Dr. Orurin Kas, Google LLC, Roen Tech; 3:00 SARS-CoV-2, Recursion; 9:00 Recursion, RXRX1, CPPR; 12:00 Recursion, ICML; 15:00 Hugging Face, Open Phenom, RXRX3.

What is mentioned in this Broad Institute episode?

15 people, works and ideas are mentioned, including Recursion, Dr. Orurin Kas, SARS-CoV-2, Open Phenom, RXRX1, RXRX3. Each one on this page links to its own page with every episode that mentions it.

How long is "Machine Learning in Drug Discovery: Oren Kraus"?

The episode runs 19 min; the transcript is about 3,210 words in 24 timestamped sections.

More from Broad Institute

About this transcript: the episode's spoken words as captioned on YouTube, split into timestamped sections. Reclip.Me locates every person, work and idea in the text, keeps the time it is first named, and quotes the words around it verbatim. Last updated January 24, 2026.