New AI Plat­form Elu­cid­ates Reg­u­lat­ory Activ­ity in the Genome

Research­ers at the Uni­ver­sity of Cali­for­nia, San Diego (UCSD) have developed a deep learn­ing soft­ware, named EUGENe (gen­om­ic ele­ments with neur­al nets), to aid in the study of gene reg­u­lat­ory mech­an­isms. Detailed in Nature Com­pu­ta­tion­al Sci­ence, EUGENe is designed to stream­line the exe­cu­tion of extract­ing, trans­form­ing sequence data, train­ing com­pu­ta­tion­al mod­els, and inter­pret­ing res­ults in reg­u­lat­ory genomics.

Research­ers invest­ig­at­ing the com­plex gene reg­u­lat­ory mech­an­isms involved in healthy and dis­ordered bio­lo­gic­al pro­cesses now have a new tool in their kit. Research­ers at the Uni­ver­sity of Cali­for­nia, San Diego (UCSD), and else­where have developed deep learn­ing soft­ware that they claim can be adap­ted to work for vari­ous gen­om­ics pro­jects. Details of the soft­ware, dubbed gen­om­ic ele­ments with neur­al nets or EUGENe, are provided in Nature Com­pu­ta­tion­al Sci­ence in a paper titled, ​“Pre­dict­ive ana­lys­is of reg­u­lat­ory sequences with EUGENe.”

Accord­ing to the paper, EUGENe’ com­prises vari­ous mod­ules and sub­pack­ages for extract­ing and trans­form­ing sequence data, instan­ti­at­ing and train­ing com­pu­ta­tion­al mod­els, and eval­u­at­ing and inter­pret­ing how the mod­els behave after train­ing. ​“The major goal of EUGENe is to stream­line the end-to-end exe­cu­tion of these three stages to pro­mote the effect­ive design, imple­ment­a­tion, val­id­a­tion, and inter­pret­a­tion of deep-learn­ing solu­tions in reg­u­lat­ory gen­om­ics,” the sci­ent­ists wrote. 

Deep learn­ing is cer­tainly not new to the gen­om­ics com­munity. As an example, the tech­no­logy has suc­cess­fully been used to detect DNA and RNA pro­tein bind­ing motifs and to make pre­dic­tions about chro­mat­in states and tran­scrip­tion­al activ­ity. But design­ing and deploy­ing deep learn­ing-based work­flows for gen­om­ics stud­ies has always been chal­len­ging even for exper­i­enced research­ers. That’s at least in part because ​“nuances spe­cif­ic to gen­om­ics data cre­ate an espe­cially high learn­ing curve for per­form­ing ana­lyses in this space. On top of this, the het­ero­gen­eity in imple­ment­a­tions of most code asso­ci­ated with pub­lic­a­tions greatly hinders extens­ib­il­ity and repro­du­cib­il­ity,” the authors wrote. 

Adam Klie, a PhD stu­dent at UCSD School of Medi­cine and the study’s first author, designed the soft­ware to mit­ig­ate those chal­lenges which he also exper­i­enced in his own work. ​“A lot of exist­ing plat­forms require many hours of cod­ing and data wrangling to use,” he said. EUGENe is much sim­pler to oper­ate. “[Y]ou give an algorithm a sequence of DNA and ask it to make pre­dic­tions about any­thing you’d expect that DNA could pre­dict, such as wheth­er a par­tic­u­lar DNA sequence is func­tion­al or wheth­er it reg­u­lates a gene in a cer­tain bio­lo­gic­al con­text.” Sci­ent­ists can use the soft­ware to explore the vari­ous prop­er­ties of the sequence in ques­tion and what hap­pens when things are modified. 

The research­ers put EUGENe through its paces by attempt­ing to repro­duce the res­ults of three reg­u­lat­ory gen­om­ics stud­ies that use dif­fer­ent types of sequen­cing data. These data­sets came from an assay of plant pro­moters, RNA bind­ing pro­tein spe­cificity data, and ChIP-sequen­cing data from the ENCODE pro­ject. Ana­lyz­ing dif­fer­ent types of data would typ­ic­ally require mix­ing and match­ing mul­tiple tech­no­logy plat­forms. How­ever, the sci­ent­ists were able to suc­cess­fully adapt EUGENe to each data type and repro­duce the find­ings of each study. 

The abil­ity to do this type of repro­du­cible ana­lys­is is crit­ic­al in sci­entif­ic research but can be chal­len­ging for stud­ies that use deep learn­ing, Han­nah Carter, PhD, asso­ci­ate pro­fess­or at UCSD School of Medi­cine and one of the authors on the paper, noted. ​“EUGENe is already show­ing a lot of prom­ise in how adapt­able it is to dif­fer­ent types of DNA sequen­cing data and sup­port­ing a lot of dif­fer­ent deep learn­ing mod­els. We hope it will evolve into a plat­form that can sup­port col­lab­or­at­ive tool devel­op­ment by the research com­munity and accel­er­ate gen­om­ics research.”

At the moment, the solu­tion works with DNA and RNA data but ​“does not have ded­ic­ated func­tions for hand­ling pro­tein sequence or mul­timod­al inputs,” the research­ers wrote. They plan to expand it to include new data types such as single-cell sequencing. 

They will also make the solu­tion avail­able more broadly to the sci­entif­ic com­munity. ​“Deep learn­ing can provide valu­able insights into the bio­lo­gic­al machinery driv­ing this vari­ety, but it can be chal­len­ging to imple­ment for research­ers without extens­ive com­puter sci­ence expert­ise,” Carter said. ​“We wanted to cre­ate a plat­form that can help gen­om­ics research­ers stream­line their deep learn­ing data ana­lys­is to make pre­dic­tions from raw data.”