Kanvas Bio's State of the Art AI Models & Reflections from JPM 2024


Kanvas Bio's State of the Art AI Models & Reflections from JPM 2024

by Kanvas Biosciences

AI Use Cas­es are Broad­en­ing: Key Take­aways from the 2024 J.P. Mor­gan Health­care Con­fer­ence and a Deep Dive into Kan­vas Bio’s State of the Art Models 

Post­ed by Kan­vas Bio

We had the priv­i­lege of attend­ing the J.P. Mor­gan Health­care Con­fer­ence ear­li­er this month, which brings togeth­er thou­sands of glob­al health­care lead­ers, emerg­ing com­pa­nies and investors. Giv­en the scale and cal­iber of the event, many major media out­lets pub­lished in-depth recaps; how­ev­er, as an ear­ly-stage start­up oper­at­ing in the trench­es of the health tech sec­tor, we want­ed to chime in with our own takeaways: 

  • One of the most com­mon top­ics of con­ver­sa­tion was how health­care M&A will increase in 2024. Large phar­ma­ceu­ti­cal com­pa­nies will need to fill their late-stage pipelines to off­set patent expiries and poten­tial fur­ther pric­ing pres­sure from the Infla­tion Reduc­tion Act. For star­tups, rais­ing cap­i­tal will remain an issue until inter­est rates low­er and/​or VCs achieve more exits through M&As or IPOs. 

  • Every­one we spoke to said they’re using AI, but that’s no longer good enough. The con­sen­sus was, com­pa­nies need to be able to explain what they’re using AI for, and they need to be able to defend if their mod­els are high-qual­i­ty, if those mod­els are trained on high-qual­i­ty data, and how they’re solv­ing for AI bias and hallucination. 

  • There was pal­pa­ble buzz around cell and gene ther­a­py. We’ll be espe­cial­ly inter­est­ed to see if the live bio­ther­a­peu­tic prod­uct (LBP) space can ben­e­fit from the path human cell ther­a­pies are cur­rent­ly forging.

Unsur­pris­ing­ly, AI was a hot top­ic and we were excit­ed to wit­ness a broad­en­ing num­ber of AI use cas­es that extend across the R&D val­ue chain (clin­i­cal tri­al recruit­ment, pro­to­col writ­ing, study design and pre­dict­ing func­tion from struc­ture). We also had sev­er­al con­ver­sa­tions about the trade­offs between train­ing as a gen­er­al­ist tool (like Chat­G­PT) with poten­tial hal­lu­ci­na­tions vs. train­ing for very spe­cif­ic tasks with no risk of hallucinations. 

Giv­en the focus on AI at the con­fer­ence, we want­ed to address how Kan­vas Bio is using AI, and share a first hand look into how we’re train­ing our mod­els and the chal­lenges we’ve had to overcome:

Kan­vas Bio is using AI for pro­cess­ing and inter­pret­ing hyper­spec­tral image data at the sin­gle-cell lev­el. More specif­i­cal­ly, we’re using our own pro­pri­etary AI mod­els to decode the tax­on­o­my of microbes and iden­ti­ty of host cells from high-dimen­sion­al spec­tral images of bio­log­i­cal sam­ples, such as tis­sues and fecal mat­ter. This hypoth­e­sis-free frame­work allows us to build microbe-microbe and host-microbe inter­ac­tion maps, and we’re using these maps to cre­ate large atlases of host-micro­bio­me inter­ac­tions. These high qual­i­ty atlases will be used to gen­er­ate hypothe­ses to under­stand the mech­a­nisms of Kan­vas Bio’s nov­el therapeutics. 

Our AI mod­els and train­ing require­ments are unique, due to the nov­el nature and mas­sive amounts of data we’re work­ing with. But at Kan­vas Bio, we have the advan­tage of using spec­tral flu­o­res­cent images to improve our train­ing data. We’re lever­ag­ing this high-dimen­sion­al spec­tral data to gen­er­ate high-qual­i­ty anno­ta­tions of com­plex micro­bial envi­ron­ments, which would oth­er­wise be chal­leng­ing with tra­di­tion­al non-spec­tral flu­o­res­cent images. 

Because Kan­vas Bio is lead­ing the field of map­ping host-micro­bio­me inter­ac­tions, often we’re gen­er­at­ing data that’s nev­er been seen before – and that presents chal­lenges. For exam­ple, we had to adapt our analy­sis strat­e­gy for tis­sue sam­ples to address prob­lems such as high den­si­ty of microbes that are often stacked on each oth­er, pres­ence of bright food par­ti­cles, and back­ground aut­o­flu­o­res­cence. All of these chal­lenges require care­ful anno­ta­tion of ground truth data for train­ing an AI mod­el, which is an active area of development. 

To suc­cess­ful­ly lever­age com­plex data that’s nev­er been seen before, we use Omni­pose, a deep-learn­ing-based image seg­men­ta­tion plat­form cre­at­ed by one of our team mem­bers. Omni­pose uses a deep neur­al net­work archi­tec­ture called a U‑net to trans­form images into vec­tor field com­po­nents that define cell bound­aries and cell inte­ri­ors. This approach works well on com­plex tis­sue images not only because a U‑net can be trained on any image type, but also because the net­work of Omni­pose specif­i­cal­ly is trained to pre­dict a vec­tor field that is uni­ver­sal­ly applic­a­ble to any cell mor­phol­o­gy. In oth­er words, the U‑net can be trained to trans­form any image type, and its trans­for­ma­tions can describe any cell type. 

Because our biggest chal­lenge is the cura­tion of ground-truth train­ing data, we plan to use state of the art image syn­the­sis, such as dif­fu­sion mod­els, in com­bi­na­tion with opti­cal physics sim­u­la­tions to gen­er­ate syn­thet­ic train­ing data for image seg­men­ta­tion and clas­si­fi­ca­tion. This approach of syn­the­siz­ing labels and cor­re­spond­ing images will allow us to build anno­tat­ed datasets much larg­er than those we could cre­ate man­u­al­ly, which can be used to build mod­els that are vast­ly more capa­ble and generalizable.

Look­ing ahead, we’re espe­cial­ly excit­ed to lever­age the unprece­dent­ed host-micro­bio­me insights that we’re gen­er­at­ing by build­ing AI drug dis­cov­ery mod­els trained on our pro­pri­etary data to under­stand the design prin­ci­ples for clin­i­cal­ly ben­e­fi­cial micro­bial com­mu­ni­ties. Togeth­er, these prin­ci­ples set the stage for design­er pre­ci­sion micro­bio­me ther­a­peu­tics that go well beyond what can be achieved with fecal micro­bio­ta trans­plants (FMTs) and con­ven­tion­al LBPs. The potent com­bi­na­tion of our pro­pri­etary strain libraries and cut­ting edge AI mod­els will pave the way for the cre­ation of super­charged” design­er micro­bio­mes, where key func­tion­al and meta­bol­ic capa­bil­i­ties are pre­cise­ly added to exist­ing microflo­ra. Ulti­mate­ly, Kan­vas Bio is on a mis­sion to pro­duce pre­ci­sion micro­bio­me ther­a­peu­tics that exceed the lim­i­ta­tions of fecal micro­bio­ta trans­plan­ta­tion (FMT) and con­ven­tion­al live bac­te­r­i­al prod­ucts (LBPs), and con­tin­u­al­ly refin­ing our AI mod­els is a key com­po­nent of achiev­ing that mission.

To learn more about how Kan­vas Bio is using AI to accel­er­ate the dis­cov­ery and devel­op­ment of micro­bio­me-based ther­a­peu­tics, reach out to us direct­ly!